fix some bugs of transformer_tts and fastspeech.

examples/fastspeech/alignments/get_alignments.py

@@ -115,15 +115,10 @@ def alignments(args):
             mel_input = fluid.layers.unsqueeze(dg.to_variable(mel_input), [0])
             mel_lens = mel_input.shape[1]
 
-            dec_slf_mask = get_triu_tensor(mel_input,
-                                           mel_input).astype(np.float32)
-            dec_slf_mask = np.expand_dims(dec_slf_mask, axis=0)
-            dec_slf_mask = fluid.layers.cast(
-                dg.to_variable(dec_slf_mask != 0), np.float32) * (-2**32 + 1)
             pos_mel = np.arange(1, mel_input.shape[1] + 1)
             pos_mel = fluid.layers.unsqueeze(dg.to_variable(pos_mel), [0])
             mel_pred, postnet_pred, attn_probs, stop_preds, attn_enc, attn_dec = model(
-                text, mel_input, pos_text, pos_mel, dec_slf_mask)
+                text, mel_input, pos_text, pos_mel)
             mel_input = fluid.layers.concat(
                 [mel_input, postnet_pred[:, -1:, :]], axis=1)
 

examples/fastspeech/configs/ljspeech.yaml

@@ -29,5 +29,5 @@ train:
   grad_clip_thresh: 0.1 #the threshold of grad clip.
 
   checkpoint_interval: 1000
-  max_epochs: 10000
+  max_iteration: 500000
 

examples/fastspeech/train.py

@@ -62,7 +62,8 @@ def main(args):
         cfg = yaml.load(f, Loader=yaml.Loader)
 
     global_step = 0
-    place = fluid.CUDAPlace(local_rank) if args.use_gpu else fluid.CPUPlace()
+    place = fluid.CUDAPlace(dg.parallel.Env()
+                            .dev_id) if args.use_gpu else fluid.CPUPlace()
     fluid.enable_dygraph(place)
 
     if not os.path.exists(args.output):
@@ -88,7 +89,8 @@ def main(args):
         cfg['train']['batch_size'],
         nranks,
         local_rank,
-        shuffle=True).reader()
+        shuffle=True).reader
+    iterator = iter(tqdm(reader))
 
     # Load parameters.
     global_step = io.load_parameters(
@@ -103,12 +105,14 @@ def main(args):
         strategy = dg.parallel.prepare_context()
         model = fluid.dygraph.parallel.DataParallel(model, strategy)
 
-    for epoch in range(cfg['train']['max_epochs']):
+    while global_step <= cfg['train']['max_iteration']:
-        pbar = tqdm(reader)
+        try:
+            batch = next(iterator)
+        except StopIteration as e:
+            iterator = iter(tqdm(reader))
+            batch = next(iterator)
 
-        for i, data in enumerate(pbar):
+        (character, mel, pos_text, pos_mel, alignment) = batch
-            pbar.set_description('Processing at epoch %d' % epoch)
-            (character, mel, pos_text, pos_mel, alignment) = data
 
         global_step += 1
 
@@ -120,8 +124,7 @@ def main(args):
         mel_postnet_loss = layers.mse_loss(mel_output_postnet, mel)
         duration_loss = layers.mean(
             layers.abs(
-                    layers.elementwise_sub(duration_predictor_output,
+                layers.elementwise_sub(duration_predictor_output, alignment)))
-                                           alignment)))
         total_loss = mel_loss + mel_postnet_loss + duration_loss
 
         if local_rank == 0:
@@ -147,8 +150,8 @@ def main(args):
         if local_rank == 0 and global_step % cfg['train'][
                 'checkpoint_interval'] == 0:
             io.save_parameters(
-                    os.path.join(args.output, 'checkpoints'), global_step,
+                os.path.join(args.output, 'checkpoints'), global_step, model,
-                    model, optimizer)
+                optimizer)
 
     if local_rank == 0:
         writer.close()

examples/transformer_tts/README.md

@@ -53,7 +53,7 @@ During synthesis, results are saved in `${output}/samples` and tensorboard log i
 TransformerTTS model can be trained by running ``train_transformer.py``.
 
 ```bash
-python train_trasformer.py \
+python train_transformer.py \
 --use_gpu=1 \
 --data=${DATAPATH} \
 --output='./experiment' \

examples/transformer_tts/configs/ljspeech.yaml

@@ -31,7 +31,7 @@ train:
   checkpoint_interval: 1000
   image_interval: 2000
 
-  max_epochs: 10000
+  max_iteration: 500000
 
 
 

examples/transformer_tts/train_transformer.py

@@ -102,16 +102,21 @@ def main(args):
         cfg['train']['batch_size'],
         nranks,
         local_rank,
-        shuffle=True).reader()
+        shuffle=True).reader
 
-    for epoch in range(cfg['train']['max_epochs']):
+    iterator = iter(tqdm(reader))
-        pbar = tqdm(reader)
-        for i, data in enumerate(pbar):
-            pbar.set_description('Processing at epoch %d' % epoch)
-            character, mel, mel_input, pos_text, pos_mel = data
 
     global_step += 1
 
+    while global_step <= cfg['train']['max_iteration']:
+        try:
+            batch = next(iterator)
+        except StopIteration as e:
+            iterator = iter(tqdm(reader))
+            batch = next(iterator)
+
+        character, mel, mel_input, pos_text, pos_mel = batch
+
         mel_pred, postnet_pred, attn_probs, stop_preds, attn_enc, attn_dec = model(
             character, mel_input, pos_text, pos_mel)
 
@@ -134,8 +139,7 @@ def main(args):
             }, global_step)
 
             if cfg['network']['stop_token']:
-                    writer.add_scalar('stop_loss',
+                writer.add_scalar('stop_loss', stop_loss.numpy(), global_step)
-                                      stop_loss.numpy(), global_step)
 
             if parallel:
                 writer.add_scalars('alphas', {
@@ -157,7 +161,7 @@ def main(args):
                     for j in range(cfg['network']['decoder_num_head']):
                         x = np.uint8(
                             cm.viridis(prob.numpy()[j * cfg['train'][
-                                    'batch_size'] // 2]) * 255)
+                                'batch_size'] // nranks]) * 255)
                         writer.add_image(
                             'Attention_%d_0' % global_step,
                             x,
@@ -168,7 +172,7 @@ def main(args):
                     for j in range(cfg['network']['encoder_num_head']):
                         x = np.uint8(
                             cm.viridis(prob.numpy()[j * cfg['train'][
-                                    'batch_size'] // 2]) * 255)
+                                'batch_size'] // nranks]) * 255)
                         writer.add_image(
                             'Attention_enc_%d_0' % global_step,
                             x,
@@ -179,7 +183,7 @@ def main(args):
                     for j in range(cfg['network']['decoder_num_head']):
                         x = np.uint8(
                             cm.viridis(prob.numpy()[j * cfg['train'][
-                                    'batch_size'] // 2]) * 255)
+                                'batch_size'] // nranks]) * 255)
                         writer.add_image(
                             'Attention_dec_%d_0' % global_step,
                             x,
@@ -199,8 +203,9 @@ def main(args):
         if local_rank == 0 and global_step % cfg['train'][
                 'checkpoint_interval'] == 0:
             io.save_parameters(
-                    os.path.join(args.output, 'checkpoints'), global_step,
+                os.path.join(args.output, 'checkpoints'), global_step, model,
-                    model, optimizer)
+                optimizer)
+        global_step += 1
 
     if local_rank == 0:
         writer.close()

parakeet/models/fastspeech/length_regulator.py

@@ -94,7 +94,8 @@ class LengthRegulator(dg.Layer):
         else:
             duration_predictor_output = layers.round(duration_predictor_output)
             output = self.LR(x, duration_predictor_output, alpha)
-            mel_pos = dg.to_variable(np.arange(1, output.shape[1] + 1))
+            mel_pos = dg.to_variable(np.arange(1, output.shape[1] + 1)).astype(
+                np.int64)
             mel_pos = layers.unsqueeze(mel_pos, [0])
             return output, mel_pos