Spaces:
Running
on
Zero
Running
on
Zero
Commit
•
39767da
1
Parent(s):
c75a416
Merge the contributions from Tango 2 full and the changes on Tango 2
Browse filesHi
@soujanyaporia
,
With this PR, this space will still have the new pipeline and will also have the new features from _Tango 2 full_ like:
- Simultaneous generations
- Seed parameter
_Successfully tested_
Regards,
Fabrice
app.py
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
import gradio as gr
|
|
|
2 |
import json
|
3 |
import torch
|
4 |
import wavio
|
@@ -23,7 +24,6 @@ from tqdm import tqdm
|
|
23 |
|
24 |
|
25 |
|
26 |
-
|
27 |
class Tango2Pipeline(DiffusionPipeline):
|
28 |
|
29 |
|
@@ -169,6 +169,7 @@ class Tango2Pipeline(DiffusionPipeline):
|
|
169 |
|
170 |
return AudioPipelineOutput(audios=wave)
|
171 |
|
|
|
172 |
|
173 |
# Automatic device detection
|
174 |
if torch.cuda.is_available():
|
@@ -249,21 +250,73 @@ pipe = Tango2Pipeline(vae=tango.vae,
|
|
249 |
scheduler=tango.scheduler
|
250 |
)
|
251 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
|
253 |
@spaces.GPU(duration=60)
|
254 |
-
def gradio_generate(
|
255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
256 |
#output_wave = tango.generate(prompt, steps, guidance)
|
257 |
# output_filename = f"{prompt.replace(' ', '_')}_{steps}_{guidance}"[:250] + ".wav"
|
258 |
-
output_wave = output_wave.audios[0]
|
259 |
-
output_filename = "temp.wav"
|
260 |
-
wavio.write(output_filename, output_wave, rate=16000, sampwidth=2)
|
261 |
-
|
262 |
-
if (output_format == "mp3"):
|
263 |
-
AudioSegment.from_wav("temp.wav").export("temp.mp3", format = "mp3")
|
264 |
-
output_filename = "temp.mp3"
|
265 |
|
266 |
-
return
|
|
|
|
|
|
|
|
|
267 |
|
268 |
# description_text = """
|
269 |
# <p><a href="https://huggingface.co/spaces/declare-lab/tango/blob/main/app.py?duplicate=true"> <img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> For faster inference without waiting in queue, you may duplicate the space and upgrade to a GPU in the settings. <br/><br/>
|
@@ -285,53 +338,130 @@ def gradio_generate(prompt, output_format, steps, guidance):
|
|
285 |
# <p/>
|
286 |
# """
|
287 |
description_text = """
|
|
|
288 |
<p><a href="https://huggingface.co/spaces/declare-lab/tango2/blob/main/app.py?duplicate=true"> <img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> For faster inference without waiting in queue, you may duplicate the space and upgrade to a GPU in the settings. <br/><br/>
|
289 |
Generate audio using Tango2 by providing a text prompt. Tango2 was built from Tango and was trained on <a href="https://huggingface.co/datasets/declare-lab/audio-alpaca">Audio-alpaca</a>
|
290 |
<br/><br/> This is the demo for Tango2 for text to audio generation: <a href="https://arxiv.org/abs/2404.09956">Read our paper.</a>
|
291 |
<p/>
|
292 |
"""
|
293 |
-
# Gradio input and output components
|
294 |
-
input_text = gr.Textbox(lines=2, label="Prompt")
|
295 |
-
output_format = gr.Radio(label = "Output format", info = "The file you can dowload", choices = ["mp3", "wav"], value = "wav")
|
296 |
-
output_audio = gr.Audio(label="Generated Audio", type="filepath")
|
297 |
-
denoising_steps = gr.Slider(minimum=100, maximum=200, value=100, step=1, label="Steps", interactive=True)
|
298 |
-
guidance_scale = gr.Slider(minimum=1, maximum=10, value=3, step=0.1, label="Guidance Scale", interactive=True)
|
299 |
|
300 |
# Gradio interface
|
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 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import random
|
3 |
import json
|
4 |
import torch
|
5 |
import wavio
|
|
|
24 |
|
25 |
|
26 |
|
|
|
27 |
class Tango2Pipeline(DiffusionPipeline):
|
28 |
|
29 |
|
|
|
169 |
|
170 |
return AudioPipelineOutput(audios=wave)
|
171 |
|
172 |
+
max_64_bit_int = 2**63 - 1
|
173 |
|
174 |
# Automatic device detection
|
175 |
if torch.cuda.is_available():
|
|
|
250 |
scheduler=tango.scheduler
|
251 |
)
|
252 |
|
253 |
+
|
254 |
+
def update_seed(is_randomize_seed, seed):
|
255 |
+
if is_randomize_seed:
|
256 |
+
return random.randint(0, max_64_bit_int)
|
257 |
+
return seed
|
258 |
+
|
259 |
+
def check(
|
260 |
+
prompt,
|
261 |
+
output_format,
|
262 |
+
output_number,
|
263 |
+
steps,
|
264 |
+
guidance,
|
265 |
+
is_randomize_seed,
|
266 |
+
seed
|
267 |
+
):
|
268 |
+
if prompt is None or prompt == "":
|
269 |
+
raise gr.Error("Please provide a prompt input.")
|
270 |
+
if not output_number in [1, 2, 3]:
|
271 |
+
raise gr.Error("Please ask for 1, 2 or 3 output files.")
|
272 |
+
|
273 |
+
def update_output(output_format, output_number):
|
274 |
+
return [
|
275 |
+
gr.update(format = output_format),
|
276 |
+
gr.update(format = output_format, visible = (2 <= output_number)),
|
277 |
+
gr.update(format = output_format, visible = (output_number == 3))
|
278 |
+
]
|
279 |
+
|
280 |
+
def generate_output(output_wave, output_format, output_number, output_index):
|
281 |
+
if (output_number < output_index):
|
282 |
+
return gr.update(format = output_format, visible = False)
|
283 |
+
|
284 |
+
output_wave = output_wave.audios[output_index - 1]
|
285 |
+
output_filename = "tmp" + str(output_index) + ".wav"
|
286 |
+
wavio.write(output_filename, output_wave, rate=16000, sampwidth=2)
|
287 |
+
|
288 |
+
if (output_format == "mp3"):
|
289 |
+
AudioSegment.from_wav("tmp" + str(output_index) + ".wav").export("tmp" + str(output_index) + ".mp3", format = "mp3")
|
290 |
+
output_filename = "tmp" + str(output_index) + ".mp3"
|
291 |
+
|
292 |
+
return gr.update(value = output_filename, format = output_format, visible = True)
|
293 |
|
294 |
@spaces.GPU(duration=60)
|
295 |
+
def gradio_generate(
|
296 |
+
prompt,
|
297 |
+
output_format,
|
298 |
+
output_number,
|
299 |
+
steps,
|
300 |
+
guidance,
|
301 |
+
is_randomize_seed,
|
302 |
+
seed
|
303 |
+
):
|
304 |
+
if seed is None:
|
305 |
+
seed = random.randint(0, max_64_bit_int)
|
306 |
+
|
307 |
+
random.seed(seed)
|
308 |
+
torch.manual_seed(seed)
|
309 |
+
|
310 |
+
output_wave = pipe(prompt, steps, guidance, samples = output_number) ## Using pipeline automatically uses flash attention for torch2.0 above
|
311 |
+
|
312 |
#output_wave = tango.generate(prompt, steps, guidance)
|
313 |
# output_filename = f"{prompt.replace(' ', '_')}_{steps}_{guidance}"[:250] + ".wav"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
314 |
|
315 |
+
return [
|
316 |
+
generate_output(output_wave, output_format, output_number, 1),
|
317 |
+
generate_output(output_wave, output_format, output_number, 2),
|
318 |
+
generate_output(output_wave, output_format, output_number, 3)
|
319 |
+
]
|
320 |
|
321 |
# description_text = """
|
322 |
# <p><a href="https://huggingface.co/spaces/declare-lab/tango/blob/main/app.py?duplicate=true"> <img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> For faster inference without waiting in queue, you may duplicate the space and upgrade to a GPU in the settings. <br/><br/>
|
|
|
338 |
# <p/>
|
339 |
# """
|
340 |
description_text = """
|
341 |
+
<h1><center>Tango 2: Aligning Diffusion-based Text-to-Audio Generations through Direct Preference Optimization</center></h1>
|
342 |
<p><a href="https://huggingface.co/spaces/declare-lab/tango2/blob/main/app.py?duplicate=true"> <img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> For faster inference without waiting in queue, you may duplicate the space and upgrade to a GPU in the settings. <br/><br/>
|
343 |
Generate audio using Tango2 by providing a text prompt. Tango2 was built from Tango and was trained on <a href="https://huggingface.co/datasets/declare-lab/audio-alpaca">Audio-alpaca</a>
|
344 |
<br/><br/> This is the demo for Tango2 for text to audio generation: <a href="https://arxiv.org/abs/2404.09956">Read our paper.</a>
|
345 |
<p/>
|
346 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
347 |
|
348 |
# Gradio interface
|
349 |
+
with gr.Blocks() as interface:
|
350 |
+
gr.HTML(description_text)
|
351 |
+
with gr.Row():
|
352 |
+
with gr.Column():
|
353 |
+
input_text = gr.Textbox(lines=2, label="Prompt")
|
354 |
+
output_format = gr.Radio(label = "Output format", info = "The file you can dowload", choices = ["mp3", "wav"], value = "wav")
|
355 |
+
output_number = gr.Slider(label = "Number of generations", info = "1, 2 or 3 output files", minimum = 1, maximum = 3, value = 1, step = 1, interactive = True)
|
356 |
+
denoising_steps = gr.Slider(minimum=10, maximum=200, value=100, step=1, label="Steps", interactive=True)
|
357 |
+
guidance_scale = gr.Slider(minimum=1, maximum=10, value=3, step=0.1, label="Guidance Scale", interactive=True)
|
358 |
+
randomize_seed = gr.Checkbox(label = "\U0001F3B2 Randomize seed", value = True, info = "If checked, result is always different")
|
359 |
+
seed = gr.Slider(minimum = 0, maximum = max_64_bit_int, step = 1, randomize = True, label = "Seed")
|
360 |
+
|
361 |
+
submit = gr.Button("Generate", variant = "primary")
|
362 |
+
|
363 |
+
with gr.Column():
|
364 |
+
output_audio_1 = gr.Audio(label = "Generated Audio #1/3", format = "wav", type="numpy")
|
365 |
+
output_audio_2 = gr.Audio(label = "Generated Audio #2/3", format = "wav", type="numpy")
|
366 |
+
output_audio_3 = gr.Audio(label = "Generated Audio #3/3", format = "wav", type="numpy")
|
367 |
+
|
368 |
+
submit.click(fn = update_seed, inputs = [
|
369 |
+
randomize_seed,
|
370 |
+
seed
|
371 |
+
], outputs = [
|
372 |
+
seed
|
373 |
+
], queue = False, show_progress = False).then(fn = check, inputs = [
|
374 |
+
input_text,
|
375 |
+
output_format,
|
376 |
+
output_number,
|
377 |
+
denoising_steps,
|
378 |
+
guidance_scale,
|
379 |
+
randomize_seed,
|
380 |
+
seed
|
381 |
+
], outputs = [], queue = False, show_progress = False).success(fn = update_output, inputs = [
|
382 |
+
output_format,
|
383 |
+
output_number
|
384 |
+
], outputs = [
|
385 |
+
output_audio_1,
|
386 |
+
output_audio_2,
|
387 |
+
output_audio_3
|
388 |
+
], queue = False, show_progress = False).success(fn = gradio_generate, inputs = [
|
389 |
+
input_text,
|
390 |
+
output_format,
|
391 |
+
output_number,
|
392 |
+
denoising_steps,
|
393 |
+
guidance_scale,
|
394 |
+
randomize_seed,
|
395 |
+
seed
|
396 |
+
], outputs = [
|
397 |
+
output_audio_1,
|
398 |
+
output_audio_2,
|
399 |
+
output_audio_3
|
400 |
+
], scroll_to_output = True)
|
401 |
+
|
402 |
+
gr.Examples(
|
403 |
+
fn = gradio_generate,
|
404 |
+
inputs = [
|
405 |
+
input_text,
|
406 |
+
output_format,
|
407 |
+
output_number,
|
408 |
+
denoising_steps,
|
409 |
+
guidance_scale,
|
410 |
+
randomize_seed,
|
411 |
+
seed
|
412 |
+
],
|
413 |
+
outputs = [
|
414 |
+
output_audio_1,
|
415 |
+
output_audio_2,
|
416 |
+
output_audio_3
|
417 |
+
],
|
418 |
+
examples = [
|
419 |
+
["Quiet speech and then airplane flying away", "wav", 3, 200, 3, False, 123],
|
420 |
+
["A bicycle peddling on dirt and gravel followed by a man speaking then laughing", "wav", 3, 200, 3, False, 123],
|
421 |
+
["Ducks quack and water splashes with some animal screeching in the background", "wav", 3, 200, 3, False, 123],
|
422 |
+
["Describe the sound of the ocean", "wav", 3, 200, 3, False, 123],
|
423 |
+
["A woman and a baby are having a conversation", "wav", 3, 200, 3, False, 123],
|
424 |
+
["A man speaks followed by a popping noise and laughter", "wav", 3, 200, 3, False, 123],
|
425 |
+
["A cup is filled from a faucet", "wav", 3, 200, 3, False, 123],
|
426 |
+
["An audience cheering and clapping", "wav", 3, 200, 3, False, 123],
|
427 |
+
["Rolling thunder with lightning strikes", "wav", 3, 200, 3, False, 123],
|
428 |
+
["A dog barking and a cat mewing and a racing car passes by", "wav", 3, 200, 3, False, 123],
|
429 |
+
["Gentle water stream, birds chirping and sudden gun shot", "wav", 3, 200, 3, False, 123],
|
430 |
+
["A man talking followed by a goat baaing then a metal gate sliding shut as ducks quack and wind blows into a microphone.", 3, 200, 3, False, 123],
|
431 |
+
["A dog barking", "wav", 3, 200, 3, False, 123],
|
432 |
+
["A cat meowing", "wav", 3, 200, 3, False, 123],
|
433 |
+
["Wooden table tapping sound while water pouring", "wav", 3, 200, 3, False, 123],
|
434 |
+
["Applause from a crowd with distant clicking and a man speaking over a loudspeaker", "wav", 3, 200, 3, False, 123],
|
435 |
+
["two gunshots followed by birds flying away while chirping", "wav", 3, 200, 3, False, 123],
|
436 |
+
["Whistling with birds chirping", "wav", 3, 200, 3, False, 123],
|
437 |
+
["A person snoring", "wav", 3, 200, 3, False, 123],
|
438 |
+
["Motor vehicles are driving with loud engines and a person whistles", "wav", 3, 200, 3, False, 123],
|
439 |
+
["People cheering in a stadium while thunder and lightning strikes", "wav", 3, 200, 3, False, 123],
|
440 |
+
["A helicopter is in flight", "wav", 3, 200, 3, False, 123],
|
441 |
+
["A dog barking and a man talking and a racing car passes by", "wav", 3, 200, 3, False, 123],
|
442 |
+
],
|
443 |
+
cache_examples = "lazy",
|
444 |
+
)
|
445 |
+
|
446 |
+
gr.Markdown(
|
447 |
+
"""
|
448 |
+
## How to prompt your sound
|
449 |
+
You can use round brackets to increase the importance of a part:
|
450 |
+
```
|
451 |
+
Peaceful and (calming) ambient music with singing bowl and other instruments
|
452 |
+
```
|
453 |
+
You can use several levels of round brackets to even more increase the importance of a part:
|
454 |
+
```
|
455 |
+
(Peaceful) and ((calming)) ambient music with singing bowl and other instruments
|
456 |
+
```
|
457 |
+
You can use number instead of several round brackets:
|
458 |
+
```
|
459 |
+
(Peaceful:1.5) and ((calming)) ambient music with singing bowl and other instruments
|
460 |
+
```
|
461 |
+
You can do the same thing with square brackets to decrease the importance of a part:
|
462 |
+
```
|
463 |
+
(Peaceful:1.5) and ((calming)) ambient music with [singing:2] bowl and other instruments
|
464 |
+
"""
|
465 |
+
)
|
466 |
+
|
467 |
+
interface.queue(10).launch()
|