Merge pull request #178 from r9y9/fix-pytorch13

Fixes for pytorch 1.3

README.md

@@ -84,9 +84,9 @@ python synthesis.py --preset=20180505_deepvoice3_ljspeech.json \
 
 ## Requirements
 
-- Python 3 (<= 3.6)
+- Python >= 3.5
 - CUDA >= 8.0
-- PyTorch >= v0.4.0
+- PyTorch >= v1.0.0
 - [nnmnkwii](https://github.com/r9y9/nnmnkwii) >= v0.0.11
 - [MeCab](http://taku910.github.io/mecab/) (Japanese only)
 

deepvoice3_pytorch/modules.py

@@ -235,7 +235,7 @@ def get_mask_from_lengths(memory, memory_lengths):
         memory: (batch, max_time, dim)
         memory_lengths: array like
     """
-    mask = memory.data.new(memory.size(0), memory.size(1)).byte().zero_()
-    for idx, l in enumerate(memory_lengths):
-        mask[idx][:l] = 1
+    max_len = max(memory_lengths)
+    mask = torch.arange(max_len).expand(memory.size(0), max_len) < torch.tensor(memory_lengths).unsqueeze(-1)
+    mask = mask.to(memory.device)
     return ~mask

setup.py

@@ -79,7 +79,7 @@ def create_readme_rst():
       install_requires=[
           "numpy",
           "scipy",
-          "torch >= 0.4.0",
+          "torch >= 1.0.0",
           "unidecode",
           "inflect",
           "librosa",

tests/test_deepvoice3.py

@@ -26,7 +26,7 @@
 
 def _get_model(n_speakers=1, speaker_embed_dim=None,
                force_monotonic_attention=False,
-               use_decoder_state_for_postnet_input=False):
+               use_decoder_state_for_postnet_input=False, use_memory_mask=False):
     model = deepvoice3(n_vocab=n_vocab,
                        embed_dim=32,
                        mel_dim=num_mels,
@@ -42,6 +42,7 @@ def _get_model(n_speakers=1, speaker_embed_dim=None,
                        converter_channels=32,
                        force_monotonic_attention=force_monotonic_attention,
                        use_decoder_state_for_postnet_input=use_decoder_state_for_postnet_input,
+                       use_memory_mask=use_memory_mask,
                        )
     return model
 
@@ -62,7 +63,7 @@ def _test_data():
     x = torch.LongTensor(seqs)
     y = torch.rand(x.size(0), 12, 80)
 
-    return x, y
+    return x, y, input_lengths
 
 
 def _deepvoice3(n_vocab, embed_dim=256, mel_dim=80,
@@ -110,11 +111,14 @@ def _deepvoice3(n_vocab, embed_dim=256, mel_dim=80,
 
 
 def test_single_speaker_deepvoice3():
-    x, y = _test_data()
+    x, y, lengths = _test_data()
 
     for v in [False, True]:
         model = _get_model(use_decoder_state_for_postnet_input=v)
-        mel_outputs, linear_outputs, alignments, done = model(x, y)
+        mel_outputs, linear_outputs, alignments, done = model(x, y, input_lengths=lengths)
+
+    model = _get_model(use_memory_mask=True)
+    mel_outputs, linear_outputs, alignments, done = model(x, y, input_lengths=lengths)
 
 
 def _pad_2d(x, max_len, b_pad=0):
@@ -192,7 +196,7 @@ def test_incremental_correctness():
         assert max_target_len % r == 0
     mel = _pad_2d(mel, max_target_len)
     mel = torch.from_numpy(mel)
-    mel_reshaped = mel.view(1, -1, mel_dim * r)
+    mel_reshaped = mel.contiguous().view(1, -1, mel_dim * r)
     frame_positions = np.arange(1, mel_reshaped.size(1) + 1).reshape(1, mel_reshaped.size(1))
 
     x = torch.LongTensor(seqs)
@@ -269,7 +273,7 @@ def test_incremental_forward():
         assert max_target_len % r == 0
     mel = _pad_2d(mel, max_target_len)
     mel = torch.from_numpy(mel)
-    mel_reshaped = mel.view(1, -1, mel_dim * r)
+    mel_reshaped = mel.contiguous().view(1, -1, mel_dim * r)
 
     frame_positions = np.arange(1, mel_reshaped.size(1) + 1).reshape(1, mel_reshaped.size(1))
 

train.py

@@ -340,8 +340,10 @@ def collate_fn(batch):
     s, e = 1, max_decoder_target_len + 1
     # if b_pad > 0:
     #    s, e = s - 1, e - 1
+    # NOTE: needs clone to supress RuntimeError in dataloarder...
+    # ref: https://github.com/pytorch/pytorch/issues/10756
     frame_positions = torch.arange(s, e).long().unsqueeze(0).expand(
-        len(batch), max_decoder_target_len)
+        len(batch), max_decoder_target_len).clone()
 
     # done flags
     done = np.array([_pad(np.zeros(len(x[1]) // r // downsample_step - 1),
@@ -963,7 +965,7 @@ def restore_parts(path, model):
     data_loader = data_utils.DataLoader(
         dataset, batch_size=hparams.batch_size,
         num_workers=hparams.num_workers, sampler=sampler,
-        collate_fn=collate_fn, pin_memory=hparams.pin_memory)
+        collate_fn=collate_fn, pin_memory=hparams.pin_memory, drop_last=True)
 
     device = torch.device("cuda" if use_cuda else "cpu")