From 4988c815c9ba0e9f3fa5b5ffce8a33d0c989ab03 Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Fri, 11 Nov 2022 22:56:14 +0800 Subject: [PATCH] Use more attention heads in slowest layer. --- egs/librispeech/ASR/pruned_transducer_stateless7/train.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py index 79860ffbc..af7503ada 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py @@ -127,7 +127,7 @@ def add_model_arguments(parser: argparse.ArgumentParser): parser.add_argument( "--num-heads", type=str, - default="8", + default="8,8,8,16,8,8", help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", )