end-2-end reborn model for mls-italian unsupervised phoneme recognition (iter2-stage1)

config.json

@@ -0,0 +1,79 @@
+{
+  "architectures": [
+    "RebornUASRModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_reborn.RebornUASRConfig",
+    "AutoModel": "modeling_reborn.RebornUASRModel"
+  },
+  "discriminator_act_after_linear": false,
+  "discriminator_causal": true,
+  "discriminator_depth": 1,
+  "discriminator_dilation": 1,
+  "discriminator_dim": 256,
+  "discriminator_dropout": 0.0,
+  "discriminator_input_dim": 42,
+  "discriminator_kernel": 3,
+  "discriminator_linear_emb": false,
+  "discriminator_max_pool": false,
+  "discriminator_spectral_norm": false,
+  "discriminator_weight_norm": false,
+  "generator_bias": false,
+  "generator_bn_apply": false,
+  "generator_bn_init_weight": 30.0,
+  "generator_dilation": 1,
+  "generator_dropout": 0.0,
+  "generator_input_dim": 512,
+  "generator_kernel": 4,
+  "generator_output_dim": 42,
+  "generator_stride": 1,
+  "model_type": "reborn_uasr",
+  "phones": [
+    "a",
+    "e",
+    "i",
+    "o",
+    "r",
+    "n",
+    "t",
+    "l",
+    "k",
+    "s",
+    "m",
+    "d",
+    "\u02d0",
+    "\u026a",
+    "p",
+    "v",
+    "\u028a",
+    "\u025b",
+    "b",
+    "f",
+    "z",
+    "t\u0283",
+    "j",
+    "\u0261",
+    "\u0254",
+    "d\u0292",
+    "ss",
+    "ts",
+    "u",
+    "\u027e",
+    "w",
+    "\u028e",
+    "\u0272",
+    "\u0283",
+    "\u014b",
+    "dz",
+    "a\u026a",
+    "<SIL>"
+  ],
+  "segmenter_dropout": 0.1,
+  "segmenter_hidden_dim": 512,
+  "segmenter_input_dim": 512,
+  "segmenter_kernel_size": 7,
+  "segmenter_type": "cnn",
+  "special_token_nums": 4,
+  "torch_dtype": "float32",
+  "transformers_version": "4.24.0"
+}

configuration_reborn.py

@@ -0,0 +1,105 @@
+import os
+import json
+from transformers import PretrainedConfig
+
+class RebornUASRConfig(PretrainedConfig):
+    '''
+    We can use this class to define the configuration of the reborn model. 
+    The reborn UASR is composed of a segmenter, a discriminator, and a generator.
+    We only include the required configurations for the discriminator and the generator from fairseq's wav2vec-U model configuration. 
+    '''
+    model_type = "reborn_uasr"
+    
+    def __init__(self, 
+        segmenter_type: str = "cnn",
+        segmenter_input_dim: int = 512,
+        segmenter_hidden_dim: int = 512,
+        segmenter_dropout: float = 0.1,
+        segmenter_kernel_size: int = 7,
+
+        discriminator_input_dim: int = 512,
+        discriminator_kernel: int = 3,
+        discriminator_dilation: int = 1,
+        discriminator_dim: int = 256,
+        discriminator_causal: bool = True,
+        discriminator_linear_emb: bool = False,
+        discriminator_depth: int = 1,
+        discriminator_max_pool: bool = False,
+        discriminator_act_after_linear: bool = False,
+        discriminator_dropout: float = 0.0,
+        discriminator_spectral_norm: bool = False,
+        discriminator_weight_norm: bool = False,
+
+        generator_input_dim: int = 512,
+        generator_output_dim: int = 40,
+        generator_kernel: int = 4,
+        generator_dilation: int = 1,
+        generator_stride: int = 1,
+        generator_bias: bool = False,
+        generator_dropout: float = 0.0,
+        generator_bn_apply: bool = False,
+        generator_bn_init_weight: float = 30.0,
+
+        phones: list = [],
+        dict_fpath: str = "",
+        special_token_nums: int = 4, # [<s>, <pad>, </s>, <unk>]
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        # read in all the configurations
+        self.segmenter_type = segmenter_type
+        self.segmenter_input_dim = segmenter_input_dim
+        self.segmenter_hidden_dim = segmenter_hidden_dim
+        self.segmenter_dropout = segmenter_dropout
+        self.segmenter_kernel_size = segmenter_kernel_size
+
+        self.discriminator_input_dim = discriminator_input_dim
+        self.discriminator_kernel = discriminator_kernel
+        self.discriminator_dilation = discriminator_dilation
+        self.discriminator_dim = discriminator_dim
+        self.discriminator_causal = discriminator_causal
+        self.discriminator_linear_emb = discriminator_linear_emb
+        self.discriminator_depth = discriminator_depth
+        self.discriminator_max_pool = discriminator_max_pool
+        self.discriminator_act_after_linear = discriminator_act_after_linear
+        self.discriminator_dropout = discriminator_dropout
+        self.discriminator_spectral_norm = discriminator_spectral_norm
+        self.discriminator_weight_norm = discriminator_weight_norm
+
+        self.generator_input_dim = generator_input_dim
+        self.generator_output_dim = generator_output_dim
+        self.generator_kernel = generator_kernel
+        self.generator_dilation = generator_dilation
+        self.generator_stride = generator_stride
+        self.generator_bias = generator_bias
+        self.generator_dropout = generator_dropout
+        self.generator_bn_apply = generator_bn_apply
+        self.generator_bn_init_weight = generator_bn_init_weight
+
+        self.special_token_nums = special_token_nums
+        if os.path.isfile(dict_fpath):
+            self.phones = self.read_phns_dict_from_fpath(dict_fpath)
+        else:
+            self.phones = phones
+        if len(self.phones) > 0:
+            self.generator_output_dim = len(self.phones) + self.special_token_nums
+            self.discriminator_input_dim = self.generator_output_dim
+    
+    def read_phns_dict_from_fpath(self, fpath: str):
+        phns = []
+        with open(fpath, "r", encoding="utf-8") as f:
+            for l in f:
+                phn = l.strip().split('\t')[0].split(' ')[0]
+                phns.append(phn)
+        return phns
+
+def main():
+    config = RebornUASRConfig(dict_fpath="/home/andybi7676/Desktop/uasr-rl/data2/it_mls/text/prep_sep/phones/dict.phn.txt")
+    print(config)
+    output_fpath = "./reborn_uasr_configs/config_mls-it.json"
+    with open(output_fpath, 'w', encoding='utf-8') as fw:
+        config_json_string = json.dumps(config.to_dict(), indent=2, sort_keys=True) + "\n"
+        fw.write(config_json_string)
+
+if __name__ == "__main__":
+    main()

modeling_reborn.py

@@ -0,0 +1,381 @@
+import torch
+import torch.nn as nn
+from transformers import PreTrainedModel
+from .configuration_reborn import RebornUASRConfig
+from typing import Optional, Tuple, Union, List
+
+class RebornSegmenter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.conv1 = nn.Conv1d(config.segmenter_input_dim, config.segmenter_hidden_dim, config.segmenter_kernel_size, padding=config.segmenter_kernel_size//2)
+        self.conv2 = nn.Conv1d(config.segmenter_hidden_dim, config.segmenter_hidden_dim, 3, padding=1)
+        self.conv3 = nn.Conv1d(config.segmenter_hidden_dim, 2, 1)
+        self.dropout = nn.Dropout(config.segmenter_dropout)
+        self.relu = nn.ReLU()
+
+    def forward(self, x):
+        """
+        Input:
+            x: (B, T, C)
+            padding_mask: (B, T) # 0: not padding; 1: padding
+        Output:
+            boundary: (B, T, 2) # 0: not boundary; 1: boundary
+        """
+        x = x.transpose(1, 2)
+        x = self.dropout(self.relu(self.conv1(x)))
+        x = self.dropout(self.relu(self.conv2(x)))
+        x = self.conv3(x)
+        x = x.transpose(1, 2)
+        return x
+    
+    def boundary_predict(self, x, padding_mask, deterministic=False):
+        """
+        Input:
+            x: (B, T, C)
+            padding_mask: (B, T)
+        Output:
+            boundary: (B, T) # 0: not boundary; 1: boundary
+            boundary_logits: (B, T, 2) # 0: not boundary; 1: boundary
+        """
+        boundary_logits = self.forward(x)
+        if deterministic:
+            boundary = boundary_logits.argmax(-1)
+            boundary[padding_mask] = -1
+        else:
+            boundary = torch.distributions.Categorical(logits=boundary_logits).sample()
+            boundary[padding_mask] = -1
+        return boundary, boundary_logits
+    
+    def pre_segment(self, logits, padding_mask, return_boundary=False, deterministic=True):
+        """
+        Input:
+            logits: (B, T, C)
+            padding_mask: (B, T)
+        Output:
+            new_logits: (B, T', C)
+            new_padding_mask: (B, T')
+        """
+        
+        bsz, tsz, csz = logits.size()
+        
+        boundary, boundary_logits = self.boundary_predict(logits, padding_mask, deterministic=deterministic)
+        
+        # max boundary number
+        # print("boundary", boundary)
+        # print(torch.sum(boundary==1, dim=1))
+        new_tsz = int(torch.max(torch.sum(boundary==1, dim=1)).item())+1 # add <bos>
+        new_logits = logits.new_zeros(bsz, new_tsz, csz)
+        new_pad = padding_mask.new_zeros(bsz, new_tsz)
+        
+        for b in range(bsz):
+            # merge consecutive segments when meeting a boundary (mean_pool_join)
+            new_idx = 0
+            count = 0
+            for t in range(tsz):
+                if padding_mask[b, t] == 1:
+                    break
+                if boundary[b, t] == 1:
+                    new_logits[b, new_idx] /= count
+                    new_idx += 1
+                    count = 0
+                new_logits[b, new_idx] += logits[b, t]
+                count += 1
+            if count > 0:
+                # last segment
+                new_logits[b, new_idx] /= count
+                new_idx += 1
+                count = 0
+            if new_idx < new_tsz:
+                pad = new_tsz - new_idx
+                new_logits[b, -pad:] = 0
+                new_pad[b, -pad:] = True
+
+        if return_boundary:
+            return new_logits, new_pad, boundary, boundary_logits
+        return new_logits, new_pad
+
+class RebornGenerator(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+        self.output_dim = config.generator_output_dim
+        self.stride = config.generator_stride
+        self.dropout = nn.Dropout(config.generator_dropout)
+        cnn_input_dim = config.generator_input_dim
+        cnn_output_dim = config.generator_output_dim
+
+        padding = config.generator_kernel // 2
+        self.proj = nn.Sequential(
+            nn.Conv1d(
+                cnn_input_dim,
+                cnn_output_dim,
+                kernel_size=config.generator_kernel,
+                stride=config.generator_stride,
+                dilation=config.generator_dilation,
+                padding=padding,
+                bias=config.generator_bias,
+            ),
+        )
+
+    def forward(self, dense_x, tokens, dense_padding_mask):
+        dense_x = self.dropout(dense_x)
+        # (B, T, C) -> (B, C, T)
+        dense_x = dense_x.transpose(-2, -1)
+
+        dense_x = self.proj(dense_x)
+        # (B, C, T) -> (B, T, C)
+        dense_x = dense_x.transpose(-2, -1)
+        if self.stride > 1:
+            dense_padding_mask = dense_padding_mask[:, :: self.stride]
+
+        if dense_padding_mask.size(1) != dense_x.size(1):
+            new_padding = dense_padding_mask.new_zeros(dense_x.shape[:-1])
+            diff = new_padding.size(1) - dense_padding_mask.size(1)
+            assert (
+                diff > 0
+            ), f"{new_padding.shape}, {dense_padding_mask.shape}, {dense_x.shape}, {diff}"
+            if diff > 0:
+                new_padding[:, diff:] = dense_padding_mask
+            else:
+                assert diff < 0
+                new_padding = dense_padding_mask[:, :diff]
+
+            dense_padding_mask = new_padding
+
+        result = {}
+
+        token_x = None
+        if tokens is not None:
+            token_x = dense_x.new_zeros(tokens.numel(), self.output_dim)
+            token_x.scatter_(1, tokens.view(-1, 1).long(), 1)
+            token_x = token_x.view(tokens.shape + (self.output_dim,))
+
+        result["dense_x"] = dense_x
+        result["token_x"] = token_x
+        result["dense_padding_mask"] = dense_padding_mask
+
+        return result
+
+def get_item(tensor):
+    # tpu-comment: making this a no-op for xla devices.
+    if torch.is_tensor(tensor) and tensor.device.type == "xla":
+        return tensor.detach()
+    if hasattr(tensor, "item"):
+        return tensor.item()
+    if hasattr(tensor, "__getitem__"):
+        return tensor[0]
+    return tensor
+
+def post_process(sentence: str, symbol: str):
+    if symbol == "sentencepiece":
+        sentence = sentence.replace(" ", "").replace("\u2581", " ").strip()
+    elif symbol == "wordpiece":
+        sentence = sentence.replace(" ", "").replace("_", " ").strip()
+    elif symbol == "letter":
+        sentence = sentence.replace(" ", "").replace("|", " ").strip()
+    elif symbol == "silence":
+        import re
+        sentence = sentence.replace("<SIL>", "")
+        sentence = re.sub(' +', ' ', sentence).strip()
+    elif symbol == "_EOW":
+        sentence = sentence.replace(" ", "").replace("_EOW", " ").strip()
+    elif symbol in {"subword_nmt", "@@ ", "@@"}:
+        if symbol == "subword_nmt":
+            symbol = "@@ "
+        sentence = (sentence + " ").replace(symbol, "").rstrip()
+    elif symbol == "none":
+        pass
+    elif symbol is not None:
+        raise NotImplementedError(f"Unknown post_process option: {symbol}")
+    return sentence
+
+class SimpleTokenizer(object):
+    def __init__(self,
+        phones: List[str],
+        bos="<s>",
+        pad="<pad>",
+        eos="</s>",
+        unk="<unk>",
+        extra_special_symbols=None,
+    ) -> None:
+        self.bos_word, self.unk_word, self.pad_word, self.eos_word = bos, unk, pad, eos
+        self.symbols = []
+        self.count = []
+        self.indices = {}
+        self.bos_index = self.add_symbol(bos)
+        self.pad_index = self.add_symbol(pad)
+        self.eos_index = self.add_symbol(eos)
+        self.unk_index = self.add_symbol(unk)
+        if extra_special_symbols:
+            for s in extra_special_symbols:
+                self.add_symbol(s)
+        self.nspecial = len(self.symbols)
+        for phone in phones:
+            self.add_symbol(phone)
+        self.postprocess_code = "silence"
+    
+    def add_symbol(self, word, n=1, overwrite=False):
+        """Adds a word to the dictionary"""
+        if word in self.indices and not overwrite:
+            idx = self.indices[word]
+            self.count[idx] = self.count[idx] + n
+            return idx
+        else:
+            idx = len(self.symbols)
+            self.indices[word] = idx
+            self.symbols.append(word)
+            self.count.append(n)
+            return idx
+
+    def __eq__(self, other):
+        return self.indices == other.indices
+
+    def __getitem__(self, idx):
+        if idx < len(self.symbols):
+            return self.symbols[idx]
+        return self.unk_word
+
+    def get_count(self, idx):
+        return self.count[idx]
+
+    def __len__(self):
+        """Returns the number of symbols in the dictionary"""
+        return len(self.symbols)
+
+    def __contains__(self, sym):
+        return sym in self.indices
+
+    def index(self, sym):
+        """Returns the index of the specified symbol"""
+        assert isinstance(sym, str)
+        if sym in self.indices:
+            return self.indices[sym]
+        return self.unk_index
+
+    def string(
+        self,
+        tensor,
+        bpe_symbol=None,
+        escape_unk=False,
+        extra_symbols_to_ignore=None,
+        unk_string=None,
+        include_eos=False,
+        separator=" ",
+    ):
+        """Helper for converting a tensor of token indices to a string.
+
+        Can optionally remove BPE symbols or escape <unk> words.
+        """
+        if torch.is_tensor(tensor) and tensor.dim() == 2:
+            return "\n".join(
+                self.string(
+                    t,
+                    bpe_symbol,
+                    escape_unk,
+                    extra_symbols_to_ignore,
+                    include_eos=include_eos,
+                )
+                for t in tensor
+            )
+
+        extra_symbols_to_ignore = set(extra_symbols_to_ignore or [])
+        if not include_eos:
+            extra_symbols_to_ignore.add(self.eos())
+
+        def token_string(i):
+            if i == self.unk():
+                if unk_string is not None:
+                    return unk_string
+                else:
+                    return self.unk_string(escape_unk)
+            else:
+                return self[i]
+
+        if hasattr(self, "bos_index"):
+            extra_symbols_to_ignore.add(self.bos())
+
+        sent = separator.join(
+            token_string(i)
+            for i in tensor
+            if get_item(i) not in extra_symbols_to_ignore
+        )
+
+        return post_process(sent, bpe_symbol)
+
+    def unk_string(self, escape=False):
+        """Return unknown string, optionally escaped as: <<unk>>"""
+        if escape:
+            return "<{}>".format(self.unk_word)
+        else:
+            return self.unk_word
+    
+    def bos(self):
+        """Helper to get index of beginning-of-sentence symbol"""
+        return self.bos_index
+
+    def pad(self):
+        """Helper to get index of pad symbol"""
+        return self.pad_index
+
+    def eos(self):
+        """Helper to get index of end-of-sentence symbol"""
+        return self.eos_index
+
+    def unk(self):
+        """Helper to get index of unk symbol"""
+        return self.unk_index
+
+
+class RebornUASRModel(PreTrainedModel):
+    config_class = RebornUASRConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.pca = nn.Linear(1024, 512)
+        self.segmenter = RebornSegmenter(config)
+        self.generator = RebornGenerator(config)
+        self.tokenizer = None
+        if len(config.phones) > 0:
+            self.tokenizer = SimpleTokenizer(config.phones)
+
+    def forward(
+        self,
+        x: Optional[torch.Tensor], # (B, T, C)
+        padding_mask: Optional[torch.Tensor], # (B, T)
+    ):
+        x_reduced = self.pca(x)
+        x_segmented, segmented_padding_mask = self.segmenter.pre_segment(x_reduced, padding_mask, deterministic=True)
+        x_generated = self.generator(x_segmented, None, segmented_padding_mask)
+
+        return {
+            'x_reduced': x_reduced,
+            'x_segmented': x_segmented,
+            'x_generated': x_generated
+        }
+    
+    def generate(self, x, padding_mask, merge_consecutive=True, remove_silence=True):
+        res = self.forward(x, padding_mask)
+        y_raw_logits = res['x_generated']['dense_x']
+        y_raw_padding = res['x_generated']['dense_padding_mask']
+        y_raw_logits[y_raw_padding][..., self.tokenizer.pad_index] = float('inf')
+        preds = y_raw_logits.argmax(-1)
+        hyps = []
+        postprocess_code = "silence" if remove_silence else "none"
+        for pred in preds:
+            if merge_consecutive:
+                # merge consecutive predictions
+                pred = torch.unique_consecutive(pred)
+            hyp = self.tokenizer.string(pred, bpe_symbol=postprocess_code)
+            hyps.append(hyp)
+        return hyps
+
+def main():
+    model_config = RebornUASRConfig.from_pretrained("/home/andybi7676/Desktop/uasr-rl/reborn_uasr/config.json")
+    print(model_config)
+    model = RebornUASRModel(model_config)
+    print(model.tokenizer.indices)
+
+if __name__ == "__main__":
+    main()

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f17a8d31d2feaa20625ff018b82a15732abb16df27b4d8f1158c4db7a6c871d
+size 12940237