diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.optim.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.optim.py.swp index 94f3371ee..c1a86c4de 100644 Binary files a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.optim.py.swp and b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.optim.py.swp differ diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp index a74222b6b..54fdf3b9a 100644 Binary files a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp and b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp differ diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py index 0543e1bf2..ea57db442 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py @@ -954,6 +954,8 @@ def train_one_epoch( ) loss_info.reduce(loss.device) + logging.info(loss_info) + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) loss *= numel ## normalize loss over utts(batch size)