mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add different debug info.

Browse Source
This commit is contained in:
Daniel Povey 2022-10-14 21:16:23 +08:00
parent 90953537ad
commit 1812f6cb28
1 changed files with 8 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py
View File

@ -1248,8 +1248,14 @@ class RelPositionMultiheadAttention(nn.Module):
attn_output = attn_output - attn_output_mean
attn_covar = torch.matmul(attn_output.transpose(1, 2), attn_output) / (bsz * seq_len)
# attn_covar: (num_heads, head_dim, head_dim)
eigs, _ = torch.symeig(attn_covar)
logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}")
#eigs, _ = torch.symeig(attn_covar)
#logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}")
attn_covar = attn_covar.mean(dim=1).sum(dim=1) # (num_heads,)
embed_dim = self.in_proj2.weight.shape[1]
in_proj_covar = (self.in_proj2.weight.reshape(num_heads, head_dim, embed_dim) ** 2).mean(dim=(1,2))
out_proj_covar = (self.out_proj2.weight.reshape(embed_dim, num_heads, head_dim) ** 2).mean(dim=(0,2))
logging.info(f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}")
class FeedforwardModule(nn.Module):