🏄 Fix MultiGPU utt_ids (#309).

Author: dathudeptrai

examples/fastspeech/train_fastspeech.py

@@ -120,20 +120,22 @@ def generate_and_save_intermediate_result(self, batch):
 
         mels_before, mels_after, *_ = outputs
         mel_gts = batch["mel_gts"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             mels_before = mels_before.values[0].numpy()
             mels_after = mels_after.values[0].numpy()
             mel_gts = mel_gts.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             mels_before = mels_before.numpy()
             mels_after = mels_after.numpy()
             mel_gts = mel_gts.numpy()
+            utt_ids = utt_ids.numpy()
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)

examples/fastspeech2/train_fastspeech2.py

@@ -132,20 +132,22 @@ def generate_and_save_intermediate_result(self, batch):
 
         mels_before, mels_after, *_ = outputs
         mel_gts = batch["mel_gts"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             mels_before = mels_before.values[0].numpy()
             mels_after = mels_after.values[0].numpy()
             mel_gts = mel_gts.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             mels_before = mels_before.numpy()
             mels_after = mels_after.numpy()
             mel_gts = mel_gts.numpy()
+            utt_ids = utt_ids.numpy()
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)

examples/fastspeech2_libritts/train_fastspeech2.py

@@ -142,17 +142,20 @@ def generate_and_save_intermediate_result(self, batch):
 
         mels_before, mels_after, *_ = outputs
         mel_gts = batch["mel_gts"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             mels_before = mels_before.values[0].numpy()
             mels_after = mels_after.values[0].numpy()
             mel_gts = mel_gts.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             mels_before = mels_before.numpy()
             mels_after = mels_after.numpy()
             mel_gts = mel_gts.numpy()
+            utt_ids = utt_ids.numpy()
 
         # check directory
         if self.use_griffin:
@@ -167,22 +170,25 @@ def generate_and_save_intermediate_result(self, batch):
         for idx, (mel_gt, mel_before, mel_after) in enumerate(
             zip(mel_gts, mels_before, mels_after), 1
         ):
-
+            
+            
             if self.use_griffin:
+                utt_id = utt_ids[idx]
                 grif_before = self.griffin_lim_tf(tf.reshape(mel_before, [-1, 80])[tf.newaxis, :], n_iter=32)
                 grif_after = self.griffin_lim_tf(tf.reshape(mel_after, [-1, 80])[tf.newaxis, :], n_iter=32)
                 grif_gt = self.griffin_lim_tf(tf.reshape(mel_gt, [-1, 80])[tf.newaxis, :], n_iter=32)
-                self.griffin_lim_tf.save_wav(grif_before, griff_dir_name, f"{idx}_before")
-                self.griffin_lim_tf.save_wav(grif_after, griff_dir_name, f"{idx}_after")
-                self.griffin_lim_tf.save_wav(grif_gt, griff_dir_name, f"{idx}_gt")
-
+                self.griffin_lim_tf.save_wav(grif_before, griff_dir_name, f"{utt_id}_before")
+                self.griffin_lim_tf.save_wav(grif_after, griff_dir_name, f"{utt_id}_after")
+                self.griffin_lim_tf.save_wav(grif_gt, griff_dir_name, f"{utt_id}_gt")
+            
+            utt_id = utt_ids[idx]
             mel_gt = tf.reshape(mel_gt, (-1, 80)).numpy()  # [length, 80]
             mel_before = tf.reshape(mel_before, (-1, 80)).numpy()  # [length, 80]
             mel_after = tf.reshape(mel_after, (-1, 80)).numpy()  # [length, 80]
 
 
             # plit figure and save it
-            figname = os.path.join(dirname, f"{idx}.png")
+            figname = os.path.join(dirname, f"{utt_id}.png")
             fig = plt.figure(figsize=(10, 8))
             ax1 = fig.add_subplot(311)
             ax2 = fig.add_subplot(312)

examples/melgan/train_melgan.py

@@ -189,18 +189,20 @@ def generate_and_save_intermediate_result(self, batch):
         # generate
         y_batch_ = self.one_step_predict(batch)
         y_batch = batch["audios"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             y_batch_ = y_batch_.values[0].numpy()
             y_batch = y_batch.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             y_batch_ = y_batch_.numpy()
             y_batch = y_batch.numpy()
+            utt_ids = utt_ids.numpy()
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)

examples/multiband_melgan/train_multiband_melgan.py

@@ -203,20 +203,22 @@ def generate_and_save_intermediate_result(self, batch):
 
         y_mb_batch_ = self.one_step_predict(batch)  # [B, T // subbands, subbands]
         y_batch = batch["audios"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             y_mb_batch_ = y_mb_batch_.values[0].numpy()
             y_batch = y_batch.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             y_mb_batch_ = y_mb_batch_.numpy()
             y_batch = y_batch.numpy()
+            utt_ids = utt_ids.numpy()
 
         y_batch_ = self.pqmf.synthesis(y_mb_batch_).numpy()  # [B, T, 1]
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)

examples/multiband_pwgan/train_multiband_pwgan.py

@@ -217,20 +217,22 @@ def generate_and_save_intermediate_result(self, batch):
 
         y_mb_batch_ = self.one_step_predict(batch)  # [B, T // subbands, subbands]
         y_batch = batch["audios"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             y_mb_batch_ = y_mb_batch_.values[0].numpy()
             y_batch = y_batch.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             y_mb_batch_ = y_mb_batch_.numpy()
             y_batch = y_batch.numpy()
+            utt_ids = utt_ids.numpy()
 
         y_batch_ = self.pqmf.synthesis(y_mb_batch_).numpy()  # [B, T, 1]
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)

examples/parallel_wavegan/train_parallel_wavegan.py

@@ -189,18 +189,20 @@ def generate_and_save_intermediate_result(self, batch):
         # generate
         y_batch_ = self.one_step_predict(batch)
         y_batch = batch["audios"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
         try:
             y_batch_ = y_batch_.values[0].numpy()
             y_batch = y_batch.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             y_batch_ = y_batch_.numpy()
             y_batch = y_batch.numpy()
+            utt_ids = utt_ids.numpy()
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)

examples/tacotron2/train_tacotron2.py

@@ -197,6 +197,7 @@ def generate_and_save_intermediate_result(self, batch):
             alignment_historys,
         ) = outputs
         mel_gts = batch["mel_gts"]
+        utt_ids = batch["utt_ids"]
 
         # convert to tensor.
         # here we just take a sample at first replica.
@@ -205,14 +206,15 @@ def generate_and_save_intermediate_result(self, batch):
             mels_after = mel_outputs.values[0].numpy()
             mel_gts = mel_gts.values[0].numpy()
             alignment_historys = alignment_historys.values[0].numpy()
+            utt_ids = utt_ids.values[0].numpy()
         except Exception:
             mels_before = decoder_output.numpy()
             mels_after = mel_outputs.numpy()
             mel_gts = mel_gts.numpy()
             alignment_historys = alignment_historys.numpy()
+            utt_ids = utt_ids.numpy()
 
         # check directory
-        utt_ids = batch["utt_ids"].numpy()
         dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
         if not os.path.exists(dirname):
             os.makedirs(dirname)