From 3d47335ab62c5ef05db4a0dacf1cc05ae45e2741 Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Tue, 15 Nov 2022 11:53:09 +0800 Subject: [PATCH] Double the duration of layer skipping warmup, from 2k to 4k. --- egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py index 3127c5d2a..053c54ce9 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py @@ -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: