Update eduport_tts_mal.py
Browse files- eduport_tts_mal.py +81 -44
eduport_tts_mal.py
CHANGED
@@ -113,6 +113,7 @@ decoder_config = GPT2Config(vocab_size=len(tokenizer))
|
|
113 |
decoder_config.add_cross_attention=True
|
114 |
decoder = GPT2LMHeadModel(config=decoder_config)
|
115 |
|
|
|
116 |
class SpeechRecognitionModel(torch.nn.Module):
|
117 |
def __init__(self, encoder, decoder):
|
118 |
super().__init__()
|
@@ -120,57 +121,93 @@ class SpeechRecognitionModel(torch.nn.Module):
|
|
120 |
self.decoder = decoder
|
121 |
|
122 |
def forward(self, audio_input, text_input):
|
|
|
123 |
encoder_output = self.encoder(audio_input).last_hidden_state
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
|
130 |
-
#
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
-
#
|
135 |
-
|
136 |
-
|
|
|
137 |
|
138 |
-
|
|
|
139 |
|
140 |
-
for
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
optimizer.zero_grad()
|
145 |
|
146 |
-
|
147 |
-
audio_input = audio_input.squeeze(1).to(device)
|
148 |
-
text_input = text_input.to(device)
|
149 |
|
150 |
-
#
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
# Validation step
|
161 |
-
model.eval()
|
162 |
-
val_loss = 0
|
163 |
-
with torch.no_grad():
|
164 |
-
for audio_input, text_input in val_loader:
|
165 |
-
audio_input = audio_input.to(device)
|
166 |
-
text_input = text_input.to(device)
|
167 |
-
output = model(audio_input, text_input)
|
168 |
-
loss = torch.nn.CrossEntropyLoss()(output.logits.view(-1, output.logits.size(-1)), text_input.view(-1))
|
169 |
-
val_loss += loss.item()
|
170 |
|
171 |
-
|
172 |
-
|
173 |
|
174 |
-
|
175 |
|
176 |
-
#
|
|
|
|
113 |
decoder_config.add_cross_attention=True
|
114 |
decoder = GPT2LMHeadModel(config=decoder_config)
|
115 |
|
116 |
+
# Model Architecture with Improved FP16 Support
|
117 |
class SpeechRecognitionModel(torch.nn.Module):
|
118 |
def __init__(self, encoder, decoder):
|
119 |
super().__init__()
|
|
|
121 |
self.decoder = decoder
|
122 |
|
123 |
def forward(self, audio_input, text_input):
|
124 |
+
# Extract encoder hidden states
|
125 |
encoder_output = self.encoder(audio_input).last_hidden_state
|
126 |
+
|
127 |
+
# Create an attention mask for the encoder output
|
128 |
+
encoder_attention_mask = torch.ones(
|
129 |
+
encoder_output.shape[:2],
|
130 |
+
dtype=torch.long,
|
131 |
+
device=encoder_output.device
|
132 |
+
)
|
133 |
+
|
134 |
+
# Forward pass through the decoder with cross-attention
|
135 |
+
outputs = self.decoder(
|
136 |
+
input_ids=text_input,
|
137 |
+
encoder_hidden_states=encoder_output,
|
138 |
+
encoder_attention_mask=encoder_attention_mask
|
139 |
+
)
|
140 |
+
|
141 |
+
return outputs
|
142 |
|
143 |
+
# Training Loop with Improved Mixed Precision
|
144 |
+
def train_model(num_epochs=10):
|
145 |
+
# Prepare the models
|
146 |
+
# Use float32 for most of the model, let autocast handle precision
|
147 |
+
encoder = Wav2Vec2Model.from_pretrained('facebook/wav2vec2-base-960h')
|
148 |
+
|
149 |
+
# Modify the decoder configuration
|
150 |
+
decoder_config = GPT2Config(
|
151 |
+
vocab_size=len(tokenizer),
|
152 |
+
add_cross_attention=True
|
153 |
+
)
|
154 |
+
decoder = GPT2LMHeadModel(config=decoder_config)
|
155 |
+
|
156 |
+
# Initialize the model
|
157 |
+
model = SpeechRecognitionModel(encoder, decoder)
|
158 |
+
|
159 |
+
# Move to GPU
|
160 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
161 |
+
model = model.to(device)
|
162 |
+
|
163 |
+
# Optimizer and learning rate scheduler
|
164 |
+
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
|
165 |
+
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min')
|
166 |
+
|
167 |
+
# Gradient scaler for mixed precision training
|
168 |
+
scaler = GradScaler()
|
169 |
+
|
170 |
+
# Training loop
|
171 |
+
for epoch in range(num_epochs):
|
172 |
+
model.train()
|
173 |
+
train_loss = 0
|
174 |
+
for audio_input, text_input in train_loader:
|
175 |
+
optimizer.zero_grad()
|
176 |
+
|
177 |
+
# Move tensors to device
|
178 |
+
audio_input = audio_input.squeeze(1).to(device)
|
179 |
+
text_input = text_input.to(device)
|
180 |
|
181 |
+
# Use autocast for mixed precision training
|
182 |
+
with autocast(dtype=torch.float16):
|
183 |
+
# Forward pass
|
184 |
+
output = model(audio_input, text_input)
|
185 |
|
186 |
+
# Compute loss
|
187 |
+
loss = torch.nn.CrossEntropyLoss()(output.logits.view(-1, output.logits.size(-1)), text_input.view(-1))
|
188 |
|
189 |
+
# Scaled loss for mixed precision
|
190 |
+
scaler.scale(loss).backward()
|
191 |
+
scaler.step(optimizer)
|
192 |
+
scaler.update()
|
|
|
193 |
|
194 |
+
train_loss += loss.item()
|
|
|
|
|
195 |
|
196 |
+
# Validation step
|
197 |
+
model.eval()
|
198 |
+
val_loss = 0
|
199 |
+
with torch.no_grad(), autocast(dtype=torch.float16):
|
200 |
+
for audio_input, text_input in val_loader:
|
201 |
+
audio_input = audio_input.to(device)
|
202 |
+
text_input = text_input.to(device)
|
203 |
+
output = model(audio_input, text_input)
|
204 |
+
loss = torch.nn.CrossEntropyLoss()(output.logits.view(-1, output.logits.size(-1)), text_input.view(-1))
|
205 |
+
val_loss += loss.item()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
206 |
|
207 |
+
# Update scheduler
|
208 |
+
scheduler.step(val_loss)
|
209 |
|
210 |
+
print(f'Epoch {epoch}: Train Loss: {train_loss / len(train_loader)}, Val Loss: {val_loss / len(val_loader)}')
|
211 |
|
212 |
+
# Run the training
|
213 |
+
train_model()
|