Bug fixes

.compute

@@ -1,3 +1,3 @@
 #!/bin/bash
 source ../tmp/venv/bin/activate
-python train.py --config_path config.json
+python train.py --config_path config.json --debug true

config.json

@@ -1,11 +1,13 @@
 {
-    "model_name": "best-model",
+    "model_name": "audio-update-l2-loss",
     "num_mels": 80,
     "num_freq": 1025,
     "sample_rate": 20000,
     "frame_length_ms": 50,
     "frame_shift_ms": 12.5,
     "preemphasis": 0.97,
+    "min_mel_freq": 125,
+    "max_mel_freq": 7600,
     "min_level_db": -100,
     "ref_level_db": 20,
     "embedding_size": 256,
@@ -25,7 +27,7 @@
 
     "checkpoint": true,
     "save_step": 376,
-    "print_step": 50,
+    "print_step": 10,
     "data_path": "/snakepit/shared/data/keithito/LJSpeech-1.1/",
     "min_seq_len": 0,
     "output_path": "experiments/"

layers/attention.py

@@ -38,11 +38,11 @@ class LocationSensitiveAttention(nn.Module):
         super(LocationSensitiveAttention, self).__init__()
         self.kernel_size = kernel_size
         self.filters = filters
-        padding = int((kernel - 1) / 2)
-        self.loc_conv =  nn.Conv1d(2, filters,
+        padding = int((kernel_size - 1) / 2)
+        self.loc_conv =  nn.Conv1d(1, filters,
                                    kernel_size=kernel_size, stride=1,
                                    padding=padding, bias=False)
-        self.loc_linear = nn.Linear(loc_dim, hidden_dim)
+        self.loc_linear = nn.Linear(filters, hidden_dim)
         self.query_layer = nn.Linear(query_dim, hidden_dim, bias=True)
         self.annot_layer = nn.Linear(annot_dim, hidden_dim, bias=True)
         self.v = nn.Linear(hidden_dim, 1, bias=False)
@@ -79,7 +79,7 @@ class AttentionRNNCell(nn.Module):
             memory_dim (int): memory vector (decoder autogression) feature dimension.
             align_model (str): 'b' for Bahdanau, 'ls' Location Sensitive alignment.
         """
-        super(AttentionRNN, self).__init__()
+        super(AttentionRNNCell, self).__init__()
         self.align_model = align_model
         self.rnn_cell = nn.GRUCell(out_dim + memory_dim, out_dim)
         # pick bahdanau or location sensitive attention

layers/tacotron.py

@@ -275,7 +275,7 @@ class Decoder(nn.Module):
                                       # dim=1)
             attention_rnn_hidden, current_context_vec, attention = self.attention_rnn(
                 processed_memory, current_context_vec, attention_rnn_hidden,
-                inputs, attention, input_lens)
+                inputs, attention.unsqueeze(1), input_lens)
             # attention_cum += attention
             # Concat RNN output and attention context vector
             decoder_input = self.project_to_decoder_in(

train.py

@@ -25,7 +25,7 @@ from utils.generic_utils import (Progbar, remove_experiment_folder,
 from utils.visual import plot_alignment, plot_spectrogram
 from datasets.LJSpeech import LJSpeechDataset
 from models.tacotron import Tacotron
-from layers.losses import L2LossMasked
+from layers.losses import L1LossMasked
 
 torch.manual_seed(1)
 use_cuda = torch.cuda.is_available()
@@ -338,6 +338,7 @@ def evaluate(model, criterion, criterion_st, data_loader, current_step):
 
 def main(args):
 
+    # Setup the dataset
     # Setup the dataset
     train_dataset = LJSpeechDataset(os.path.join(c.data_path, 'metadata_train.csv'),
                                     os.path.join(c.data_path, 'wavs'),

utils/generic_utils.py

@@ -37,7 +37,7 @@ def get_commit_hash():
 
 def create_experiment_folder(root_path, model_name, debug):
     """ Create a folder with the current date and time """
-    date_str = datetime.datetime.now().strftime("%B-%d-%Y_%I:%M%p")
+    date_str = datetime.datetime.now().strftime("%B-%d-%Y_%I+%M%p")
     if debug:
         commit_hash = 'debug'
     else: