zchen/b2txt25
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

tpu

Browse Source
This commit is contained in:
Zchen
2025-10-14 23:11:54 +08:00
parent f67ed2b820
commit 989ba67618
2 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
model_training_nnn/rnn_model.py
View File

@@ -447,7 +447,8 @@ class TripleGRUDecoder(nn.Module):
x_processed = self._apply_preprocessing(x, day_idx)
# 3. Clean speech model processes denoised signal
denoised_input = x_processed - noise_output # Residual connection in processed space
# Ensure dtype consistency for mixed precision training in residual connection
denoised_input = x_processed - noise_output.to(x_processed.dtype) # Residual connection in processed space
# Clean speech model will apply its own preprocessing, so we pass the denoised processed data
# But we need to reverse the preprocessing first, then let clean model do its own
# Actually, it's simpler to pass the residual directly to clean model after bypassing its preprocessing
@@ -476,7 +477,8 @@ class TripleGRUDecoder(nn.Module):
x_processed = self._apply_preprocessing(x, day_idx)
# 3. Process denoised signal
denoised_input = x_processed - noise_output
# Ensure dtype consistency for mixed precision training in residual connection
denoised_input = x_processed - noise_output.to(x_processed.dtype)
clean_logits = self._clean_forward_with_processed_input(denoised_input, day_idx,
states['clean'] if states else None)