diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py index 5de7102d3..aa345dd84 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py @@ -927,7 +927,8 @@ def run(rank, world_size, args): model = DDP(model, device_ids=[rank]) optimizer = ScaledAdam(model.parameters(), - lr=params.initial_lr) + lr=params.initial_lr, + clipping_scale=2.0) scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)