File size: 9,000 Bytes
2e6f087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch as T\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "import torchaudio\n",
    "from utils import load_ckpt, print_colored\n",
    "from tokenizer import make_tokenizer\n",
    "from model import get_hertz_dev_config\n",
    "import matplotlib.pyplot as plt\n",
    "from IPython.display import Audio, display\n",
    "\n",
    "\n",
    "# If you get an error like \"undefined symbol: __nvJitLinkComplete_12_4, version libnvJitLink.so.12\",\n",
    "# you need to install PyTorch with the correct CUDA version. Run:\n",
    "# `pip3 uninstall torch torchaudio && pip3 install torch torchaudio --index-url https://download.pytorch.org/whl/cu121`\n",
    "\n",
    "device = 'cuda' if T.cuda.is_available() else 'cpu'\n",
    "T.cuda.set_device(0)\n",
    "print_colored(f\"Using device: {device}\", \"grey\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# This code will automatically download them if it can't find them.\n",
    "audio_tokenizer = make_tokenizer(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "# We have different checkpoints for the single-speaker and two-speaker models\n",
    "# Set to True to load and run inference with the two-speaker model\n",
    "TWO_SPEAKER = False\n",
    "USE_PURE_AUDIO_ABLATION = False # We trained a base model with no text initialization at all. Toggle this to enable it.\n",
    "assert not (USE_PURE_AUDIO_ABLATION and TWO_SPEAKER) # We only have a single-speaker version of this model.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_config = get_hertz_dev_config(is_split=TWO_SPEAKER, use_pure_audio_ablation=USE_PURE_AUDIO_ABLATION)\n",
    "\n",
    "generator = model_config()\n",
    "generator = generator.eval().to(T.bfloat16).to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def load_and_preprocess_audio(audio_path):\n",
    "    print_colored(\"Loading and preprocessing audio...\", \"blue\", bold=True)\n",
    "    # Load audio file\n",
    "    audio_tensor, sr = torchaudio.load(audio_path)\n",
    "    print_colored(f\"Loaded audio shape: {audio_tensor.shape}\", \"grey\")\n",
    "    \n",
    "    if TWO_SPEAKER:\n",
    "        if audio_tensor.shape[0] == 1:\n",
    "            print_colored(\"Converting mono to stereo...\", \"grey\")\n",
    "            audio_tensor = audio_tensor.repeat(2, 1)\n",
    "            print_colored(f\"Stereo audio shape: {audio_tensor.shape}\", \"grey\")\n",
    "    else:\n",
    "        if audio_tensor.shape[0] == 2:\n",
    "            print_colored(\"Converting stereo to mono...\", \"grey\")\n",
    "            audio_tensor = audio_tensor.mean(dim=0).unsqueeze(0)\n",
    "            print_colored(f\"Mono audio shape: {audio_tensor.shape}\", \"grey\")\n",
    "        \n",
    "    # Resample to 16kHz if needed\n",
    "    if sr != 16000:\n",
    "        print_colored(f\"Resampling from {sr}Hz to 16000Hz...\", \"grey\")\n",
    "        resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)\n",
    "        audio_tensor = resampler(audio_tensor)\n",
    "        \n",
    "    # Clip to 5 minutes if needed\n",
    "    max_samples = 16000 * 60 * 5\n",
    "    if audio_tensor.shape[1] > max_samples:\n",
    "        print_colored(\"Clipping audio to 5 minutes...\", \"grey\")\n",
    "        audio_tensor = audio_tensor[:, :max_samples]\n",
    "\n",
    "    \n",
    "    print_colored(\"Audio preprocessing complete!\", \"green\")\n",
    "    return audio_tensor.unsqueeze(0)\n",
    "\n",
    "def display_audio(audio_tensor):\n",
    "    audio_tensor = audio_tensor.cpu().squeeze()\n",
    "    if audio_tensor.ndim == 1:\n",
    "        audio_tensor = audio_tensor.unsqueeze(0)\n",
    "    audio_tensor = audio_tensor.float()\n",
    "\n",
    "    # Make a waveform plot\n",
    "    plt.figure(figsize=(4, 1))\n",
    "    plt.plot(audio_tensor.numpy()[0], linewidth=0.5)\n",
    "    plt.axis('off')\n",
    "    plt.show()\n",
    "\n",
    "    # Make an audio player\n",
    "    display(Audio(audio_tensor.numpy(), rate=16000))\n",
    "    print_colored(f\"Audio ready for playback ↑\", \"green\", bold=True)\n",
    "    \n",
    "    \n",
    "\n",
    "# Our model is very prompt-sensitive, so we recommend experimenting with a diverse set of prompts.\n",
    "prompt_audio = load_and_preprocess_audio('./prompts/toaskanymore.wav')\n",
    "display_audio(prompt_audio)\n",
    "prompt_len_seconds = 3\n",
    "prompt_len = prompt_len_seconds * 8"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print_colored(\"Encoding prompt...\", \"blue\")\n",
    "with T.autocast(device_type='cuda', dtype=T.bfloat16):\n",
    "    if TWO_SPEAKER:\n",
    "        encoded_prompt_audio_ch1 = audio_tokenizer.latent_from_data(prompt_audio[:, 0:1].to(device))\n",
    "        encoded_prompt_audio_ch2 = audio_tokenizer.latent_from_data(prompt_audio[:, 1:2].to(device))\n",
    "        encoded_prompt_audio = T.cat([encoded_prompt_audio_ch1, encoded_prompt_audio_ch2], dim=-1)\n",
    "    else:\n",
    "        encoded_prompt_audio = audio_tokenizer.latent_from_data(prompt_audio.to(device))\n",
    "print_colored(f\"Encoded prompt shape: {encoded_prompt_audio.shape}\", \"grey\")\n",
    "print_colored(\"Prompt encoded successfully!\", \"green\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_completion(encoded_prompt_audio, prompt_len, gen_len=None):\n",
    "    prompt_len_seconds = prompt_len / 8\n",
    "    print_colored(f\"Prompt length: {prompt_len_seconds:.2f}s\", \"grey\")\n",
    "    print_colored(\"Completing audio...\", \"blue\")\n",
    "    encoded_prompt_audio = encoded_prompt_audio[:, :prompt_len]\n",
    "    with T.autocast(device_type='cuda', dtype=T.bfloat16):\n",
    "        completed_audio_batch = generator.completion(\n",
    "            encoded_prompt_audio, \n",
    "            temps=(.8, (0.5, 0.1)), # (token_temp, (categorical_temp, gaussian_temp))\n",
    "            use_cache=True,\n",
    "            gen_len=gen_len)\n",
    "\n",
    "        completed_audio = completed_audio_batch\n",
    "        print_colored(f\"Decoding completion...\", \"blue\")\n",
    "        if TWO_SPEAKER:\n",
    "            decoded_completion_ch1 = audio_tokenizer.data_from_latent(completed_audio[:, :, :32].bfloat16())\n",
    "            decoded_completion_ch2 = audio_tokenizer.data_from_latent(completed_audio[:, :, 32:].bfloat16())\n",
    "            decoded_completion = T.cat([decoded_completion_ch1, decoded_completion_ch2], dim=0)\n",
    "        else:\n",
    "            decoded_completion = audio_tokenizer.data_from_latent(completed_audio.bfloat16())\n",
    "        print_colored(f\"Decoded completion shape: {decoded_completion.shape}\", \"grey\")\n",
    "\n",
    "    print_colored(\"Preparing audio for playback...\", \"blue\")\n",
    "\n",
    "    audio_tensor = decoded_completion.cpu().squeeze()\n",
    "    if audio_tensor.ndim == 1:\n",
    "        audio_tensor = audio_tensor.unsqueeze(0)\n",
    "    audio_tensor = audio_tensor.float()\n",
    "\n",
    "    if audio_tensor.abs().max() > 1:\n",
    "        audio_tensor = audio_tensor / audio_tensor.abs().max()\n",
    "\n",
    "    return audio_tensor[:, max(prompt_len*2000 - 16000, 0):]\n",
    "\n",
    "num_completions = 10\n",
    "print_colored(f\"Generating {num_completions} completions...\", \"blue\")\n",
    "for _ in range(num_completions):\n",
    "    completion = get_completion(encoded_prompt_audio, prompt_len, gen_len=20*8) # 20 seconds of generation\n",
    "    display_audio(completion)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": ".venv",
   "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": "3.10.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}