Post PR commit - Also fixed #5

r9y9#53 (comment) issue solved in PyTorch 0.4

.gitignore

@@ -11,6 +11,7 @@ generated
 data
 text
 datasets
+testout
 
 # Created by https://www.gitignore.io
 
@@ -200,9 +201,3 @@ Temporary Items
 
 # Linux trash folder which might appear on any partition or disk
 .Trash-*
-vctk_preprocess/WorkingHowToUseThis.txt
-GoTBook1.01.txt
-presets/deepvoice3_got.json
-presets/deepvoice3_gotOnly.json
-presets/deepvoice3_stest.json
-presets/deepvoice3_test.json

deepvoice3_pytorch/conv.py

@@ -40,7 +40,7 @@ def incremental_forward(self, input):
                 self.input_buffer[:, :-1, :] = self.input_buffer[:, 1:, :].clone()
             # append next input
             self.input_buffer[:, -1, :] = input[:, -1, :]
-            input = torch.Tensor(self.input_buffer)
+            input = self.input_buffer.clone()
             if dilation > 1:
                 input = input[:, 0::dilation, :].contiguous()
         output = F.linear(input.view(bsz, -1), weight, self.bias)

gentle_web_align.py

@@ -3,7 +3,7 @@
 Created on Sat Apr 21 09:06:37 2018
 Phoneme alignment and conversion in HTK-style label file using Web-served Gentle
 This works on any type of english dataset.
-This allows its usage on Windows (Via Docker) and external server.
+Unlike prepare_htk_alignments_vctk.py, this is Python3 and Windows(with Docker) compatible.
 Preliminary results show that gentle has better performance with noisy dataset
 (e.g. movie extracted audioclips)
 *This work was derived from vctk_preprocess/prepare_htk_alignments_vctk.py
@@ -109,7 +109,7 @@ def gentle_request(wav_path,txt_path, server_addr, port, debug=False):
     server_addr = arguments['--server_addr']
     port = int(arguments['--port'])
     max_unalign  = float(arguments['--max_unalign'])
-    if arguments['--nested-directories'] == None:
+    if arguments['--nested-directories'] is None:
         wav_paths = sorted(glob(arguments['--wav_pattern']))
         txt_paths = sorted(glob(arguments['--txt_pattern']))    
     else:

hparams.py

@@ -125,14 +125,14 @@
     # Forced garbage collection probability
     # Use only when MemoryError continues in Windows (Disabled by default)
     #gc_probability = 0.001,
-	
-	# json_meta mode only
-	# 0: "use all",
+    
+    # json_meta mode only
+    # 0: "use all",
     # 1: "ignore only unmatched_alignment",
     # 2: "fully ignore recognition",
     ignore_recognition_level = 2,
-    min_text=20,
-    process_only_htk_aligned = False,
+    min_text=20, # when dealing with non-dedicated speech dataset(e.g. movie excerpts), setting min_text above 15 is desirable. Can be adjusted by dataset.
+    process_only_htk_aligned = False, # if true, data without phoneme alignment file(.lab) will be ignored
 )
 
 

json_meta.py

@@ -86,7 +86,7 @@ def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
         # ignore_recognition_level check
         for path in info.keys():
             is_aligned[path] = True
-            if type(info[path]) == list:
+            if isinstance(info[path], list):
                 if hparams.ignore_recognition_level == 1 and len(info[path]) == 1 or \
                         hparams.ignore_recognition_level == 2:
                     # flag the path to be 'non-aligned' text
@@ -96,7 +96,7 @@ def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
         # Reserve for future processing
         queue_count = 0
         for audio_path, text in info.items():
-            if type(text)==list:
+            if isinstance(text, list):
                 if hparams.ignore_recognition_level == 0:
                     text = text[-1]
                 else:

nikl_m.py

@@ -4,7 +4,6 @@
 import os
 import audio
 import re
-from hparams import hparams
 
 from hparams import hparams
 
@@ -73,11 +72,7 @@ def _process_utterance(out_dir, index, speaker_id, wav_path, text):
         wav = wav / np.abs(wav).max() * hparams.rescaling_max
 
     # Compute the linear-scale spectrogram from the wav:
-    try:
-        spectrogram = audio.spectrogram(wav).astype(np.float32)
-    except:
-        print(wav_path)
-        print(wav)
+    spectrogram = audio.spectrogram(wav).astype(np.float32)
     n_frames = spectrogram.shape[1]
 
     # Compute a mel-scale spectrogram from the wav:

setup.py

@@ -82,11 +82,12 @@ def create_readme_rst():
           "torch >= 0.3.0",
           "unidecode",
           "inflect",
-          "librosa == 0.5.1",
+          "librosa",
           "numba",
           "lws <= 1.0",
           "nltk",
-		  "requests",
+          "requests",
+          "PyQt5",
       ],
       extras_require={
           "train": [

train.py

@@ -59,6 +59,9 @@
 
 fs = hparams.sample_rate
 
+# Prevent Issue #5
+plt.switch_backend('Qt5Agg')
+
 global_step = 0
 global_epoch = 0
 use_cuda = torch.cuda.is_available()
@@ -916,8 +919,7 @@ def restore_parts(path, model):
     # Preventing Windows specific error such as MemoryError
     # Also reduces the occurrence of THAllocator.c 0x05 error in Widows build of PyTorch
     if platform.system() == "Windows":
-        print("Windows Detected - num_workers set to 1")
-        hparams.set_hparam('num_workers', 1)
+        print(" [!] Windows Detected - IF THAllocator.c 0x05 error occurs SET num_workers to 1")
 
     assert hparams.name == "deepvoice3"
     print(hparams_debug_string())