From a41e93437c608f2061f72796c7260e3d5ff7bc7c Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Wed, 6 Apr 2022 12:36:58 +0800 Subject: [PATCH] Change some defaults in LR-setting rule. --- egs/librispeech/ASR/pruned_transducer_stateless2/optim.py | 2 +- egs/librispeech/ASR/pruned_transducer_stateless2/train.py | 5 ++--- 2 files changed, 3 insertions(+), 4 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py index 2b40dda45..a2e0463da 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py @@ -60,7 +60,7 @@ class Eve(Optimizer): """ def __init__(self, params, lr=1e-3, betas=(0.9, 0.98), eps=1e-8, - weight_decay=3e-4, target_rms=0.1): + weight_decay=1e-3, target_rms=0.1): if not 0.0 <= lr: raise ValueError("Invalid learning rate: {}".format(lr)) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py index 3b8f0499f..306a2195b 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py @@ -161,7 +161,7 @@ def get_parser(): parser.add_argument( "--lr-power", type=float, - default=0.5, + default=0.75, help="Power in LR-setting rule", ) @@ -780,8 +780,7 @@ def run(rank, world_size, args): optimizer = Eve( model.parameters(), - lr=params.initial_lr, betas=(0.9, 0.98), - eps=1e-9, weight_decay=3e-04, target_rms=0.1) + lr=params.initial_lr) scheduler = torch.optim.lr_scheduler.LambdaLR( optimizer, lambda step: ((params.lr_num_steps/(step + params.lr_num_steps)) ** params.lr_power))