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

Mask larger regions

Browse Source
This commit is contained in:
Daniel Povey 2023-03-31 17:07:22 +08:00
parent d41b73000e
commit cd0f48f508
1 changed files with 9 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
View File

@ -292,7 +292,14 @@ class Zipformer2(EncoderInterface):
assert self.encoder_dim[0] == _encoder_dims0
max_downsampling_factor = max(self.downsampling_factor)
# this is a heuristic, the regions over which we masked frames are 4 times the
# largest of the downsampling_factors. Can be any integer >= 1.
downsampling_multiple = 4
downsampling_factor = [ downsampling_multiple * i for i in self.downsampling_factor ]
max_downsampling_factor = max(downsampling_factor)
num_frames_max = (num_frames0 + max_downsampling_factor - 1)
@ -318,7 +325,7 @@ class Zipformer2(EncoderInterface):
feature_masks = []
for i in range(num_encoders):
ds = self.downsampling_factor[i]
ds = downsampling_factor[i]
upsample_factor = (max_downsampling_factor // ds)
frame_mask = (frame_mask_max.unsqueeze(1).expand(num_frames_max, upsample_factor,