app.py

#%%
import os
os.system("git clone https://github.com/v-iashin/SpecVQGAN")
os.system("conda env create -f ./SpecVQGAN/conda_env.yml")
os.system("conda activate specvqgan")
os.system("pip install pytorch-lightning==1.2.10 omegaconf==2.0.6 streamlit==0.80 matplotlib==3.4.1 albumentations==0.5.2 SoundFile torch torchvision librosa gdown")


# %%

import sys
sys.path.append('./SpecVQGAN')
import time
from pathlib import Path

import IPython.display as display_audio
import soundfile
import torch
from IPython import display
from matplotlib import pyplot as plt
from torch.utils.data.dataloader import default_collate
from torchvision.utils import make_grid
from tqdm import tqdm

from feature_extraction.demo_utils import (ExtractResNet50, check_video_for_audio,
                                           extract_melspectrogram, load_model,
                                           show_grid, trim_video)
from sample_visualization import (all_attention_to_st, get_class_preditions,
                                  last_attention_to_st, spec_to_audio_to_st,
                                  tensor_to_plt)
from specvqgan.data.vggsound import CropImage

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# load model
model_name = '2021-07-30T21-34-25_vggsound_transformer'
log_dir = './logs'
os.chdir("./SpecVQGAN/")
config, sampler, melgan, melception = load_model(model_name, log_dir, device)
# %%

def extract_thumbnails(video_path):
  # Trim the video
  start_sec = 0  # to start with 01:35 use 95 seconds
  video_path = trim_video(video_path, start_sec, trim_duration=10)

  # Extract Features
  extraction_fps = 21.5
  feature_extractor = ExtractResNet50(extraction_fps, config.data.params, device)
  visual_features, resampled_frames = feature_extractor(video_path)

  # Show the selected frames to extract features for
  if not config.data.params.replace_feats_with_random:
      fig = show_grid(make_grid(resampled_frames))
      fig.show()

  # Prepare Input
  batch = default_collate([visual_features])
  batch['feature'] = batch['feature'].to(device)
  c = sampler.get_input(sampler.cond_stage_key, batch)
  return c, video_path

# %%
import numpy as np

def generate_audio(video_path, temperature = 1.0):
  # Define Sampling Parameters
  W_scale = 1
  mode = 'full'
  top_x = sampler.first_stage_model.quantize.n_e // 2
  update_every = 0  # use > 0 value, e.g. 15, to see the progress of generation (slows down the sampling speed)
  full_att_mat = True

  c, video_path = extract_thumbnails(video_path)

  # Start sampling
  with torch.no_grad():
      start_t = time.time()

      quant_c, c_indices = sampler.encode_to_c(c)
      # crec = sampler.cond_stage_model.decode(quant_c)

      patch_size_i = 5
      patch_size_j = 53

      B, D, hr_h, hr_w = sampling_shape = (1, 256, 5, 53*W_scale)

      z_pred_indices = torch.zeros((B, hr_h*hr_w)).long().to(device)

      if mode == 'full':
          start_step = 0
      else:
          start_step = (patch_size_j // 2) * patch_size_i
          z_pred_indices[:, :start_step] = z_indices[:, :start_step]

      pbar = tqdm(range(start_step, hr_w * hr_h), desc='Sampling Codebook Indices')
      for step in pbar:
          i = step % hr_h
          j = step // hr_h

          i_start = min(max(0, i - (patch_size_i // 2)), hr_h - patch_size_i)
          j_start = min(max(0, j - (patch_size_j // 2)), hr_w - patch_size_j)
          i_end = i_start + patch_size_i
          j_end = j_start + patch_size_j

          local_i = i - i_start
          local_j = j - j_start

          patch_2d_shape = (B, D, patch_size_i, patch_size_j)

          pbar.set_postfix(
              Step=f'({i},{j}) | Local: ({local_i},{local_j}) | Crop: ({i_start}:{i_end},{j_start}:{j_end})'
          )

          patch = z_pred_indices \
              .reshape(B, hr_w, hr_h) \
              .permute(0, 2, 1)[:, i_start:i_end, j_start:j_end].permute(0, 2, 1) \
              .reshape(B, patch_size_i * patch_size_j)

          # assuming we don't crop the conditioning and just use the whole c, if not desired uncomment the above
          cpatch = c_indices
          logits, _, attention = sampler.transformer(patch[:, :-1], cpatch)
          # remove conditioning
          logits = logits[:, -patch_size_j*patch_size_i:, :]

          local_pos_in_flat = local_j * patch_size_i + local_i
          logits = logits[:, local_pos_in_flat, :]

          logits = logits / temperature
          logits = sampler.top_k_logits(logits, top_x)

          # apply softmax to convert to probabilities
          probs = torch.nn.functional.softmax(logits, dim=-1)

          # sample from the distribution
          ix = torch.multinomial(probs, num_samples=1)
          z_pred_indices[:, j * hr_h + i] = ix

          if update_every > 0 and step % update_every == 0:
              z_pred_img = sampler.decode_to_img(z_pred_indices, sampling_shape)
              # fliping the spectrogram just for illustration purposes (low freqs to bottom, high - top)
              z_pred_img_st = tensor_to_plt(z_pred_img, flip_dims=(2,))
              display.clear_output(wait=True)
              display.display(z_pred_img_st)

              if full_att_mat:
                  att_plot = all_attention_to_st(attention, placeholders=None, scale_by_prior=True)
                  display.display(att_plot)
                  plt.close()
              else:
                  quant_z_shape = sampling_shape
                  c_length = cpatch.shape[-1]
                  quant_c_shape = quant_c.shape
                  c_att_plot, z_att_plot = last_attention_to_st(
                      attention, local_pos_in_flat, c_length, sampler.first_stage_permuter,
                      sampler.cond_stage_permuter, quant_c_shape, patch_2d_shape,
                      placeholders=None, flip_c_dims=None, flip_z_dims=(2,))
                  display.display(c_att_plot)
                  display.display(z_att_plot)
                  plt.close()
                  plt.close()
              plt.close()

      # quant_z_shape = sampling_shape
      z_pred_img = sampler.decode_to_img(z_pred_indices, sampling_shape)

      # showing the final image
      z_pred_img_st = tensor_to_plt(z_pred_img, flip_dims=(2,))
      display.clear_output(wait=True)
      display.display(z_pred_img_st)

      if full_att_mat:
          att_plot = all_attention_to_st(attention, placeholders=None, scale_by_prior=True)
          display.display(att_plot)
          plt.close()
      else:
          quant_z_shape = sampling_shape
          c_length = cpatch.shape[-1]
          quant_c_shape = quant_c.shape
          c_att_plot, z_att_plot = last_attention_to_st(
              attention, local_pos_in_flat, c_length, sampler.first_stage_permuter,
              sampler.cond_stage_permuter, quant_c_shape, patch_2d_shape,
              placeholders=None, flip_c_dims=None, flip_z_dims=(2,)
          )
          display.display(c_att_plot)
          display.display(z_att_plot)
          plt.close()
          plt.close()
      plt.close()

      print(f'Sampling Time: {time.time() - start_t:3.2f} seconds')
      waves = spec_to_audio_to_st(z_pred_img, config.data.params.spec_dir_path,
                                  config.data.params.sample_rate, show_griffin_lim=False,
                                  vocoder=melgan, show_in_st=False)
      print(f'Sampling Time (with vocoder): {time.time() - start_t:3.2f} seconds')
      print(f'Generated: {len(waves["vocoder"]) / config.data.params.sample_rate:.2f} seconds')

      # Melception opinion on the class distribution of the generated sample
      topk_preds = get_class_preditions(z_pred_img, melception)
      print(topk_preds)

  audio_path = os.path.join(log_dir, Path(video_path).stem + '.wav')
  audio = waves['vocoder']
  audio = np.repeat([audio], 2, axis=0).T
  print(audio.shape)
  soundfile.write(audio_path, audio, config.data.params.sample_rate, 'PCM_24')
  print(f'The sample has been saved @ {audio_path}')


  video_out_path = os.path.join(log_dir, Path(video_path).stem + '_audio.mp4')
  print(video_path, audio_path, video_out_path)
  os.system("ffmpeg -i %s -i %s -map 0:v -map 1:a -c:v copy -shortest %s" % (video_path, audio_path, video_out_path))
  
  return video_out_path
#  return config.data.params.sample_rate, audio

# %%
#generate_audio("../kiss.avi")
#%%
import gradio as gr

iface = gr.Interface(generate_audio, "video", ["playable_video"], 
            description="Generate audio based on the video input")
iface.launch()

# %%