update

.gitignore

 .remi
 output/
 *.zip
+generation.py

generation.py

@@ -73,6 +73,8 @@ nfft = wlen + zp_percent*wlen # number of points of the DFT
 win = np.sin(np.arange(.5,wlen-.5+1)/wlen*np.pi); # sine analysis window
 
 num_batches = int(args.num_generate // args.batch_size)
+stft_gen_path = os.path.join(args.out_dir, 'stft_gen')
+os.makedirs(stft_gen_path, exist_ok=True)
 
 num_gen = 1
 for idx in range(num_batches):
@@ -80,6 +82,15 @@ for idx in range(num_batches):
         generated_sequences = net.generation_from_prior()
     seq_len, bs, x_dim = generated_sequences.shape
     for i in range(bs):
+
+        plt.figure()
+        librosa.display.specshow(librosa.power_to_db(generated_sequences[:, i, :].detach().cpu().numpy().squeeze().T), 
+            y_axis='linear', sr=fs, hop_length=hop, vmin=-40, vmax=40)
+        plt.set_cmap('magma')
+        plt.colorbar()
+        plt.title('generated spectrogram')
+        plt.savefig(os.path.join(stft_gen_path, 'generated_speech_{}.png'.format(num_gen)))
+
         generated_power = generated_sequences[:, i, :].detach().permute(1,0)
         generated_mag = torch.sqrt(generated_power).cpu().numpy()
 
@@ -93,6 +104,7 @@ for idx in range(num_batches):
         X_abs_2 = np.abs(spec_after_griffinlim)**2
 
         num_gen+=1
+    import pdb; pdb.set_trace()
 
 logger.info('Generation finished.')
 logger.info('Generated %d sequences.' % (num_gen))

script.sh

+# python train.py --exp-name VB_l4h1_inverse_stochastic_with_residual_mergedec --resume-pth ./output_backup/VB_l4h1_inverse_stochastic_with_residual_mergedec/net_best.pth --model-name 'light_dvae'
+python generation.py --exp-name VB_l4h1_inverse_stochastic_with_residual_mergedec --num-generate 7200 --batch-size 20 --seq-len 100 --resume-pth ./output_backup/VB_l4h1_inverse_stochastic_with_residual_mergedec/net_best.pth --model-name 'light_dvae'

src/utils/option.py

@@ -60,7 +60,7 @@ def get_args_parser():
     parser.add_argument('--exp-name', type=str, default='hitdvae_speech_debug', help='name of the experiment, will create a file inside out-dir')
 
     ## other
-    parser.add_argument('--model-name', type=str, default='light_dvae', choices=['light_dvae', 'hit_dvae'], help='dataset name')
+    parser.add_argument('--model-name', type=str, default='light_dvae', choices=['light_dvae', 'hit_dvae'], help='model name')
     parser.add_argument('--resume-pth', type=str, default=None, help='resume pth')
     parser.add_argument('--print-iter', type=int, default=500, help='print frequency')
     parser.add_argument('--eval-iter', type=int, default=5000, help='evaluation frequency')

train.py

@@ -20,6 +20,7 @@ from torch.utils.tensorboard import SummaryWriter
 from src.model import hit_dvae, light_dvae
 from src.dataset import speech_dataset
 from src.utils import option, optim_goal, utils
+import matplotlib.pyplot as plt
 import librosa.display