From d389524d457a4ebd9a3f216dfb4f4ded4eceb07f Mon Sep 17 00:00:00 2001 From: Wei Kang Date: Tue, 1 Nov 2022 11:09:56 +0800 Subject: [PATCH] remove tail padding for non-streaming models (#625) --- .../ASR/pruned_transducer_stateless2/decode.py | 13 ++++++------- .../ASR/pruned_transducer_stateless3/decode.py | 13 ++++++------- .../ASR/pruned_transducer_stateless4/decode.py | 13 ++++++------- .../ASR/pruned_transducer_stateless5/decode.py | 13 ++++++------- 4 files changed, 24 insertions(+), 28 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py index 7852dafc9..3b834b919 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py @@ -380,14 +380,13 @@ def decode_one_batch( supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"].to(device) - feature_lens += params.left_context - feature = torch.nn.functional.pad( - feature, - pad=(0, 0, 0, params.left_context), - value=LOG_EPS, - ) - if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( x=feature, x_lens=feature_lens, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py index fa4f1e7d9..0f30792e3 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py @@ -462,14 +462,13 @@ def decode_one_batch( supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"].to(device) - feature_lens += params.left_context - feature = torch.nn.functional.pad( - feature, - pad=(0, 0, 0, params.left_context), - value=LOG_EPS, - ) - if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( x=feature, x_lens=feature_lens, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py index 13697008f..85097a01a 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py @@ -411,14 +411,13 @@ def decode_one_batch( supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"].to(device) - feature_lens += params.left_context - feature = torch.nn.functional.pad( - feature, - pad=(0, 0, 0, params.left_context), - value=LOG_EPS, - ) - if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( x=feature, x_lens=feature_lens, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py index 96103500b..632932214 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py @@ -378,14 +378,13 @@ def decode_one_batch( supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"].to(device) - feature_lens += params.left_context - feature = torch.nn.functional.pad( - feature, - pad=(0, 0, 0, params.left_context), - value=LOG_EPS, - ) - if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( x=feature, x_lens=feature_lens,