added app file

.gitattributes

@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tunisian.arpa filter=lfs diff=lfs merge=lfs -text

Untitled.ipynb

@@ -0,0 +1,242 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "71d69be2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torchaudio\n",
+    "import numpy as np \n",
+    "import torch\n",
+    "import pandas as pd\n",
+    "import os"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "eb5c6da2",
+   "metadata": {
+    "scrolled": true
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "['Salah1.wav',\n",
+       " 'Salah2.wav',\n",
+       " 'Salah3.wav',\n",
+       " 'Salah4.wav',\n",
+       " 'Salah5.wav',\n",
+       " 'Salah6.wav',\n",
+       " 'Salah7.wav',\n",
+       " 'Salah8.wav',\n",
+       " 'Salah9.wav',\n",
+       " 'Salah10.wav']"
+      ]
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "files = os.listdir(\"./\")\n",
+    "files = [x for x in files if \".wav\" in x]\n",
+    "files = [f\"Salah{i}.wav\" for i in range(1,11)]\n",
+    "files"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "b2be1d8e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "words = {}\n",
+    "words[1] =  \"نحب ماكلة بنينة كسكروت نظيف و رخيص\"\n",
+    "words[2]=  \"باهي وقتاش نمشيو ال تونس\"\n",
+    "words[3] = \"اعطيني خمسة الاف و خمسة ميا بلاهي\"\n",
+    "words[4] = \"تعبت هاني راكش في الدار\"\n",
+    "words[5] = \"نهار السبت ماشي نقرى ان شاء الله\"\n",
+    "words[6]= \"زعما نلقى أحمد في الستاد ولا ماهوش هوني\"\n",
+    "words[7]= \"نحب نمشي ال بنزرت نرتاح شوية\"\n",
+    "words[8] = \"حكيت مع لولاد قالولي كل شي مريقل نهار السبت\"\n",
+    "words[9] = \"ناكل كفتاجي و نجم نشري شوية حوت زادة\"\n",
+    "words[10] = \"انتي خويا و عشيري صالح نحبك\"\n",
+    "words = [words[i] for i in range(1,11)]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "c46588ba",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "torch.Size([1, 238080])\n",
+      "torch.Size([1, 184320])\n",
+      "torch.Size([1, 207360])\n",
+      "torch.Size([1, 168960])\n",
+      "torch.Size([1, 168960])\n",
+      "torch.Size([1, 192000])\n",
+      "torch.Size([1, 184320])\n",
+      "torch.Size([1, 199680])\n",
+      "torch.Size([1, 230400])\n",
+      "torch.Size([1, 192000])\n"
+     ]
+    }
+   ],
+   "source": [
+    "durations= []\n",
+    "path_jz = \"samples/\"\n",
+    "paths  = [os.path.join(path_jz,x) for x in files]\n",
+    "srs= [48000 for x in paths]\n",
+    "IDs=[]\n",
+    "for f in files: \n",
+    "    x,sr = torchaudio.load(f)\n",
+    "    new_audio = torch.mean(x, dim=0).unsqueeze(0)\n",
+    "    print(new_audio.shape)\n",
+    "    torchaudio.save(os.path.join(\"monoaudiotun\", f), new_audio, sr)\n",
+    "    duration = float(x.shape[1]) / sr\n",
+    "    durations.append(duration)\n",
+    "    IDs.append(f.split(\".\")[0])\n",
+    "    \n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "b71db098",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "test_salah = pd.DataFrame(\n",
+    "    {'ID': IDs,\n",
+    "     'duration': durations,\n",
+    "     'wav': paths,\n",
+    "     \"sr\": srs,\n",
+    "     \"wrd\": words\n",
+    "    })\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "b3fdd365",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "test_salah.to_csv(\"test_salah_local.csv\", index=False)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 28,
+   "id": "f6ac8451",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "%WER 45.59 [ 31 / 68, 3 ins, 7 del, 21 sub ]\n",
+      "%SER 90.00 [ 9 / 10 ]\n",
+      "Scored 10 sentences, 0 not present in hyp.\n",
+      "================================================================================\n",
+      "ALIGNMENTS\n",
+      "\n",
+      "Format:\n",
+      "<utterance-id>, WER DETAILS\n",
+      "<eps> ; reference  ; on ; the ; first ;  line\n",
+      "  I   ;     S      ; =  ;  =  ;   S   ;   D  \n",
+      " and  ; hypothesis ; on ; the ; third ; <eps>\n",
+      "================================================================================\n",
+      "Salah4, %WER 0.00 [ 0 / 5, 0 ins, 0 del, 0 sub ]\n",
+      "تعبت ; هاني ; راكش ; في ; الدار\n",
+      " =   ;  =   ;  =   ; =  ;   =  \n",
+      "تعبت ; هاني ; راكش ; في ; الدار\n",
+      "================================================================================\n",
+      "Salah5, %WER 57.14 [ 4 / 7, 0 ins, 1 del, 3 sub ]\n",
+      "نهار ; السبت ; ماشي ; نقرى ; ان ;   شاء   ;  ا��له\n",
+      " =   ;   =   ;  =   ;  S   ; S  ;    S    ;   D  \n",
+      "نهار ; السبت ; ماشي ; نقرا ; إن ; شاءالله ; <eps>\n",
+      "================================================================================\n",
+      "Salah2, %WER 60.00 [ 3 / 5, 0 ins, 1 del, 2 sub ]\n",
+      "باهي ; وقتاش ; نمشيو  ;   ال  ;  تونس\n",
+      " =   ;   =   ;   S    ;   S   ;   D  \n",
+      "باهي ; وقتاش ; نمشيوا ; لتونس ; <eps>\n",
+      "================================================================================\n",
+      "Salah7, %WER 33.33 [ 2 / 6, 0 ins, 1 del, 1 sub ]\n",
+      "نحب ; نمشي ;   ال   ; بنزرت ; نرتاح ; شوية\n",
+      " =  ;  =   ;   S    ;   D   ;   =   ;  =  \n",
+      "نحب ; نمشي ; لبنزرت ; <eps> ; نرتاح ; شوية\n",
+      "================================================================================\n",
+      "Salah6, %WER 37.50 [ 3 / 8, 0 ins, 0 del, 3 sub ]\n",
+      "زعما ; نلقى ; أحمد ; في ; الستاد ; ولا  ; ماهوش ; هوني\n",
+      " S   ;  =   ;  =   ; =  ;   S    ;  S   ;   =   ;  =  \n",
+      "زعمة ; نلقى ; أحمد ; في ; السعد  ; وإلا ; ماهوش ; هوني\n",
+      "================================================================================\n",
+      "Salah10, %WER 83.33 [ 5 / 6, 1 ins, 1 del, 3 sub ]\n",
+      "انتي ; <eps> ; خويا ;   و    ; عشيري ;  صالح ; نحبك\n",
+      " S   ;   I   ;  =   ;   S    ;   S   ;   D   ;  =  \n",
+      "إنت  ;   ي   ; خويا ; وعشيلي ;  صلاح ; <eps> ; نحبك\n",
+      "================================================================================\n",
+      "Salah8, %WER 44.44 [ 4 / 9, 2 ins, 0 del, 2 sub ]\n",
+      "حكيت ; مع ;  لولاد  ; قالولي ; كل ; شي ; مريقل ; <eps> ; <eps> ; نهار ; السبت\n",
+      " =   ; =  ;    S    ;   =    ; =  ; =  ;   S   ;   I   ;   I   ;  =   ;   =  \n",
+      "حكيت ; مع ; الأولاد ; قالولي ; كل ; شي ;   مر  ;   ي   ;   ل   ; نهار ; السبت\n",
+      "================================================================================\n",
+      "Salah3, %WER 85.71 [ 6 / 7, 0 ins, 1 del, 5 sub ]\n",
+      "اعطيني ; خمسة ; الاف ;   و   ; خمسة ;   ميا   ; بلاهي\n",
+      "  S    ;  =   ;  S   ;   S   ;  S   ;    S    ;   D  \n",
+      "أعطيني ; خمسة ; آلاف ; وخمسة ; ملا  ; باللاهي ; <eps>\n",
+      "================================================================================\n",
+      "Salah9, %WER 25.00 [ 2 / 8, 0 ins, 1 del, 1 sub ]\n",
+      "ناكل ; كفتاجي ;   و   ;  نجم  ; نشري ; شوية ; حوت ; زادة\n",
+      " =   ;   =    ;   S   ;   D   ;  =   ;  =   ;  =  ;  =  \n",
+      "ناكل ; كفتاجي ; وننجم ; <eps> ; نشري ; شوية ; حوت ; زادة\n",
+      "================================================================================\n",
+      "Salah1, %WER 28.57 [ 2 / 7, 0 ins, 1 del, 1 sub ]\n",
+      "نحب ; ماكلة ; بنينة ; كسكروت ; نظيف ;   و   ;  رخيص\n",
+      " =  ;   =   ;   =   ;   =    ;  =   ;   S   ;   D  \n",
+      "نحب ; ماكلة ; بنينة ; كسكروت ; نظيف ; ورخيص ; <eps>\n"
+     ]
+    }
+   ],
+   "source": [
+    "filein = \"wer_test_salah.txt\"\n",
+    "with open(filein, \"r\") as wer : \n",
+    "    lines = wer.read().splitlines()\n",
+    "    print(\"\\n\".join(lines))"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.8.5"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

app.py

@@ -0,0 +1,371 @@
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+import torchaudio
+import numpy as np
+
+from pyctcdecode import build_ctcdecoder
+hparams_file, run_opts, overrides = sb.parse_arguments(["wavlm_partly_frozen.yaml"])
+
+# If distributed_launch=True then
+# create ddp_group with the right communication protocol
+sb.utils.distributed.ddp_init_group(run_opts)
+
+with open(hparams_file) as fin:
+    hparams = load_hyperpyyaml(fin, overrides)
+
+# Create experiment directory
+sb.create_experiment_directory(
+    experiment_directory=hparams["output_folder"],
+    hyperparams_to_save=hparams_file,
+    overrides=overrides,
+)
+def read_labels_file(labels_file): 
+    with open(labels_file, "r") as lf: 
+        lines = lf.read().splitlines()
+        division = "==="
+        numbers = {}
+        for line in lines : 
+            if division in line : 
+                break
+            string, number = line.split("=>")
+            number = int(number)
+            string = string[1:-2]
+            numbers[number] = string
+        return [numbers[x] for x in range(len(numbers))]
+labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+print(labels)
+labels = [""] + labels[1:]
+print(len(labels))
+
+# Dataset prep (parsing Librispeech)
+
+resampler_8000  = T.Resample(8000, 16000, dtype=torch.float)
+
+resampler_44100 =T.Resample(44100, 16000, dtype=torch.float)
+resampler_48000 =T.Resample(48000, 16000, dtype=torch.float)
+
+
+resamplers = {"8000": resampler_8000, "44100":resampler_44100, "48000": resampler_48000}
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav", "sr")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav, sr):
+        sig = sb.dataio.dataio.read_audio(wav)
+        sig = resamplers[sr](sig)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        print(wavs)
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+        x = self.modules.enc(feats)
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def treat_wav(self,sig):
+        feats = self.modules.wav2vec2(sig.to(self.device))
+        x = self.modules.enc(feats)
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        predicted_words =[]
+        for logs in p_ctc:
+            text = decoder.decode(logs.detach().cpu().numpy())
+            predicted_words.append(text.split(" "))
+        return " ".join(predicted_words[0])
+
+
+
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words =[]
+            for logs in p_ctc:
+                text = decoder.decode(logs.detach().cpu().numpy())
+                predicted_words.append(text.split(" "))
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+
+# We dynamicaly add the tokenizer to our brain class.
+# NB: This tokenizer corresponds to the one used for the LM!!
+decoder = build_ctcdecoder(
+    labels,
+    kenlm_model_path="tunisian.arpa",  # either .arpa or .bin file
+    alpha=0.5,  # tuned on a val set
+    beta=1,  # tuned on a val set
+)
+
+asr_brain = ASR(
+    modules=hparams["modules"],
+    hparams=hparams,
+    run_opts=run_opts,
+    checkpointer=hparams["checkpointer"],
+)
+asr_brain.device= "cpu"
+asr_brain.modules.to("cpu")
+asr_brain.tokenizer = label_encoder
+
+from enum import Enum, auto
+class Stage(Enum):
+    TRAIN = auto()
+    VALID = auto()
+    TEST = auto()
+
+asr_brain.on_evaluate_start()
+asr_brain.modules.eval()
+import gradio as gr
+def treat_wav_file(file_mic, file_upload, resamplers = resamplers,asr=asr_brain, device="cpu") :
+    
+    if (file_mic is not None) and (file_upload is not None):
+       warn_output = "WARNING: You've uploaded an audio file and used the microphone. The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
+       wav = file_mic
+    elif (file_mic is None) and (file_upload is None):
+       return "ERROR: You have to either use the microphone or upload an audio file"
+    elif file_mic is not None:
+       wav = file_mic
+    else:
+       wav = file_upload
+    sig, sr = torchaudio.load(wav)
+    tensor_wav = sig.to(device)
+    resampled = resamplers[str(sr)](tensor_wav)
+    sentence = asr_brain.treat_wav(resampled)
+    return sentence
+
+gr.Interface(
+    fn=treat_wav_file, 
+    inputs=[gr.inputs.Audio(source="microphone", type='filepath', optional=True),
+        gr.inputs.Audio(source="upload", type='filepath', optional=True)]
+    ,outputs="text").launch()
+
+
+

ctc_train.py

@@ -0,0 +1,339 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+logger = logging.getLogger(__name__)
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+        x = self.modules.enc(feats)
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav", "sr")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav, sr):
+        sig = sb.dataio.dataio.read_audio(wav)
+        sig = resamplers[sr](sig)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+
+    # Dataset prep (parsing Librispeech)
+
+    resampler_8000  = T.Resample(8000, 16000, dtype=torch.float)
+
+    resampler_44100 =T.Resample(44100, 16000, dtype=torch.float)
+    resampler_32000 =T.Resample(32000, 16000, dtype=torch.float)
+    resampler_48000 =T.Resample(48000, 16000, dtype=torch.float)
+
+
+    resamplers = {"48000": resampler_48000,"8000": resampler_8000, "44100":resampler_44100, "32000":resampler_32000}
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+    asr_brain.device= "cpu"
+    asr_brain.modules.to("cpu")
+
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

debugging.csv

@@ -0,0 +1,11 @@
+ID,duration,wav,sr,wrd
+Salah1,4.96,samples/Salah1.wav,48000,نحب ماكلة بنينة كسكروت نظيف و رخيص
+Salah2,3.84,samples/Salah2.wav,48000,باهي وقتاش نمشيو ال تونس
+Salah3,4.32,samples/Salah3.wav,48000,اعطيني خمسة الاف و خمسة ميا بلاهي
+Salah4,3.52,samples/Salah4.wav,48000,تعبت هاني راكش في الدار
+Salah5,3.52,samples/Salah5.wav,48000,نهار السبت ماشي نقرى ان شاء الله
+Salah6,4.0,samples/Salah6.wav,48000,زعما نلقى أحمد في الستاد ولا ماهوش هوني
+Salah7,3.84,samples/Salah7.wav,48000,نحب نمشي ال بنزرت نرتاح شوية
+Salah8,4.16,samples/Salah8.wav,48000,حكيت مع لولاد قالولي كل شي مريقل نهار السبت
+Salah9,4.8,samples/Salah9.wav,48000,ناكل كفتاجي و نجم نشري شوية حوت زادة
+Salah10,4.0,samples/Salah10.wav,48000,انتي خويا و عشيري صالح نحبك

lm_decoded_ctc.py

@@ -0,0 +1,376 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+from pyctcdecode import build_ctcdecoder
+
+logger = logging.getLogger(__name__)
+
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+
+        x = self.modules.enc(feats.detach())[0]
+        #x = self.modules.enc(feats.detach())
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            predicted_words =[]
+            for logs in p_ctc: 
+                text = decoder.decode(logs.detach().cpu().numpy())
+                predicted_words.append(text.split(" "))
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav):
+        sig = sb.dataio.dataio.read_audio(wav)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+    def read_labels_file(labels_file): 
+        with open(labels_file, "r") as lf: 
+            lines = lf.read().splitlines()
+            division = "==="
+            numbers = {}
+            for line in lines : 
+                if division in line : 
+                    break
+                string, number = line.split("=>")
+                number = int(number)
+                string = string[1:-2]
+                numbers[number] = string
+            return [numbers[x] for x in range(len(numbers))]
+    labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+    print(labels)
+    labels = [""] + labels[1:]
+    print(len(labels))
+    decoder = build_ctcdecoder(
+        labels,
+        kenlm_model_path="/gpfsstore/rech/nou/uzn19yk/4-gram.arpa",  # either .arpa or .bin file
+        alpha=0.5,  # tuned on a val set
+        beta=1.0,  # tuned on a val set
+    )
+
+    # Dataset prep (parsing Librispeech)
+
+    # multi-gpu (ddp) save data preparation
+    """
+    run_on_main(
+        prepare_librispeech,
+        kwargs={
+            "data_folder": hparams["data_folder"],
+            "tr_splits": hparams["train_splits"],
+            "dev_splits": hparams["dev_splits"],
+            "te_splits": hparams["test_splits"],
+            "save_folder": hparams["output_folder"],
+            "merge_lst": hparams["train_splits"],
+            "merge_name": "train.csv",
+            "skip_prep": hparams["skip_prep"],
+        },
+    )
+    """
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+    
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

lm_tunisian.py

@@ -0,0 +1,361 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+
+from pyctcdecode import build_ctcdecoder
+logger = logging.getLogger(__name__)
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+        x = self.modules.enc(feats)
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words =[]
+            for logs in p_ctc: 
+                text = decoder.decode(logs.detach().cpu().numpy())
+                predicted_words.append(text.split(" "))
+
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav", "sr")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav, sr):
+        sig = sb.dataio.dataio.read_audio(wav)
+        sig = resamplers[sr](sig)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+    def read_labels_file(labels_file): 
+        with open(labels_file, "r") as lf: 
+            lines = lf.read().splitlines()
+            division = "==="
+            numbers = {}
+            for line in lines : 
+                if division in line : 
+                    break
+                string, number = line.split("=>")
+                number = int(number)
+                string = string[1:-2]
+                numbers[number] = string
+            return [numbers[x] for x in range(len(numbers))]
+    labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+    print(labels)
+    labels = [""] + labels[1:]
+    print(len(labels))
+    decoder = build_ctcdecoder(
+        labels,
+        kenlm_model_path="tunisian.arpa",  # either .arpa or .bin file
+        alpha=0.5,  # tuned on a val set
+        beta=1.0,  # tuned on a val set
+    )
+
+    # Dataset prep (parsing Librispeech)
+
+    resampler_8000  = T.Resample(8000, 16000, dtype=torch.float)
+
+    resampler_44100 =T.Resample(44100, 16000, dtype=torch.float)
+    resampler_48000 =T.Resample(48000, 16000, dtype=torch.float)
+    resamplers = {"8000": resampler_8000, "44100":resampler_44100, "48000": resampler_48000}
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+    asr_brain.device= "cpu"
+    asr_brain.modules.to("cpu")
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

partly_frozen_splitted_wavlm/1986/ctc_train.py

@@ -0,0 +1,339 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+logger = logging.getLogger(__name__)
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+        x = self.modules.enc(feats)
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav", "sr")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav, sr):
+        sig = sb.dataio.dataio.read_audio(wav)
+        sig = resamplers[sr](sig)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+
+    # Dataset prep (parsing Librispeech)
+
+    resampler_8000  = T.Resample(8000, 16000, dtype=torch.float)
+
+    resampler_44100 =T.Resample(44100, 16000, dtype=torch.float)
+    resampler_32000 =T.Resample(32000, 16000, dtype=torch.float)
+    resampler_48000 =T.Resample(48000, 16000, dtype=torch.float)
+
+
+    resamplers = {"48000": resampler_48000,"8000": resampler_8000, "44100":resampler_44100, "32000":resampler_32000}
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+    asr_brain.device= "cpu"
+    asr_brain.modules.to("cpu")
+
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

partly_frozen_splitted_wavlm/1986/env.log

@@ -0,0 +1,402 @@
+SpeechBrain system description
+==============================
+Python version:
+3.8.5 (default, Sep  4 2020, 07:30:14) 
+[GCC 7.3.0]
+==============================
+Installed Python packages:
+abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altair==4.2.0
+altgraph==0.17
+antlr4-python3-runtime==4.8
+anyio==3.6.2
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fastapi==0.89.0
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+ffmpy==0.3.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+gradio==3.16.0
+greenlet==1.1.2
+grpcio==1.32.0
+h11==0.14.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+httpcore==0.16.3
+httpx==0.23.3
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywebrtc==0.6.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+linkify-it-py==1.0.3
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+markdown-it-py==2.1.0
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+orjson==3.8.4
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pycryptodome==3.16.0
+pyctcdecode==0.4.0
+pydantic==1.10.4
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-multipart==0.0.5
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+sniffio==1.3.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+starlette==0.22.0
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+toolz==0.12.0
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==4.4.0
+uc-micro-py==1.0.1
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+uvicorn==0.20.0
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+websockets==10.4
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+==============================
+Could not get git revision==============================
+CUDA version:
+11.7

partly_frozen_splitted_wavlm/1986/hyperparams.yaml

@@ -0,0 +1,162 @@
+# Generated 2023-01-08 from:
+# /home/salah/kenlm_train/to_copy/wavlm_partly_frozen.yaml
+# yamllint disable
+# ################################
+# Model: wav2vec2 + DNN + CTC
+# Augmentation: SpecAugment
+# Authors: Sung-Lin Yeh 2021
+# ################################
+
+# Seed needs to be set at top of yaml, before objects with parameters are made
+seed: 1986
+__set_seed: !apply:torch.manual_seed [1986]
+output_folder: partly_frozen_splitted_wavlm/1986/
+wer_file: partly_frozen_splitted_wavlm/1986//wer.txt
+save_folder: partly_frozen_splitted_wavlm/1986//save
+train_log: partly_frozen_splitted_wavlm/1986//train_log.txt
+
+# URL for the biggest Fairseq english wav2vec2 model.
+
+# Data files
+data_folder: /gpfsscratch/rech/nou/uzn19yk/Libri/LibriSpeech/ # e,g./path/to/LibriSpeech
+# noise/ris dataset will automatically be downloaded
+data_folder_rirs: /gpfsscratch/rech/nou/uzn19yk/Libri/LibriSpeech/
+train_splits: [train-clean-100]
+dev_splits: [dev-clean]
+test_splits: [test-clean, test-other]
+skip_prep: false
+ckpt_interval_minutes: 25 # save checkpoint every N min
+csv_folder: /gpfsstore/rech/nou/uzn19yk/iwslt/splitted_clean_tunisian_csvs/
+train_csv: test_salah_local.csv
+valid_csv: test_salah_local.csv
+test_csv:
+- test_salah_local.csv
+
+# Training parameters
+number_of_epochs: 12
+lr: 1
+lr_wav2vec: 0.0001
+sorting: ascending
+auto_mix_prec: false
+sample_rate: 16000
+
+avoid_if_longer_than: 10
+# With data_parallel batch_size is split into N jobs
+# With DDP batch_size is multiplied by N jobs
+# Must be 3 per GPU to fit 32GB of VRAM
+batch_size: 1
+test_batch_size: 1
+
+# Dataloader options
+train_dataloader_opts:
+  batch_size: 1
+
+valid_dataloader_opts:
+  batch_size: 1
+
+test_dataloader_opts:
+  batch_size: 1
+
+# Model parameters
+activation: &id001 !name:torch.nn.LeakyReLU
+dnn_layers: 2
+dnn_neurons: 1024
+freeze_wav2vec: false
+
+# Outputs
+output_neurons: 41  # BPE size, index(blank/eos/bos) = 0
+
+# Decoding parameters
+blank_index: 0
+bos_index: 1
+eos_index: 2
+
+#
+# Functions and classes
+#
+epoch_counter: &id008 !new:speechbrain.utils.epoch_loop.EpochCounter
+
+  limit: 12
+
+augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
+  sample_rate: 16000
+  speeds: [95, 100, 105]
+
+enc: &id003 !new:speechbrain.lobes.models.VanillaNN.VanillaNN
+  input_shape: [null, null, 1024]
+  activation: *id001
+  dnn_blocks: 2
+  dnn_neurons: 1024
+
+wav2vec2: &id002 !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
+  source: wavlm-large/
+  output_norm: true
+  freeze: false
+  freeze_feature_extractor: true
+  save_path: partly_frozen_splitted_wavlm/1986//save/wav2vec2_hubert_checkpoint
+
+#####
+# Uncomment this block if you prefer to use a Fairseq pretrained model instead
+# of a HuggingFace one. Here, we provide an URL that is obtained from the
+# Fairseq github for the multilingual XLSR.
+#
+#wav2vec2_url: https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_960h_pl.pt
+#wav2vec2: !new:speechbrain.lobes.models.fairseq_wav2vec.FairseqWav2Vec2
+#    pretrained_path: !ref <wav2vec2_url>
+#    output_norm: True
+#    freeze: False
+#    save_path: !ref <save_folder>/wav2vec2_checkpoint/model.pt
+
+ctc_lin: &id004 !new:speechbrain.nnet.linear.Linear
+
+  input_size: 1024
+  n_neurons: 41
+
+log_softmax: !new:speechbrain.nnet.activations.Softmax
+  apply_log: true
+
+ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
+  blank_index: 0
+
+modules:
+  wav2vec2: *id002
+  enc: *id003
+  ctc_lin: *id004
+model: &id005 !new:torch.nn.ModuleList
+- [*id003, *id004]
+model_opt_class: !name:torch.optim.Adadelta
+  lr: 1
+  rho: 0.95
+  eps: 1.e-8
+
+wav2vec_opt_class: !name:torch.optim.Adam
+  lr: 0.0001
+
+lr_annealing_model: &id006 !new:speechbrain.nnet.schedulers.NewBobScheduler
+  initial_value: 1
+  improvement_threshold: 0.0025
+  annealing_factor: 0.8
+  patient: 0
+
+lr_annealing_wav2vec: &id007 !new:speechbrain.nnet.schedulers.NewBobScheduler
+  initial_value: 0.0001
+  improvement_threshold: 0.0025
+  annealing_factor: 0.9
+  patient: 0
+
+
+checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
+  checkpoints_dir: partly_frozen_splitted_wavlm/1986//save
+  recoverables:
+    wav2vec2: *id002
+    model: *id005
+    scheduler_model: *id006
+    scheduler_wav2vec: *id007
+    counter: *id008
+train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
+  save_file: partly_frozen_splitted_wavlm/1986//train_log.txt
+
+error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+
+cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+  split_tokens: true

partly_frozen_splitted_wavlm/1986/lm_decoded_ctc.py

@@ -0,0 +1,377 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+from pyctcdecode import build_ctcdecoder
+
+logger = logging.getLogger(__name__)
+
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+
+        x = self.modules.enc(feats.detach())[0]
+        #x = self.modules.enc(feats.detach())
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            predicted_words =[]
+            for logs in p_ctc: 
+                text = decoder.decode(logs.detach().cpu().numpy())
+                predicted_words.append(text.split(" "))
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav):
+        sig = sb.dataio.dataio.read_audio(wav)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+    def read_labels_file(labels_file): 
+        with open(labels_file, "r") as lf: 
+            lines = lf.read().splitlines()
+            division = "==="
+            numbers = {}
+            for line in lines : 
+                if division in line : 
+                    break
+                string, number = line.split("=>")
+                number = int(number)
+                string = string[1:-2]
+                numbers[number] = string
+            return [numbers[x] for x in range(len(numbers))]
+    labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+    print(labels)
+    labels = [""] + labels[1:]
+    print(len(labels))
+    decoder = build_ctcdecoder(
+        labels,
+        kenlm_model_path="/gpfsstore/rech/nou/uzn19yk/4-gram.arpa",  # either .arpa or .bin file
+        alpha=0.5,  # tuned on a val set
+        beta=1.0,  # tuned on a val set
+    )
+
+    # Dataset prep (parsing Librispeech)
+    from librispeech_prepare import prepare_librispeech  # noqa
+
+    # multi-gpu (ddp) save data preparation
+    """
+    run_on_main(
+        prepare_librispeech,
+        kwargs={
+            "data_folder": hparams["data_folder"],
+            "tr_splits": hparams["train_splits"],
+            "dev_splits": hparams["dev_splits"],
+            "te_splits": hparams["test_splits"],
+            "save_folder": hparams["output_folder"],
+            "merge_lst": hparams["train_splits"],
+            "merge_name": "train.csv",
+            "skip_prep": hparams["skip_prep"],
+        },
+    )
+    """
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+    
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

partly_frozen_splitted_wavlm/1986/lm_tunisian.py

@@ -0,0 +1,361 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+
+from pyctcdecode import build_ctcdecoder
+logger = logging.getLogger(__name__)
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+        x = self.modules.enc(feats)
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words =[]
+            for logs in p_ctc: 
+                text = decoder.decode(logs.detach().cpu().numpy())
+                predicted_words.append(text.split(" "))
+
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav", "sr")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav, sr):
+        sig = sb.dataio.dataio.read_audio(wav)
+        sig = resamplers[sr](sig)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+    def read_labels_file(labels_file): 
+        with open(labels_file, "r") as lf: 
+            lines = lf.read().splitlines()
+            division = "==="
+            numbers = {}
+            for line in lines : 
+                if division in line : 
+                    break
+                string, number = line.split("=>")
+                number = int(number)
+                string = string[1:-2]
+                numbers[number] = string
+            return [numbers[x] for x in range(len(numbers))]
+    labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+    print(labels)
+    labels = [""] + labels[1:]
+    print(len(labels))
+    decoder = build_ctcdecoder(
+        labels,
+        kenlm_model_path="tunisian.arpa",  # either .arpa or .bin file
+        alpha=0.5,  # tuned on a val set
+        beta=1.0,  # tuned on a val set
+    )
+
+    # Dataset prep (parsing Librispeech)
+
+    resampler_8000  = T.Resample(8000, 16000, dtype=torch.float)
+
+    resampler_44100 =T.Resample(44100, 16000, dtype=torch.float)
+    resampler_48000 =T.Resample(48000, 16000, dtype=torch.float)
+    resamplers = {"8000": resampler_8000, "44100":resampler_44100, "48000": resampler_48000}
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+    asr_brain.device= "cpu"
+    asr_brain.modules.to("cpu")
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/CKPT.yaml

@@ -0,0 +1,4 @@
+# yamllint disable
+WER: 47.68855883035507
+end-of-epoch: true
+unixtime: 1672916345.1685827

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/brain.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33809a026a2c1febce7b03c8aafaee4ddfc851b2c70f180f8c06bf1017f4df5c
+size 46

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/counter.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b51d431df5d7f141cbececcf79edf3dd861c3b4069f0b11661a3eefacbba918
+size 2

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/dataloader-TRAIN.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d53eb07a66864f0a18b5a0c5d029b33bddb11f05025a6e385c92c9fbb618edee
+size 6

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8782ec6abb3dd1a6aa05d54a8f159375707d2ef212f2d934c1209aab5f01a46b
+size 8566935

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/modelopt.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18828222ecebe54247f2e16e2eedf607c5fc34d0b0029f4cc3356f9397a10802
+size 17133057

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/scheduler_model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:524fb9db201619c436c0f51539869e8ffd90c8b60bf30384726052e65076751a
+size 623

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/scheduler_wav2vec.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4663a4d525410c34529e6c80671db216a35c9f285f1cf9a1ca7004d3e14679c7
+size 623

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/wav2vec2.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96716b4e96d96f3aa00182e1e5bf99219b86f74990e2f05cb69165d6e96c8b4e
+size 1262004913

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+11-59-05+00/wav2vec_opt.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c39bdda929f079de1ab70cd624900b7fda0a6a6682446ab2a25296442d01862e
+size 2490235001

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/CKPT.yaml

@@ -0,0 +1,4 @@
+# yamllint disable
+brain_intra_epoch_ckpt: true
+end-of-epoch: false
+unixtime: 1672917851.8581367

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/brain.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bc72ed3a1d0a5dc95a83b7a139fce806d6929b8db82b1f8dd010d42556698ca
+size 65

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/counter.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fdba35f04dc8c462986c992bcf875546257113072a909c162f7e470e581e278
+size 2

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/dataloader-TRAIN.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8228f8d57233cd4d689b31048e7ec6e2c12b409b2501ac69252bbdb65ea575d
+size 5

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:316a78537d2603b0ef0030df65934944a0cc88a0865c7983aa77a081605d1d05
+size 8566935

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/modelopt.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d721ac115b10592cb716b2641f80edf21549fb11c74bcf443f8213249c7fa7e8
+size 17133057

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/scheduler_model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:524fb9db201619c436c0f51539869e8ffd90c8b60bf30384726052e65076751a
+size 623

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/scheduler_wav2vec.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4663a4d525410c34529e6c80671db216a35c9f285f1cf9a1ca7004d3e14679c7
+size 623

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/wav2vec2.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8616b0fbaeb370bfae2df4d924367ae88d4a17a644b22f7050a607242f036fe5
+size 1262004913

partly_frozen_splitted_wavlm/1986/save/CKPT+2023-01-05+12-24-11+00/wav2vec_opt.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3767ccf04ad421c4a71a167b10e0d31b064bf2f91f769ec42f9510f675dea148
+size 2490235001

partly_frozen_splitted_wavlm/1986/save/label_encoder.txt

@@ -0,0 +1,46 @@
+'م' => 38
+'و' => 39
+' ' => 40
+'ق' => 3
+'ا' => 4
+'ع' => 5
+'د' => 6
+'ة' => 7
+'ت' => 8
+'ش' => 9
+'ي' => 10
+'ك' => 11
+'ه' => 12
+'ل' => 13
+'ح' => 14
+'ب' => 15
+'ن' => 16
+'ى' => 17
+'ر' => 18
+'ف' => 19
+'إ' => 20
+'س' => 21
+'أ' => 22
+'ض' => 23
+'ص' => 24
+'ط' => 25
+'خ' => 26
+'ج' => 27
+'ظ' => 28
+'ز' => 29
+'آ' => 30
+'ذ' => 31
+'غ' => 32
+'ث' => 33
+'ئ' => 34
+'ء' => 35
+'ؤ' => 36
+'ٱ' => 37
+'<blank>' => 0
+'<bos>' => 1
+'<eos>' => 2
+================
+'starting_index' => 0
+'bos_label' => '<bos>'
+'eos_label' => '<eos>'
+'blank_label' => '<blank>'

partly_frozen_splitted_wavlm/1986/train_log.txt

@@ -0,0 +1,86 @@
+epoch: 1, lr_model: 1, lr_wav2vec: 1.00e-04 - train loss: 1.24 - valid loss: 9.24e-01, valid CER: 26.54, valid WER: 59.12
+epoch: 2, lr_model: 1, lr_wav2vec: 1.00e-04 - train loss: 9.67e-01 - valid loss: 9.03e-01, valid CER: 25.85, valid WER: 57.02
+epoch: 3, lr_model: 1, lr_wav2vec: 1.00e-04 - train loss: 8.84e-01 - valid loss: 8.81e-01, valid CER: 24.89, valid WER: 55.02
+epoch: 4, lr_model: 1, lr_wav2vec: 1.00e-04 - train loss: 8.19e-01 - valid loss: 8.31e-01, valid CER: 22.92, valid WER: 51.69
+epoch: 5, lr_model: 1, lr_wav2vec: 1.00e-04 - train loss: 7.76e-01 - valid loss: 8.67e-01, valid CER: 23.36, valid WER: 50.67
+epoch: 6, lr_model: 8.00e-01, lr_wav2vec: 9.00e-05 - train loss: 7.20e-01 - valid loss: 8.37e-01, valid CER: 22.87, valid WER: 49.84
+epoch: 7, lr_model: 8.00e-01, lr_wav2vec: 9.00e-05 - train loss: 6.87e-01 - valid loss: 8.78e-01, valid CER: 23.90, valid WER: 51.19
+epoch: 8, lr_model: 6.40e-01, lr_wav2vec: 8.10e-05 - train loss: 6.44e-01 - valid loss: 8.68e-01, valid CER: 23.03, valid WER: 49.83
+epoch: 9, lr_model: 6.40e-01, lr_wav2vec: 8.10e-05 - train loss: 6.20e-01 - valid loss: 8.47e-01, valid CER: 22.77, valid WER: 48.42
+epoch: 10, lr_model: 6.40e-01, lr_wav2vec: 8.10e-05 - train loss: 5.98e-01 - valid loss: 9.07e-01, valid CER: 24.31, valid WER: 49.76
+epoch: 11, lr_model: 5.12e-01, lr_wav2vec: 7.29e-05 - train loss: 5.60e-01 - valid loss: 9.08e-01, valid CER: 23.75, valid WER: 49.33
+epoch: 12, lr_model: 4.10e-01, lr_wav2vec: 6.56e-05 - train loss: 5.22e-01 - valid loss: 9.08e-01, valid CER: 22.61, valid WER: 47.69
+Epoch loaded: 12 - test loss: 1.26e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 5.95e-02, test CER: 20.52, test WER: 40.71
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 10.97, test WER: 54.41
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 10.03, test WER: 48.53
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 10.03, test WER: 48.53
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 10.03, test WER: 48.53
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 10.03, test WER: 48.53
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.72, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.40, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.40, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.40, test WER: 45.59
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 8.78, test WER: 44.12
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 10.66, test WER: 54.41
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 11.60, test WER: 60.29
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 7.97e-04, test CER: 9.09, test WER: 42.65
+Epoch loaded: 12 - test loss: 6.99e-05, test CER: 8.82, test WER: 42.86
+Epoch loaded: 12 - test loss: 6.99e-05, test CER: 8.82, test WER: 42.86
+Epoch loaded: 12 - test loss: 6.99e-05, test CER: 8.82, test WER: 42.86

partly_frozen_splitted_wavlm/1986/wer_test_salah.txt

@@ -0,0 +1,61 @@
+%WER 42.65 [ 29 / 68, 1 ins, 7 del, 21 sub ]
+%SER 90.00 [ 9 / 10 ]
+Scored 10 sentences, 0 not present in hyp.
+================================================================================
+ALIGNMENTS
+
+Format:
+<utterance-id>, WER DETAILS
+<eps> ; reference  ; on ; the ; first ;  line
+  I   ;     S      ; =  ;  =  ;   S   ;   D  
+ and  ; hypothesis ; on ; the ; third ; <eps>
+================================================================================
+Salah4, %WER 0.00 [ 0 / 5, 0 ins, 0 del, 0 sub ]
+تعبت ; هاني ; راكش ; في ; الدار
+ =   ;  =   ;  =   ; =  ;   =  
+تعبت ; هاني ; راكش ; في ; الدار
+================================================================================
+Salah5, %WER 57.14 [ 4 / 7, 0 ins, 1 del, 3 sub ]
+نهار ; السبت ; ماشي ; نقرى ; ان ;   شاء   ;  الله
+ =   ;   =   ;  =   ;  S   ; S  ;    S    ;   D  
+نهار ; السبت ; ماشي ; نقرا ; إن ; شاءالله ; <eps>
+================================================================================
+Salah2, %WER 60.00 [ 3 / 5, 0 ins, 1 del, 2 sub ]
+باهي ; وقتاش ; نمشيو  ;   ال  ;  تونس
+ =   ;   =   ;   S    ;   S   ;   D  
+باهي ; وقتاش ; نمشيوا ; لتونس ; <eps>
+================================================================================
+Salah7, %WER 33.33 [ 2 / 6, 0 ins, 1 del, 1 sub ]
+نحب ; نمشي ;   ال   ; بنزرت ; نرتاح ; شوية
+ =  ;  =   ;   S    ;   D   ;   =   ;  =  
+نحب ; نمشي ; لبنزرت ; <eps> ; نرتاح ; شوية
+================================================================================
+Salah6, %WER 37.50 [ 3 / 8, 0 ins, 0 del, 3 sub ]
+زعما ; نلقى ; أحمد ; في ; الستاد ; ولا  ; ماهوش ; هوني
+ S   ;  =   ;  =   ; =  ;   =    ;  S   ;   =   ;  S  
+زعمة ; نلقى ; أحمد ; في ; الستاد ; وإلا ; ماهوش ; كوني
+================================================================================
+Salah10, %WER 66.67 [ 4 / 6, 1 ins, 1 del, 2 sub ]
+انتي ; <eps> ; خويا ;   و    ; عشيري ; صالح ; نحبك
+ S   ;   I   ;  =   ;   S    ;   D   ;  =   ;  =  
+إنت  ;   ي   ; خويا ; وعشيلي ; <eps> ; صالح ; نحبك
+================================================================================
+Salah8, %WER 11.11 [ 1 / 9, 0 ins, 0 del, 1 sub ]
+حكيت ; مع ;  لولاد  ; قالولي ; كل ; شي ; مريقل ; نهار ; السبت
+ =   ; =  ;    S    ;   =    ; =  ; =  ;   =   ;  =   ;   =  
+حكيت ; مع ; الاولاد ; قالولي ; كل ; شي ; مريقل ; نهار ; السبت
+================================================================================
+Salah3, %WER 85.71 [ 6 / 7, 0 ins, 1 del, 5 sub ]
+اعطيني ; خمسة ; الاف ;   و   ; خمسة ;   ميا   ; بلاهي
+  S    ;  =   ;  S   ;   S   ;  S   ;    S    ;   D  
+أعطيني ; خمسة ; آلاف ; وخمسة ; مية  ; باللاهي ; <eps>
+================================================================================
+Salah9, %WER 37.50 [ 3 / 8, 0 ins, 1 del, 2 sub ]
+ناكل ;  كفتاجي  ;   و   ;  نجم  ; نشري ; شوية ; حوت ; زادة
+ =   ;    S     ;   S   ;   D   ;  =   ;  =   ;  =  ;  =  
+ناكل ; الكفتاجي ; وننجم ; <eps> ; نشري ; شوية ; حوت ; زادة
+================================================================================
+Salah1, %WER 42.86 [ 3 / 7, 0 ins, 1 del, 2 sub ]
+نحب ; ماكلة ; بنينة ; كسكروت ; نظيف ;   و   ;  رخيص
+ S  ;   =   ;   =   ;   =    ;  =   ;   S   ;   D  
+لحم ; ماكلة ; بنينة ; كسكروت ; نظيف ; ورخيص ; <eps>

partly_frozen_splitted_wavlm/1986/wer_test_salah_local.txt

@@ -0,0 +1,21 @@
+%WER 42.86 [ 6 / 14, 0 ins, 2 del, 4 sub ]
+%SER 100.00 [ 2 / 2 ]
+Scored 2 sentences, 0 not present in hyp.
+================================================================================
+ALIGNMENTS
+
+Format:
+<utterance-id>, WER DETAILS
+<eps> ; reference  ; on ; the ; first ;  line
+  I   ;     S      ; =  ;  =  ;   S   ;   D  
+ and  ; hypothesis ; on ; the ; third ; <eps>
+================================================================================
+Salah1, %WER 42.86 [ 3 / 7, 0 ins, 1 del, 2 sub ]
+نحب ; ماكلة ; بنينة ; كسكروت ; نظيف ;   و   ;  رخيص
+ S  ;   =   ;   =   ;   =    ;  =   ;   S   ;   D  
+لحم ; ماكلة ; بنينة ; كسكروت ; نظيف ; ورخيص ; <eps>
+================================================================================
+Salah2, %WER 42.86 [ 3 / 7, 0 ins, 1 del, 2 sub ]
+نحب ; ماكلة ; بنينة ; كسكروت ; نظيف ;   و   ;  رخيص
+ S  ;   =   ;   =   ;   =    ;  =   ;   S   ;   D  
+لحم ; ماكلة ; بنينة ; كسكروت ; نظيف ; ورخيص ; <eps>

partly_frozen_splitted_wavlm/ctc_train.py

@@ -0,0 +1,339 @@
+#!/usr/bin/env/python3
+"""Recipe for training a wav2vec-based ctc ASR system with librispeech.
+The system employs wav2vec as its encoder. Decoding is performed with
+ctc greedy decoder.
+To run this recipe, do the following:
+> python train_with_wav2vec.py hparams/train_with_wav2vec.yaml
+The neural network is trained on CTC likelihood target and character units
+are used as basic recognition tokens. Training is performed on the full
+LibriSpeech dataset (960 h).
+
+Authors
+ * Sung-Lin Yeh 2021
+ * Titouan Parcollet 2021
+ * Ju-Chieh Chou 2020
+ * Mirco Ravanelli 2020
+ * Abdel Heba 2020
+ * Peter Plantinga 2020
+ * Samuele Cornell 2020
+"""
+
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+logger = logging.getLogger(__name__)
+
+# Define training procedure
+class ASR(sb.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        tokens_bos, _ = batch.tokens_bos
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs)
+        x = self.modules.enc(feats)
+        # Compute outputs
+        p_tokens = None
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+        if stage != sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        return p_ctc, wav_lens, p_tokens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC+NLL) given predictions and targets."""
+
+        p_ctc, wav_lens, predicted_tokens = predictions
+
+        ids = batch.id
+        tokens_eos, tokens_eos_lens = batch.tokens_eos
+        tokens, tokens_lens = batch.tokens
+
+        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
+            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
+            tokens_eos_lens = torch.cat(
+                [tokens_eos_lens, tokens_eos_lens], dim=0
+            )
+            tokens = torch.cat([tokens, tokens], dim=0)
+            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)
+
+        loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+        loss = loss_ctc
+
+        if stage != sb.Stage.TRAIN:
+            # Decode token terms to words
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+        loss = self.compute_objectives(predictions, batch, sb.Stage.TRAIN)
+        loss.backward()
+        if self.check_gradients(loss):
+            self.wav2vec_optimizer.step()
+            self.model_optimizer.step()
+
+        self.wav2vec_optimizer.zero_grad()
+        self.model_optimizer.zero_grad()
+
+        return loss.detach()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.wav2vec_optimizer, new_lr_wav2vec
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+        self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+            self.modules.wav2vec2.parameters()
+        )
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable(
+                "wav2vec_opt", self.wav2vec_optimizer
+            )
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(sort_key="duration")
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration", reverse=True
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["train_dataloader_opts"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+
+    # test is separate
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav", "sr")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav, sr):
+        sig = sb.dataio.dataio.read_audio(wav)
+        sig = resamplers[sr](sig)
+        return sig
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
+        yield tokens_bos
+        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
+        yield tokens_eos
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "bos_label": hparams["bos_index"],
+        "eos_label": hparams["eos_index"],
+        "blank_label": hparams["blank_index"],
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets,
+        ["id", "sig", "wrd", "char_list", "tokens_bos", "tokens_eos", "tokens"],
+    )
+    return train_data, valid_data, test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # CLI:
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+
+    # If distributed_launch=True then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+
+    # Dataset prep (parsing Librispeech)
+
+    resampler_8000  = T.Resample(8000, 16000, dtype=torch.float)
+
+    resampler_44100 =T.Resample(44100, 16000, dtype=torch.float)
+    resampler_32000 =T.Resample(32000, 16000, dtype=torch.float)
+    resampler_48000 =T.Resample(48000, 16000, dtype=torch.float)
+
+
+    resamplers = {"48000": resampler_48000,"8000": resampler_8000, "44100":resampler_44100, "32000":resampler_32000}
+
+    # here we create the datasets objects as well as tokenization and encoding
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+        hparams
+    )
+
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+    asr_brain.device= "cpu"
+    asr_brain.modules.to("cpu")
+
+    # We dynamicaly add the tokenizer to our brain class.
+    # NB: This tokenizer corresponds to the one used for the LM!!
+    asr_brain.tokenizer = label_encoder
+
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["train_dataloader_opts"],
+        valid_loader_kwargs=hparams["valid_dataloader_opts"],
+    )
+
+    # Testing
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_opts"]
+        )

partly_frozen_splitted_wavlm/env.log

@@ -0,0 +1,379 @@
+SpeechBrain system description
+==============================
+Python version:
+3.8.5 (default, Sep  4 2020, 07:30:14) 
+[GCC 7.3.0]
+==============================
+Installed Python packages:
+abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altgraph==0.17
+antlr4-python3-runtime==4.8
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+greenlet==1.1.2
+grpcio==1.32.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pyctcdecode==0.4.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==3.7.4.3
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+==============================
+Could not get git revision==============================
+CUDA version:
+11.7

partly_frozen_splitted_wavlm/hyperparams.yaml

@@ -0,0 +1,162 @@
+# Generated 2023-01-07 from:
+# /home/salah/kenlm_train/to_copy/wavlm_partly_frozen.yaml
+# yamllint disable
+# ################################
+# Model: wav2vec2 + DNN + CTC
+# Augmentation: SpecAugment
+# Authors: Sung-Lin Yeh 2021
+# ################################
+
+# Seed needs to be set at top of yaml, before objects with parameters are made
+seed: 1986
+__set_seed: !apply:torch.manual_seed [1986]
+output_folder: partly_frozen_splitted_wavlm
+wer_file: partly_frozen_splitted_wavlm/wer.txt
+save_folder: partly_frozen_splitted_wavlm/save
+train_log: partly_frozen_splitted_wavlm/train_log.txt
+
+# URL for the biggest Fairseq english wav2vec2 model.
+
+# Data files
+data_folder: /gpfsscratch/rech/nou/uzn19yk/Libri/LibriSpeech/ # e,g./path/to/LibriSpeech
+# noise/ris dataset will automatically be downloaded
+data_folder_rirs: /gpfsscratch/rech/nou/uzn19yk/Libri/LibriSpeech/
+train_splits: [train-clean-100]
+dev_splits: [dev-clean]
+test_splits: [test-clean, test-other]
+skip_prep: false
+ckpt_interval_minutes: 25 # save checkpoint every N min
+csv_folder: /gpfsstore/rech/nou/uzn19yk/iwslt/splitted_clean_tunisian_csvs/
+train_csv: test_salah_local.csv
+valid_csv: test_salah_local.csv
+test_csv:
+- test_salah_local.csv
+
+# Training parameters
+number_of_epochs: 12
+lr: 1
+lr_wav2vec: 0.0001
+sorting: ascending
+auto_mix_prec: false
+sample_rate: 16000
+
+avoid_if_longer_than: 10
+# With data_parallel batch_size is split into N jobs
+# With DDP batch_size is multiplied by N jobs
+# Must be 3 per GPU to fit 32GB of VRAM
+batch_size: 1
+test_batch_size: 1
+
+# Dataloader options
+train_dataloader_opts:
+  batch_size: 1
+
+valid_dataloader_opts:
+  batch_size: 1
+
+test_dataloader_opts:
+  batch_size: 1
+
+# Model parameters
+activation: &id001 !name:torch.nn.LeakyReLU
+dnn_layers: 2
+dnn_neurons: 1024
+freeze_wav2vec: false
+
+# Outputs
+output_neurons: 41  # BPE size, index(blank/eos/bos) = 0
+
+# Decoding parameters
+blank_index: 0
+bos_index: 1
+eos_index: 2
+
+#
+# Functions and classes
+#
+epoch_counter: &id008 !new:speechbrain.utils.epoch_loop.EpochCounter
+
+  limit: 12
+
+augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
+  sample_rate: 16000
+  speeds: [95, 100, 105]
+
+enc: &id003 !new:speechbrain.lobes.models.VanillaNN.VanillaNN
+  input_shape: [null, null, 1024]
+  activation: *id001
+  dnn_blocks: 2
+  dnn_neurons: 1024
+
+wav2vec2: &id002 !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
+  source: wavlm-large/
+  output_norm: true
+  freeze: false
+  freeze_feature_extractor: true
+  save_path: partly_frozen_splitted_wavlm/save/wav2vec2_hubert_checkpoint
+
+#####
+# Uncomment this block if you prefer to use a Fairseq pretrained model instead
+# of a HuggingFace one. Here, we provide an URL that is obtained from the
+# Fairseq github for the multilingual XLSR.
+#
+#wav2vec2_url: https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_960h_pl.pt
+#wav2vec2: !new:speechbrain.lobes.models.fairseq_wav2vec.FairseqWav2Vec2
+#    pretrained_path: !ref <wav2vec2_url>
+#    output_norm: True
+#    freeze: False
+#    save_path: !ref <save_folder>/wav2vec2_checkpoint/model.pt
+
+ctc_lin: &id004 !new:speechbrain.nnet.linear.Linear
+
+  input_size: 1024
+  n_neurons: 41
+
+log_softmax: !new:speechbrain.nnet.activations.Softmax
+  apply_log: true
+
+ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
+  blank_index: 0
+
+modules:
+  wav2vec2: *id002
+  enc: *id003
+  ctc_lin: *id004
+model: &id005 !new:torch.nn.ModuleList
+- [*id003, *id004]
+model_opt_class: !name:torch.optim.Adadelta
+  lr: 1
+  rho: 0.95
+  eps: 1.e-8
+
+wav2vec_opt_class: !name:torch.optim.Adam
+  lr: 0.0001
+
+lr_annealing_model: &id006 !new:speechbrain.nnet.schedulers.NewBobScheduler
+  initial_value: 1
+  improvement_threshold: 0.0025
+  annealing_factor: 0.8
+  patient: 0
+
+lr_annealing_wav2vec: &id007 !new:speechbrain.nnet.schedulers.NewBobScheduler
+  initial_value: 0.0001
+  improvement_threshold: 0.0025
+  annealing_factor: 0.9
+  patient: 0
+
+
+checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
+  checkpoints_dir: partly_frozen_splitted_wavlm/save
+  recoverables:
+    wav2vec2: *id002
+    model: *id005
+    scheduler_model: *id006
+    scheduler_wav2vec: *id007
+    counter: *id008
+train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
+  save_file: partly_frozen_splitted_wavlm/train_log.txt
+
+error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+
+cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+  split_tokens: true

partly_frozen_splitted_wavlm/log.txt

@@ -0,0 +1,1998 @@
+2023-01-07 15:57:01,210 - speechbrain.core - INFO - Beginning experiment!
+2023-01-07 15:57:01,210 - speechbrain.core - INFO - Experiment folder: partly_frozen_splitted_wavlm
+2023-01-07 15:57:01,912 - speechbrain.utils.superpowers - DEBUG - abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altgraph==0.17
+antlr4-python3-runtime==4.8
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+greenlet==1.1.2
+grpcio==1.32.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pyctcdecode==0.4.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==3.7.4.3
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+
+
+2023-01-07 15:57:02,047 - speechbrain.core - ERROR - Exception:
+Traceback (most recent call last):
+  File "ctc_train.py", line 305, in <module>
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
+  File "ctc_train.py", line 178, in dataio_prepare
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/dataio/dataset.py", line 365, in from_csv
+    data = load_data_csv(csv_path, replacements)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/dataio/dataio.py", line 127, in load_data_csv
+    with open(csv_path, newline="") as csvfile:
+FileNotFoundError: [Errno 2] No such file or directory: 'test_salah.csv'
+2023-01-07 15:57:54,234 - speechbrain.core - INFO - Beginning experiment!
+2023-01-07 15:57:54,234 - speechbrain.core - INFO - Experiment folder: partly_frozen_splitted_wavlm
+2023-01-07 15:57:54,891 - speechbrain.utils.superpowers - DEBUG - abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altgraph==0.17
+antlr4-python3-runtime==4.8
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+greenlet==1.1.2
+grpcio==1.32.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pyctcdecode==0.4.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==3.7.4.3
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+
+
+2023-01-07 15:57:54,987 - speechbrain.dataio.encoder - DEBUG - Would load categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt, but file doesn't exist yet.
+2023-01-07 15:57:54,988 - speechbrain.dataio.encoder - INFO - Moving label 'ت' from index 0, because '<blank>' was put at its place.
+2023-01-07 15:57:54,988 - speechbrain.dataio.encoder - INFO - Moving label 'ع' from index 1, because '<bos>' was put at its place.
+2023-01-07 15:57:54,989 - speechbrain.dataio.encoder - INFO - Moving label 'ب' from index 2, because '<eos>' was put at its place.
+2023-01-07 15:57:54,989 - speechbrain.dataio.encoder - INFO - Load called, but CTCTextEncoder is not empty. Loaded data will overwrite everything. This is normal if there is e.g. an unk label defined at init.
+2023-01-07 15:57:54,990 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 15:57:54,990 - speechbrain.core - INFO - Info: auto_mix_prec arg from hparam file is used
+2023-01-07 15:57:54,990 - speechbrain.core - INFO - Info: ckpt_interval_minutes arg from hparam file is used
+2023-01-07 15:57:57,073 - speechbrain.core - INFO - 313.4M trainable parameters in ASR
+2023-01-07 15:57:57,075 - speechbrain.utils.checkpoints - INFO - Would load a checkpoint here, but none found yet.
+2023-01-07 15:57:57,075 - speechbrain.utils.epoch_loop - INFO - Going into epoch 1
+2023-01-07 15:57:57,132 - speechbrain.core - ERROR - Exception:
+Traceback (most recent call last):
+  File "ctc_train.py", line 322, in <module>
+    asr_brain.fit(
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1153, in fit
+    self._fit_train(train_set=train_set, epoch=epoch, enable=enable)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1009, in _fit_train
+    loss = self.fit_batch(batch)
+  File "ctc_train.py", line 88, in fit_batch
+    predictions = self.compute_forward(batch, sb.Stage.TRAIN)
+  File "ctc_train.py", line 42, in compute_forward
+    feats = self.modules.wav2vec2(wavs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
+    return forward_call(*input, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/lobes/models/huggingface_wav2vec.py", line 266, in forward
+    return self.extract_features(wav)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/lobes/models/huggingface_wav2vec.py", line 281, in extract_features
+    out = self.model(wav)[0]
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
+    return forward_call(*input, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/transformers/models/wavlm/modeling_wavlm.py", line 1232, in forward
+    extract_features = self.feature_extractor(input_values)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
+    return forward_call(*input, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/transformers/models/wavlm/modeling_wavlm.py", line 400, in forward
+    hidden_states = conv_layer(hidden_states)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
+    return forward_call(*input, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/transformers/models/wavlm/modeling_wavlm.py", line 270, in forward
+    hidden_states = self.conv(hidden_states)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
+    return forward_call(*input, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 313, in forward
+    return self._conv_forward(input, self.weight, self.bias)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 309, in _conv_forward
+    return F.conv1d(input, weight, bias, self.stride,
+RuntimeError: Expected 2D (unbatched) or 3D (batched) input to conv1d, but got input of size: [2, 1, 168960, 1]
+2023-01-07 15:59:22,733 - speechbrain.core - INFO - Beginning experiment!
+2023-01-07 15:59:22,733 - speechbrain.core - INFO - Experiment folder: partly_frozen_splitted_wavlm
+2023-01-07 15:59:23,373 - speechbrain.utils.superpowers - DEBUG - abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altgraph==0.17
+antlr4-python3-runtime==4.8
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+greenlet==1.1.2
+grpcio==1.32.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pyctcdecode==0.4.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==3.7.4.3
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+
+
+2023-01-07 15:59:23,493 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 15:59:23,493 - speechbrain.dataio.encoder - INFO - Load called, but CTCTextEncoder is not empty. Loaded data will overwrite everything. This is normal if there is e.g. an unk label defined at init.
+2023-01-07 15:59:23,494 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 15:59:23,495 - speechbrain.core - INFO - Info: auto_mix_prec arg from hparam file is used
+2023-01-07 15:59:23,495 - speechbrain.core - INFO - Info: ckpt_interval_minutes arg from hparam file is used
+2023-01-07 15:59:24,946 - speechbrain.core - INFO - 313.4M trainable parameters in ASR
+2023-01-07 15:59:24,949 - speechbrain.utils.checkpoints - INFO - Would load a checkpoint here, but none found yet.
+2023-01-07 15:59:24,949 - speechbrain.utils.epoch_loop - INFO - Going into epoch 1
+2023-01-07 15:59:27,528 - speechbrain.core - ERROR - Exception:
+Traceback (most recent call last):
+  File "ctc_train.py", line 322, in <module>
+    asr_brain.fit(
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1153, in fit
+    self._fit_train(train_set=train_set, epoch=epoch, enable=enable)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1009, in _fit_train
+    loss = self.fit_batch(batch)
+  File "ctc_train.py", line 92, in fit_batch
+    self.wav2vec_optimizer.step()
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/optimizer.py", line 140, in wrapper
+    out = func(*args, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/optimizer.py", line 23, in _use_grad
+    ret = func(self, *args, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/adam.py", line 220, in step
+    state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 5.80 GiB total capacity; 3.82 GiB already allocated; 52.50 MiB free; 3.94 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
+2023-01-07 15:59:54,237 - speechbrain.core - INFO - Beginning experiment!
+2023-01-07 15:59:54,237 - speechbrain.core - INFO - Experiment folder: partly_frozen_splitted_wavlm
+2023-01-07 15:59:54,868 - speechbrain.utils.superpowers - DEBUG - abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altgraph==0.17
+antlr4-python3-runtime==4.8
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+greenlet==1.1.2
+grpcio==1.32.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pyctcdecode==0.4.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==3.7.4.3
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+
+
+2023-01-07 15:59:54,960 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 15:59:54,960 - speechbrain.dataio.encoder - INFO - Load called, but CTCTextEncoder is not empty. Loaded data will overwrite everything. This is normal if there is e.g. an unk label defined at init.
+2023-01-07 15:59:54,961 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 15:59:54,961 - speechbrain.core - INFO - Info: auto_mix_prec arg from hparam file is used
+2023-01-07 15:59:54,961 - speechbrain.core - INFO - Info: ckpt_interval_minutes arg from hparam file is used
+2023-01-07 15:59:56,396 - speechbrain.core - INFO - 313.4M trainable parameters in ASR
+2023-01-07 15:59:56,398 - speechbrain.utils.checkpoints - INFO - Would load a checkpoint here, but none found yet.
+2023-01-07 15:59:56,398 - speechbrain.utils.epoch_loop - INFO - Going into epoch 1
+2023-01-07 15:59:57,755 - speechbrain.core - ERROR - Exception:
+Traceback (most recent call last):
+  File "ctc_train.py", line 322, in <module>
+    asr_brain.fit(
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1153, in fit
+    self._fit_train(train_set=train_set, epoch=epoch, enable=enable)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1009, in _fit_train
+    loss = self.fit_batch(batch)
+  File "ctc_train.py", line 92, in fit_batch
+    self.wav2vec_optimizer.step()
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/optimizer.py", line 140, in wrapper
+    out = func(*args, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/optimizer.py", line 23, in _use_grad
+    ret = func(self, *args, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/adam.py", line 218, in step
+    state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 5.80 GiB total capacity; 3.86 GiB already allocated; 64.50 MiB free; 3.93 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
+2023-01-07 16:00:47,395 - speechbrain.core - INFO - Beginning experiment!
+2023-01-07 16:00:47,395 - speechbrain.core - INFO - Experiment folder: partly_frozen_splitted_wavlm
+2023-01-07 16:00:48,055 - speechbrain.utils.superpowers - DEBUG - abkhazia==1.0
+absl-py==0.11.0
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altgraph==0.17
+antlr4-python3-runtime==4.8
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+asgiref==3.6.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+boto3==1.20.2
+botocore==1.23.2
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+cryptography==38.0.4
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+datasets==1.5.0
+decorator==4.4.2
+deepspeech==0.9.1
+defusedxml==0.7.1
+denoiser==0.1.5
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+filelock==3.0.12
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.2.0
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+google-pasta==0.2.0
+greenlet==1.1.2
+grpcio==1.32.0
+h5features==1.3.2
+h5py==2.10.0
+htk-io==0.5
+huggingface-hub==0.9.1
+hydra-colorlog==0.1.4
+hydra-core==0.11.3
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+ipadic==1.0.0
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+jedi==0.17.2
+jieba==0.42.1
+Jinja2==2.11.2
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+MarkupSafe==1.1.1
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==6.0.7
+nbformat==5.1.3
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.5
+nltk==3.5
+nodeenv==1.5.0
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+oauthlib==3.1.0
+omegaconf==1.4.1
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathspec==0.8.1
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pre-commit==2.9.0
+pretty-midi==0.2.9
+prettytable==2.2.1
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+pyannote.audio==1.1.1
+pyannote.core==4.3
+pyannote.database==4.1.1
+pyannote.metrics==3.1
+pyannote.pipeline==1.5.2
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pyctcdecode==0.4.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.7.2
+pygtrie==2.5.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-Levenshtein==0.12.2
+python-utils==2.4.0
+pytorch-lightning==1.4.9
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==5.3.1
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+segments==2.1.3
+Send2Trash==1.5.0
+sentencepiece==0.1.94
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.13
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlparse==0.4.2
+stanza==1.4.2
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tkseem==0.0.3
+tokenizers==0.10.2
+toml==0.10.2
+torch==1.13.1
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.6.0
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.0.5
+transformers==4.15.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==3.7.4.3
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+
+
+2023-01-07 16:00:48,174 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 16:00:48,174 - speechbrain.dataio.encoder - INFO - Load called, but CTCTextEncoder is not empty. Loaded data will overwrite everything. This is normal if there is e.g. an unk label defined at init.
+2023-01-07 16:00:48,175 - speechbrain.dataio.encoder - DEBUG - Loaded categorical encoding from partly_frozen_splitted_wavlm/save/label_encoder.txt
+2023-01-07 16:00:48,176 - speechbrain.core - INFO - Info: auto_mix_prec arg from hparam file is used
+2023-01-07 16:00:48,176 - speechbrain.core - INFO - Info: ckpt_interval_minutes arg from hparam file is used
+2023-01-07 16:00:49,674 - speechbrain.core - INFO - 313.4M trainable parameters in ASR
+2023-01-07 16:00:50,298 - speechbrain.utils.checkpoints - INFO - Would load a checkpoint here, but none found yet.
+2023-01-07 16:00:50,298 - speechbrain.utils.epoch_loop - INFO - Going into epoch 1
+2023-01-07 16:01:08,641 - speechbrain.core - ERROR - Exception:
+Traceback (most recent call last):
+  File "ctc_train.py", line 324, in <module>
+    asr_brain.fit(
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1153, in fit
+    self._fit_train(train_set=train_set, epoch=epoch, enable=enable)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/speechbrain/core.py", line 1009, in _fit_train
+    loss = self.fit_batch(batch)
+  File "ctc_train.py", line 92, in fit_batch
+    self.wav2vec_optimizer.step()
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/optimizer.py", line 140, in wrapper
+    out = func(*args, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/optimizer.py", line 23, in _use_grad
+    ret = func(self, *args, **kwargs)
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/adam.py", line 234, in step
+    adam(params_with_grad,
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/adam.py", line 300, in adam
+    func(params,
+  File "/home/salah/anaconda3/envs/salah/lib/python3.8/site-packages/torch/optim/adam.py", line 410, in _single_tensor_adam
+    denom = (exp_avg_sq.sqrt() / bias_correction2_sqrt).add_(eps)
+KeyboardInterrupt

partly_frozen_splitted_wavlm/save/label_encoder.txt

@@ -0,0 +1,37 @@
+'ت' => 29
+'ع' => 30
+'ب' => 31
+' ' => 3
+'ه' => 4
+'ا' => 5
+'ن' => 6
+'ي' => 7
+'ر' => 8
+'ك' => 9
+'ش' => 10
+'ف' => 11
+'ل' => 12
+'د' => 13
+'س' => 14
+'م' => 15
+'ق' => 16
+'ى' => 17
+'ء' => 18
+'و' => 19
+'ح' => 20
+'ز' => 21
+'ة' => 22
+'أ' => 23
+'خ' => 24
+'ص' => 25
+'ط' => 26
+'ج' => 27
+'ظ' => 28
+'<blank>' => 0
+'<bos>' => 1
+'<eos>' => 2
+================
+'starting_index' => 0
+'bos_label' => '<bos>'
+'eos_label' => '<eos>'
+'blank_label' => '<blank>'

requirements.txt

@@ -0,0 +1,16 @@
+-r lint-requirements.txt
+huggingface_hub>=0.7.0
+hyperpyyaml>=0.0.1
+joblib>=0.14.1
+numpy>=1.17.0
+packaging
+pre-commit>=2.3.0
+scipy>=1.4.1, <1.9
+sentencepiece>=0.1.91
+SoundFile; sys_platform == 'win32'
+torch>=1.9.0
+torchaudio>=0.9.0
+tqdm>=4.42.0
+transformers==4.15
+speechbrain
+pyctcdecode