app.py

import gradio as gr
import json
import torch
import wavio
from tqdm import tqdm
from huggingface_hub import snapshot_download
from models import AudioDiffusion, DDPMScheduler
from audioldm.audio.stft import TacotronSTFT
from audioldm.variational_autoencoder import AutoencoderKL
from pydub import AudioSegment
from gradio import Markdown
import spaces

import torch
#from diffusers.models.autoencoder_kl import AutoencoderKL
from diffusers.models.unet_2d_condition import UNet2DConditionModel
from diffusers import DiffusionPipeline,AudioPipelineOutput
from transformers import CLIPTextModel, T5EncoderModel, AutoModel, T5Tokenizer, T5TokenizerFast
from typing import Union
from diffusers.utils.torch_utils import randn_tensor
from tqdm import tqdm





class TangoPipeline(DiffusionPipeline):

    
    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: T5EncoderModel,
        tokenizer: Union[T5Tokenizer, T5TokenizerFast],
        unet: UNet2DConditionModel,
        scheduler: DDPMScheduler
    ):
        
        super().__init__()
    
        self.register_modules(vae=vae,
        text_encoder=text_encoder,
        tokenizer=tokenizer,
        unet=unet,
        scheduler=scheduler
        )
        
    
    def _encode_prompt(self, prompt):
        device = self.text_encoder.device
        
        batch = self.tokenizer(
            prompt, max_length=self.tokenizer.model_max_length, padding=True, truncation=True, return_tensors="pt"
        )
        input_ids, attention_mask = batch.input_ids.to(device), batch.attention_mask.to(device)

       
        encoder_hidden_states = self.text_encoder(
                input_ids=input_ids, attention_mask=attention_mask
            )[0]

        boolean_encoder_mask = (attention_mask == 1).to(device)
        
        return encoder_hidden_states, boolean_encoder_mask
        
    def _encode_text_classifier_free(self, prompt, num_samples_per_prompt):
        device = self.text_encoder.device
        batch = self.tokenizer(
            prompt, max_length=self.tokenizer.model_max_length, padding=True, truncation=True, return_tensors="pt"
        )
        input_ids, attention_mask = batch.input_ids.to(device), batch.attention_mask.to(device)

        with torch.no_grad():
            prompt_embeds = self.text_encoder(
                input_ids=input_ids, attention_mask=attention_mask
            )[0]
                
        prompt_embeds = prompt_embeds.repeat_interleave(num_samples_per_prompt, 0)
        attention_mask = attention_mask.repeat_interleave(num_samples_per_prompt, 0)

        # get unconditional embeddings for classifier free guidance
        uncond_tokens = [""] * len(prompt)

        max_length = prompt_embeds.shape[1]
        uncond_batch = self.tokenizer(
            uncond_tokens, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt",
        )
        uncond_input_ids = uncond_batch.input_ids.to(device)
        uncond_attention_mask = uncond_batch.attention_mask.to(device)

        with torch.no_grad():
            negative_prompt_embeds = self.text_encoder(
                input_ids=uncond_input_ids, attention_mask=uncond_attention_mask
            )[0]
                
        negative_prompt_embeds = negative_prompt_embeds.repeat_interleave(num_samples_per_prompt, 0)
        uncond_attention_mask = uncond_attention_mask.repeat_interleave(num_samples_per_prompt, 0)

        # For classifier free guidance, we need to do two forward passes.
        # We concatenate the unconditional and text embeddings into a single batch to avoid doing two forward passes
        prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
        prompt_mask = torch.cat([uncond_attention_mask, attention_mask])
        boolean_prompt_mask = (prompt_mask == 1).to(device)

        return prompt_embeds, boolean_prompt_mask
        
    def prepare_latents(self, batch_size, inference_scheduler, num_channels_latents, dtype, device):
        shape = (batch_size, num_channels_latents, 256, 16)
        latents = randn_tensor(shape, generator=None, device=device, dtype=dtype)
        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * inference_scheduler.init_noise_sigma
        return latents
    
    @torch.no_grad()
    def inference(self, prompt, inference_scheduler, num_steps=20, guidance_scale=3, num_samples_per_prompt=1, 
                  disable_progress=True):
        device = self.text_encoder.device
        classifier_free_guidance = guidance_scale > 1.0
        batch_size = len(prompt) * num_samples_per_prompt

        if classifier_free_guidance:
            prompt_embeds, boolean_prompt_mask = self._encode_text_classifier_free(prompt, num_samples_per_prompt)
        else:
            prompt_embeds, boolean_prompt_mask = self._encode_text(prompt)
            prompt_embeds = prompt_embeds.repeat_interleave(num_samples_per_prompt, 0)
            boolean_prompt_mask = boolean_prompt_mask.repeat_interleave(num_samples_per_prompt, 0)

        inference_scheduler.set_timesteps(num_steps, device=device)
        timesteps = inference_scheduler.timesteps

        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(batch_size, inference_scheduler, num_channels_latents, prompt_embeds.dtype, device)

        num_warmup_steps = len(timesteps) - num_steps * inference_scheduler.order
        progress_bar = tqdm(range(num_steps), disable=disable_progress)

        for i, t in enumerate(timesteps):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = torch.cat([latents] * 2) if classifier_free_guidance else latents
            latent_model_input = inference_scheduler.scale_model_input(latent_model_input, t)

            noise_pred = self.unet(
                latent_model_input, t, encoder_hidden_states=prompt_embeds,
                encoder_attention_mask=boolean_prompt_mask
            ).sample

            # perform guidance
            if classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

            # compute the previous noisy sample x_t -> x_t-1
            latents = inference_scheduler.step(noise_pred, t, latents).prev_sample

            # call the callback, if provided
            if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % inference_scheduler.order == 0):
                progress_bar.update(1)

        return latents
        
    @torch.no_grad()
    def __call__(self, prompt, steps=100, guidance=3, samples=1, disable_progress=True):
        """ Genrate audio for a single prompt string. """
        with torch.no_grad():
            latents = self.inference([prompt], self.scheduler, steps, guidance, samples, disable_progress=disable_progress)
            mel = self.vae.decode_first_stage(latents)
            wave = self.vae.decode_to_waveform(mel)


        return AudioPipelineOutput(audios=wave)


# Automatic device detection
if torch.cuda.is_available():
    device_type = "cuda"
    device_selection = "cuda:0"
else:
    device_type = "cpu"
    device_selection = "cpu"

class Tango:
    def __init__(self, name="declare-lab/tango-music-af-ft-mc", device=device_selection):
        
        path = snapshot_download(repo_id=name)
        
        vae_config = json.load(open("{}/vae_config.json".format(path)))
        stft_config = json.load(open("{}/stft_config.json".format(path)))
        main_config = json.load(open("{}/main_config.json".format(path)))
        
        self.vae = AutoencoderKL(**vae_config).to(device)
        self.stft = TacotronSTFT(**stft_config).to(device)
        self.model = AudioDiffusion(**main_config).to(device)
        
        vae_weights = torch.load("{}/pytorch_model_vae.bin".format(path), map_location=device)
        stft_weights = torch.load("{}/pytorch_model_stft.bin".format(path), map_location=device)
        main_weights = torch.load("{}/pytorch_model_main.bin".format(path), map_location=device)
        
        self.vae.load_state_dict(vae_weights)
        self.stft.load_state_dict(stft_weights)
        self.model.load_state_dict(main_weights)

        print ("Successfully loaded checkpoint from:", name)
        
        self.vae.eval()
        self.stft.eval()
        self.model.eval()
        
        self.scheduler = DDPMScheduler.from_pretrained(main_config["scheduler_name"], subfolder="scheduler")
        
    def chunks(self, lst, n):
        """ Yield successive n-sized chunks from a list. """
        for i in range(0, len(lst), n):
            yield lst[i:i + n]
        
    def generate(self, prompt, steps=100, guidance=3, samples=1, disable_progress=True):
        """ Genrate audio for a single prompt string. """
        with torch.no_grad():
            latents = self.model.inference([prompt], self.scheduler, steps, guidance, samples, disable_progress=disable_progress)
            mel = self.vae.decode_first_stage(latents)
            wave = self.vae.decode_to_waveform(mel)
        return wave[0]
    
    def generate_for_batch(self, prompts, steps=200, guidance=3, samples=1, batch_size=8, disable_progress=True):
        """ Genrate audio for a list of prompt strings. """
        outputs = []
        for k in tqdm(range(0, len(prompts), batch_size)):
            batch = prompts[k: k+batch_size]
            with torch.no_grad():
                latents = self.model.inference(batch, self.scheduler, steps, guidance, samples, disable_progress=disable_progress)
                mel = self.vae.decode_first_stage(latents)
                wave = self.vae.decode_to_waveform(mel)
                outputs += [item for item in wave]
        if samples == 1:
            return outputs
        else:
            return list(self.chunks(outputs, samples))

# Initialize TANGO

tango = Tango(device="cpu")
tango.vae.to(device_type)
tango.stft.to(device_type)
tango.model.to(device_type)

pipe = TangoPipeline(vae=tango.vae,
                      text_encoder=tango.model.text_encoder,
                      tokenizer=tango.model.tokenizer,
                      unet=tango.model.unet,
                      scheduler=tango.scheduler
                      )

    
@spaces.GPU(duration=120)
def gradio_generate(prompt, output_format, steps, guidance):
    output_wave = pipe(prompt,steps,guidance) ## Using pipeliine automatically uses flash attention for torch2.0 above
    #output_wave = tango.generate(prompt, steps, guidance)
    # output_filename = f"{prompt.replace(' ', '_')}_{steps}_{guidance}"[:250] + ".wav"
    output_wave = output_wave.audios[0]
    output_filename = "temp.wav"
    wavio.write(output_filename, output_wave, rate=16000, sampwidth=2)

    if (output_format == "mp3"):
        AudioSegment.from_wav("temp.wav").export("temp.mp3", format = "mp3")
        output_filename = "temp.mp3"

    return output_filename


description_text = """
<p><a href="https://huggingface.co/spaces/declare-lab/Tango-Music-AF/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/>
Generate music using Tango-Music-AF by providing a text prompt. The model was trained on a combination of MusicCaps and synthetic corpus of captions for audio.
<br/><br/> This is the demo for Tango-Music-AF for text to music generation: <a href="https://arxiv.org/pdf/2406.15487">Read our paper.</a>
<p/>
"""
# Gradio input and output components
input_text = gr.Textbox(lines=2, label="Prompt")
output_format = gr.Radio(label = "Output format", info = "The file you can dowload", choices = ["mp3", "wav"], value = "wav")
output_audio = gr.Audio(label="Generated Audio", type="filepath")
denoising_steps = gr.Slider(minimum=100, maximum=200, value=100, step=1, label="Steps", interactive=True)
guidance_scale = gr.Slider(minimum=1, maximum=10, value=3, step=0.1, label="Guidance Scale", interactive=True)

# Gradio interface
gr_interface = gr.Interface(
    fn=gradio_generate,
    inputs=[input_text, output_format, denoising_steps, guidance_scale],
    outputs=[output_audio],
    title="Improving Text-To-Audio Models with Synthetic Captions",
    description=description_text,
    allow_flagging=False,
    examples=[
        ["The song has a traditional jazzy feel to it. The piano chord progression is bouncy and light. The electric guitar has a chorus applied to it, and we hear various licks being played."],
        ["This song is a fusion of alternative and folk genres, highlighting simple yet soulful melodies and minimalist instrumentals."],
        ["The instrumental music features an ensemble that resembles the orchestra. The melody is being played by a brass section while strings provide harmonic accompaniment."],
        ["This music is instrumental. The tempo is fast with a lively keyboard harmony, steady drumming, groovy bass lines and harmonica melodic. The song is fresh, groovy, sunny, happy; vivacious and spirited."],
    ],
    cache_examples="lazy", # Turn on to cache.
)

# Launch Gradio app
gr_interface.queue(10).launch()