Spaces:
Running
Running
File size: 28,500 Bytes
bd6e54b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 |
import librosa
import numpy as np
import torch
import gradio as gr
from scipy.ndimage import zoom
from model.DiffSynthSampler import DiffSynthSampler
from tools import adjust_audio_length, rescale, safe_int, pad_STFT, encode_stft
from webUI.natural_language_guided_STFT.utils import latent_representation_to_Gradio_image
from webUI.natural_language_guided_STFT.utils import InputBatch2Encode_STFT, encodeBatch2GradioOutput_STFT
def get_super_resolution_with_text_module(gradioWebUI, inpaintWithText_state):
# Load configurations
uNet = gradioWebUI.uNet
freq_resolution, time_resolution = gradioWebUI.freq_resolution, gradioWebUI.time_resolution
VAE_scale = gradioWebUI.VAE_scale
height, width, channels = int(freq_resolution/VAE_scale), int(time_resolution/VAE_scale), gradioWebUI.channels
timesteps = gradioWebUI.timesteps
VAE_encoder = gradioWebUI.VAE_encoder
VAE_quantizer = gradioWebUI.VAE_quantizer
VAE_decoder = gradioWebUI.VAE_decoder
CLAP = gradioWebUI.CLAP
CLAP_tokenizer = gradioWebUI.CLAP_tokenizer
device = gradioWebUI.device
squared = gradioWebUI.squared
sample_rate = gradioWebUI.sample_rate
noise_strategy = gradioWebUI.noise_strategy
def receive_uopoad_origin_audio(sound2sound_duration, sound2sound_origin_source, sound2sound_origin_upload, sound2sound_origin_microphone,
inpaintWithText_dict):
if sound2sound_origin_source == "upload":
origin_sr, origin_audio = sound2sound_origin_upload
else:
origin_sr, origin_audio = sound2sound_origin_microphone
origin_audio = origin_audio / np.max(np.abs(origin_audio))
width = int(time_resolution*((sound2sound_duration+1)/4) / VAE_scale)
audio_length = 256 * (VAE_scale * width - 1)
origin_audio = adjust_audio_length(origin_audio, audio_length, origin_sr, sample_rate)
D = librosa.stft(origin_audio, n_fft=1024, hop_length=256, win_length=1024)
padded_D = pad_STFT(D)
encoded_D = encode_stft(padded_D)
# Todo: justify batchsize to 1
origin_spectrogram_batch_tensor = torch.from_numpy(
np.repeat(encoded_D[np.newaxis, :, :, :], 1, axis=0)).float().to(device)
# Todo: remove hard-coding
origin_flipped_log_spectrums, origin_flipped_phases, origin_signals, origin_latent_representations, quantized_origin_latent_representations = InputBatch2Encode_STFT(
VAE_encoder, origin_spectrogram_batch_tensor, resolution=(512, width * VAE_scale), quantizer=VAE_quantizer, squared=squared)
if sound2sound_origin_source == "upload":
inpaintWithText_dict["origin_upload_latent_representations"] = origin_latent_representations.tolist()
inpaintWithText_dict[
"sound2sound_origin_upload_latent_representation_image"] = latent_representation_to_Gradio_image(
origin_latent_representations[0]).tolist()
inpaintWithText_dict[
"sound2sound_origin_upload_quantized_latent_representation_image"] = latent_representation_to_Gradio_image(
quantized_origin_latent_representations[0]).tolist()
return {sound2sound_origin_spectrogram_upload_image: origin_flipped_log_spectrums[0],
sound2sound_origin_phase_upload_image: origin_flipped_phases[0],
sound2sound_origin_spectrogram_microphone_image: gr.update(),
sound2sound_origin_phase_microphone_image: gr.update(),
sound2sound_origin_upload_latent_representation_image: latent_representation_to_Gradio_image(
origin_latent_representations[0]),
sound2sound_origin_upload_quantized_latent_representation_image: latent_representation_to_Gradio_image(
quantized_origin_latent_representations[0]),
sound2sound_origin_microphone_latent_representation_image: gr.update(),
sound2sound_origin_microphone_quantized_latent_representation_image: gr.update(),
inpaintWithText_state: inpaintWithText_dict}
else:
inpaintWithText_dict["origin_microphone_latent_representations"] = origin_latent_representations.tolist()
inpaintWithText_dict[
"sound2sound_origin_microphone_latent_representation_image"] = latent_representation_to_Gradio_image(
origin_latent_representations[0]).tolist()
inpaintWithText_dict[
"sound2sound_origin_microphone_quantized_latent_representation_image"] = latent_representation_to_Gradio_image(
quantized_origin_latent_representations[0]).tolist()
return {sound2sound_origin_spectrogram_upload_image: origin_flipped_log_spectrums[0],
sound2sound_origin_phase_upload_image: origin_flipped_phases[0],
sound2sound_origin_spectrogram_microphone_image: gr.update(),
sound2sound_origin_phase_microphone_image: gr.update(),
sound2sound_origin_upload_latent_representation_image: latent_representation_to_Gradio_image(
origin_latent_representations[0]),
sound2sound_origin_upload_quantized_latent_representation_image: latent_representation_to_Gradio_image(
quantized_origin_latent_representations[0]),
sound2sound_origin_microphone_latent_representation_image: gr.update(),
sound2sound_origin_microphone_quantized_latent_representation_image: gr.update(),
inpaintWithText_state: inpaintWithText_dict}
def sound2sound_sample(sound2sound_origin_spectrogram_upload, sound2sound_origin_spectrogram_microphone,
text2sound_prompts, text2sound_negative_prompts, sound2sound_batchsize,
sound2sound_guidance_scale, sound2sound_sampler,
sound2sound_sample_steps, sound2sound_origin_source,
sound2sound_noising_strength, sound2sound_seed, sound2sound_inpaint_area, inpaintWithText_dict
):
# input preprocessing
sound2sound_seed = safe_int(sound2sound_seed, 12345678)
sound2sound_batchsize = int(sound2sound_batchsize)
noising_strength = sound2sound_noising_strength
sound2sound_sample_steps = int(sound2sound_sample_steps)
CFG = int(sound2sound_guidance_scale)
text2sound_embedding = \
CLAP.get_text_features(**CLAP_tokenizer([text2sound_prompts], padding=True, return_tensors="pt"))[0].to(device)
if sound2sound_origin_source == "upload":
origin_latent_representations = torch.tensor(
inpaintWithText_dict["origin_upload_latent_representations"]).repeat(sound2sound_batchsize, 1, 1, 1).to(
device)
elif sound2sound_origin_source == "microphone":
origin_latent_representations = torch.tensor(
inpaintWithText_dict["origin_microphone_latent_representations"]).repeat(sound2sound_batchsize, 1, 1, 1).to(
device)
else:
print("Input source not in ['upload', 'microphone']!")
raise NotImplementedError()
high_resolution_latent_representations = torch.zeros((sound2sound_batchsize, channels, 256, 64)).to(device)
high_resolution_latent_representations[:, :, :128, :] = origin_latent_representations
latent_mask = np.ones((256, 64))
latent_mask[192:, :] = 0.0
print(f"latent_mask mean: {np.mean(latent_mask)}")
if sound2sound_inpaint_area == "inpaint masked":
latent_mask = 1 - latent_mask
latent_mask = torch.from_numpy(latent_mask).unsqueeze(0).unsqueeze(1).repeat(sound2sound_batchsize, channels, 1,
1).float().to(device)
latent_mask = torch.flip(latent_mask, [2])
mySampler = DiffSynthSampler(timesteps, height=height*2, channels=channels, noise_strategy=noise_strategy)
unconditional_condition = \
CLAP.get_text_features(**CLAP_tokenizer([text2sound_negative_prompts], padding=True, return_tensors="pt"))[0]
mySampler.activate_classifier_free_guidance(CFG, unconditional_condition.to(device))
normalized_sample_steps = int(sound2sound_sample_steps / noising_strength)
mySampler.respace(list(np.linspace(0, timesteps - 1, normalized_sample_steps, dtype=np.int32)))
# Todo: remove hard-coding
width = high_resolution_latent_representations.shape[-1]
condition = text2sound_embedding.repeat(sound2sound_batchsize, 1)
new_sound_latent_representations, initial_noise = \
mySampler.inpaint_sample(model=uNet, shape=(sound2sound_batchsize, channels, height*2, width),
seed=sound2sound_seed,
noising_strength=noising_strength,
guide_img=high_resolution_latent_representations, mask=latent_mask, return_tensor=True,
condition=condition, sampler=sound2sound_sampler)
new_sound_latent_representations = new_sound_latent_representations[-1]
# Quantize new sound latent representations
quantized_new_sound_latent_representations, loss, (_, _, _) = VAE_quantizer(new_sound_latent_representations)
new_sound_flipped_log_spectrums, new_sound_flipped_phases, new_sound_signals, _, _, _ = encodeBatch2GradioOutput_STFT(VAE_decoder,
quantized_new_sound_latent_representations,
resolution=(
1024,
width * VAE_scale),
original_STFT_batch=None
)
new_sound_latent_representation_gradio_images = []
new_sound_quantized_latent_representation_gradio_images = []
new_sound_spectrogram_gradio_images = []
new_sound_phase_gradio_images = []
new_sound_rec_signals_gradio = []
for i in range(sound2sound_batchsize):
new_sound_latent_representation_gradio_images.append(
latent_representation_to_Gradio_image(new_sound_latent_representations[i]))
new_sound_quantized_latent_representation_gradio_images.append(
latent_representation_to_Gradio_image(quantized_new_sound_latent_representations[i]))
new_sound_spectrogram_gradio_images.append(new_sound_flipped_log_spectrums[i])
new_sound_phase_gradio_images.append(new_sound_flipped_phases[i])
new_sound_rec_signals_gradio.append((sample_rate, new_sound_signals[i]))
inpaintWithText_dict[
"new_sound_latent_representation_gradio_images"] = new_sound_latent_representation_gradio_images
inpaintWithText_dict[
"new_sound_quantized_latent_representation_gradio_images"] = new_sound_quantized_latent_representation_gradio_images
inpaintWithText_dict["new_sound_spectrogram_gradio_images"] = new_sound_spectrogram_gradio_images
inpaintWithText_dict["new_sound_phase_gradio_images"] = new_sound_phase_gradio_images
inpaintWithText_dict["new_sound_rec_signals_gradio"] = new_sound_rec_signals_gradio
return {sound2sound_new_sound_latent_representation_image: latent_representation_to_Gradio_image(
new_sound_latent_representations[0]),
sound2sound_new_sound_quantized_latent_representation_image: latent_representation_to_Gradio_image(
quantized_new_sound_latent_representations[0]),
sound2sound_new_sound_spectrogram_image: new_sound_flipped_log_spectrums[0],
sound2sound_new_sound_phase_image: new_sound_flipped_phases[0],
sound2sound_new_sound_audio: (sample_rate, new_sound_signals[0]),
sound2sound_sample_index_slider: gr.update(minimum=0, maximum=sound2sound_batchsize - 1, value=0,
step=1.0,
visible=True,
label="Sample index",
info="Swipe to view other samples"),
sound2sound_seed_textbox: sound2sound_seed,
inpaintWithText_state: inpaintWithText_dict}
def show_sound2sound_sample(sound2sound_sample_index, inpaintWithText_dict):
sample_index = int(sound2sound_sample_index)
return {sound2sound_new_sound_latent_representation_image:
inpaintWithText_dict["new_sound_latent_representation_gradio_images"][sample_index],
sound2sound_new_sound_quantized_latent_representation_image:
inpaintWithText_dict["new_sound_quantized_latent_representation_gradio_images"][sample_index],
sound2sound_new_sound_spectrogram_image: inpaintWithText_dict["new_sound_spectrogram_gradio_images"][
sample_index],
sound2sound_new_sound_phase_image: inpaintWithText_dict["new_sound_phase_gradio_images"][
sample_index],
sound2sound_new_sound_audio: inpaintWithText_dict["new_sound_rec_signals_gradio"][sample_index]}
def sound2sound_switch_origin_source(sound2sound_origin_source):
if sound2sound_origin_source == "upload":
return {sound2sound_origin_upload_audio: gr.update(visible=True),
sound2sound_origin_microphone_audio: gr.update(visible=False),
sound2sound_origin_spectrogram_upload_image: gr.update(visible=True),
sound2sound_origin_phase_upload_image: gr.update(visible=True),
sound2sound_origin_spectrogram_microphone_image: gr.update(visible=False),
sound2sound_origin_phase_microphone_image: gr.update(visible=False),
sound2sound_origin_upload_latent_representation_image: gr.update(visible=True),
sound2sound_origin_upload_quantized_latent_representation_image: gr.update(visible=True),
sound2sound_origin_microphone_latent_representation_image: gr.update(visible=False),
sound2sound_origin_microphone_quantized_latent_representation_image: gr.update(visible=False)}
elif sound2sound_origin_source == "microphone":
return {sound2sound_origin_upload_audio: gr.update(visible=False),
sound2sound_origin_microphone_audio: gr.update(visible=True),
sound2sound_origin_spectrogram_upload_image: gr.update(visible=False),
sound2sound_origin_phase_upload_image: gr.update(visible=False),
sound2sound_origin_spectrogram_microphone_image: gr.update(visible=True),
sound2sound_origin_phase_microphone_image: gr.update(visible=True),
sound2sound_origin_upload_latent_representation_image: gr.update(visible=False),
sound2sound_origin_upload_quantized_latent_representation_image: gr.update(visible=False),
sound2sound_origin_microphone_latent_representation_image: gr.update(visible=True),
sound2sound_origin_microphone_quantized_latent_representation_image: gr.update(visible=True)}
else:
print("Input source not in ['upload', 'microphone']!")
with gr.Tab("Super Resolution"):
gr.Markdown("Select the area to inpaint and use the prompt to guide the synthesis of a new sound!")
with gr.Row(variant="panel"):
with gr.Column(scale=3):
text2sound_prompts_textbox = gr.Textbox(label="Positive prompt", lines=2, value="organ")
text2sound_negative_prompts_textbox = gr.Textbox(label="Negative prompt", lines=2, value="")
with gr.Column(scale=1):
sound2sound_sample_button = gr.Button(variant="primary", value="Generate", scale=1)
sound2sound_sample_index_slider = gr.Slider(minimum=0, maximum=3, value=0, step=1.0, visible=False,
label="Sample index",
info="Swipe to view other samples")
with gr.Row(variant="panel"):
with gr.Column(scale=1):
with gr.Tab("Origin sound"):
sound2sound_duration_slider = gradioWebUI.get_duration_slider()
sound2sound_origin_source_radio = gr.Radio(choices=["upload", "microphone"], value="upload",
label="Input source")
sound2sound_origin_upload_audio = gr.Audio(type="numpy", label="Upload", source="upload",
interactive=True, visible=True)
sound2sound_origin_microphone_audio = gr.Audio(type="numpy", label="Record", source="microphone",
interactive=True, visible=False)
with gr.Row(variant="panel"):
sound2sound_origin_spectrogram_upload_image = gr.Image(label="Original upload spectrogram",
type="numpy", height=600,
visible=True, tool="sketch")
sound2sound_origin_phase_upload_image = gr.Image(label="Original upload phase",
type="numpy", height=600,
visible=True)
sound2sound_origin_spectrogram_microphone_image = gr.Image(label="Original microphone spectrogram",
type="numpy", height=600,
visible=False, tool="sketch")
sound2sound_origin_phase_microphone_image = gr.Image(label="Original microphone phase",
type="numpy", height=600,
visible=False)
sound2sound_inpaint_area_radio = gr.Radio(choices=["inpaint masked", "inpaint not masked"],
value="inpaint masked")
with gr.Tab("Sound2sound settings"):
sound2sound_sample_steps_slider = gradioWebUI.get_sample_steps_slider()
sound2sound_sampler_radio = gradioWebUI.get_sampler_radio()
sound2sound_batchsize_slider = gradioWebUI.get_batchsize_slider()
sound2sound_noising_strength_slider = gradioWebUI.get_noising_strength_slider(default_noising_strength=1.0)
sound2sound_guidance_scale_slider = gradioWebUI.get_guidance_scale_slider()
sound2sound_seed_textbox = gradioWebUI.get_seed_textbox()
with gr.Column(scale=1):
sound2sound_new_sound_audio = gr.Audio(type="numpy", label="Play new sound", interactive=False)
with gr.Row(variant="panel"):
sound2sound_new_sound_spectrogram_image = gr.Image(label="New sound spectrogram", type="numpy",
height=1200, scale=1)
sound2sound_new_sound_phase_image = gr.Image(label="New sound phase", type="numpy",
height=1200, scale=1)
with gr.Row(variant="panel"):
sound2sound_origin_upload_latent_representation_image = gr.Image(label="Original latent representation",
type="numpy", height=1200,
visible=True)
sound2sound_origin_upload_quantized_latent_representation_image = gr.Image(
label="Original quantized latent representation", type="numpy", height=1200, visible=True)
sound2sound_origin_microphone_latent_representation_image = gr.Image(label="Original latent representation",
type="numpy", height=1200,
visible=False)
sound2sound_origin_microphone_quantized_latent_representation_image = gr.Image(
label="Original quantized latent representation", type="numpy", height=1200, visible=False)
sound2sound_new_sound_latent_representation_image = gr.Image(label="New latent representation",
type="numpy", height=1200)
sound2sound_new_sound_quantized_latent_representation_image = gr.Image(
label="New sound quantized latent representation", type="numpy", height=1200)
sound2sound_origin_upload_audio.change(receive_uopoad_origin_audio,
inputs=[sound2sound_duration_slider, sound2sound_origin_source_radio, sound2sound_origin_upload_audio,
sound2sound_origin_microphone_audio, inpaintWithText_state],
outputs=[sound2sound_origin_spectrogram_upload_image,
sound2sound_origin_phase_upload_image,
sound2sound_origin_spectrogram_microphone_image,
sound2sound_origin_phase_microphone_image,
sound2sound_origin_upload_latent_representation_image,
sound2sound_origin_upload_quantized_latent_representation_image,
sound2sound_origin_microphone_latent_representation_image,
sound2sound_origin_microphone_quantized_latent_representation_image,
inpaintWithText_state])
sound2sound_origin_microphone_audio.change(receive_uopoad_origin_audio,
inputs=[sound2sound_duration_slider, sound2sound_origin_source_radio, sound2sound_origin_upload_audio,
sound2sound_origin_microphone_audio, inpaintWithText_state],
outputs=[sound2sound_origin_spectrogram_upload_image,
sound2sound_origin_phase_upload_image,
sound2sound_origin_spectrogram_microphone_image,
sound2sound_origin_phase_microphone_image,
sound2sound_origin_upload_latent_representation_image,
sound2sound_origin_upload_quantized_latent_representation_image,
sound2sound_origin_microphone_latent_representation_image,
sound2sound_origin_microphone_quantized_latent_representation_image,
inpaintWithText_state])
sound2sound_sample_button.click(sound2sound_sample,
inputs=[sound2sound_origin_spectrogram_upload_image,
sound2sound_origin_spectrogram_microphone_image,
text2sound_prompts_textbox,
text2sound_negative_prompts_textbox,
sound2sound_batchsize_slider,
sound2sound_guidance_scale_slider,
sound2sound_sampler_radio,
sound2sound_sample_steps_slider,
sound2sound_origin_source_radio,
sound2sound_noising_strength_slider,
sound2sound_seed_textbox,
sound2sound_inpaint_area_radio,
inpaintWithText_state],
outputs=[sound2sound_new_sound_latent_representation_image,
sound2sound_new_sound_quantized_latent_representation_image,
sound2sound_new_sound_spectrogram_image,
sound2sound_new_sound_phase_image,
sound2sound_new_sound_audio,
sound2sound_sample_index_slider,
sound2sound_seed_textbox,
inpaintWithText_state])
sound2sound_sample_index_slider.change(show_sound2sound_sample,
inputs=[sound2sound_sample_index_slider, inpaintWithText_state],
outputs=[sound2sound_new_sound_latent_representation_image,
sound2sound_new_sound_quantized_latent_representation_image,
sound2sound_new_sound_spectrogram_image,
sound2sound_new_sound_phase_image,
sound2sound_new_sound_audio])
sound2sound_origin_source_radio.change(sound2sound_switch_origin_source,
inputs=[sound2sound_origin_source_radio],
outputs=[sound2sound_origin_upload_audio,
sound2sound_origin_microphone_audio,
sound2sound_origin_spectrogram_upload_image,
sound2sound_origin_phase_upload_image,
sound2sound_origin_spectrogram_microphone_image,
sound2sound_origin_phase_microphone_image,
sound2sound_origin_upload_latent_representation_image,
sound2sound_origin_upload_quantized_latent_representation_image,
sound2sound_origin_microphone_latent_representation_image,
sound2sound_origin_microphone_quantized_latent_representation_image])
|