From 23b3aa233c792de86fb23b04f4a0160ba74f4d51 Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Fri, 4 Mar 2022 00:42:37 +0800 Subject: [PATCH] Double learning rate of exp-scale units --- egs/librispeech/ASR/conformer_ctc/subsampling.py | 4 ++-- egs/librispeech/ASR/transducer_stateless/conformer.py | 4 ++-- egs/librispeech/ASR/transducer_stateless/train.py | 2 +- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/egs/librispeech/ASR/conformer_ctc/subsampling.py b/egs/librispeech/ASR/conformer_ctc/subsampling.py index 73493a7ea..3b35c2ebe 100644 --- a/egs/librispeech/ASR/conformer_ctc/subsampling.py +++ b/egs/librispeech/ASR/conformer_ctc/subsampling.py @@ -48,12 +48,12 @@ class Conv2dSubsampling(nn.Module): in_channels=1, out_channels=odim, kernel_size=3, stride=2 ), nn.ReLU(), - ExpScale(odim, 1, 1, speed=2.0), + ExpScale(odim, 1, 1, speed=4.0), nn.Conv2d( in_channels=odim, out_channels=odim, kernel_size=3, stride=2 ), nn.ReLU(), - ExpScale(odim, 1, 1, speed=2.0), + ExpScale(odim, 1, 1, speed=4.0), ) self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim) self.out_norm = nn.LayerNorm(odim, elementwise_affine=False) diff --git a/egs/librispeech/ASR/transducer_stateless/conformer.py b/egs/librispeech/ASR/transducer_stateless/conformer.py index 3789e02fd..59f317e90 100644 --- a/egs/librispeech/ASR/transducer_stateless/conformer.py +++ b/egs/librispeech/ASR/transducer_stateless/conformer.py @@ -157,7 +157,7 @@ class ConformerEncoderLayer(nn.Module): self.feed_forward = nn.Sequential( nn.Linear(d_model, dim_feedforward), Swish(), - ExpScale(dim_feedforward, speed=2.0), + ExpScale(dim_feedforward, speed=4.0), nn.Dropout(dropout), nn.Linear(dim_feedforward, d_model), ) @@ -165,7 +165,7 @@ class ConformerEncoderLayer(nn.Module): self.feed_forward_macaron = nn.Sequential( nn.Linear(d_model, dim_feedforward), Swish(), - ExpScale(dim_feedforward, speed=2.0), + ExpScale(dim_feedforward, speed=4.0), nn.Dropout(dropout), nn.Linear(dim_feedforward, d_model), ) diff --git a/egs/librispeech/ASR/transducer_stateless/train.py b/egs/librispeech/ASR/transducer_stateless/train.py index 183a924c6..a1ded87c6 100755 --- a/egs/librispeech/ASR/transducer_stateless/train.py +++ b/egs/librispeech/ASR/transducer_stateless/train.py @@ -110,7 +110,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="transducer_stateless/specaugmod_baseline_randcombine1_expscale", + default="transducer_stateless/specaugmod_baseline_randcombine1_expscale2", help="""The experiment dir. It specifies the directory where all training related files, e.g., checkpoints, log, etc, are saved