Double the duration of layer skipping warmup, from 2k to 4k.

This commit is contained in:
Daniel Povey 2022-11-15 11:53:09 +08:00
parent 22a1401f36
commit 3d47335ab6

View File

@ -365,8 +365,8 @@ class ZipformerEncoderLayer(nn.Module):
# layer_skip_prob will be overwritten to change warmup begin and end times.
# treating batch_index == 0.0 specially is just to get scan_pessimistic_batches_for_oom()
# to work correctly.
layer_skip_prob: FloatLike = ScheduledFloat((0.0, 0.5), (2000.0, 0.05), default=0),
dynamic_skip_prob: FloatLike = ScheduledFloat((0.0, 0.2), (2000.0, 0.0), default=0),
layer_skip_prob: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.05), default=0),
dynamic_skip_prob: FloatLike = ScheduledFloat((0.0, 0.2), (4000.0, 0.0), default=0),
bypass_min: FloatLike = ScheduledFloat((0.0, 0.75), (20000.0, 0.25), default=0),
bypass_max: FloatLike = 1.0,
) -> None: