From d777edc605b27093a3cac5b76667577de1847310 Mon Sep 17 00:00:00 2001 From: Hui Zhang Date: Mon, 17 May 2021 18:48:51 +0800 Subject: [PATCH] ctc decoding weight 0.5 (#614) * ctc decoding weight 0.5 * tiny decoding conf * more label of mergify * format doc --- .mergify.yml | 12 ++++++++++++ examples/aishell/s1/conf/conformer.yaml | 6 +++--- examples/librispeech/s1/conf/chunk_confermer.yaml | 2 +- examples/librispeech/s1/conf/chunk_transformer.yaml | 2 +- examples/librispeech/s1/conf/conformer.yaml | 6 +++--- examples/librispeech/s1/conf/transformer.yaml | 2 +- examples/tiny/s1/conf/chunk_confermer.yaml | 2 +- examples/tiny/s1/conf/chunk_transformer.yaml | 2 +- examples/tiny/s1/conf/conformer.yaml | 2 +- examples/tiny/s1/conf/transformer.yaml | 2 +- 10 files changed, 25 insertions(+), 13 deletions(-) diff --git a/.mergify.yml b/.mergify.yml index 5a1e1ff34..b11fd5c1f 100644 --- a/.mergify.yml +++ b/.mergify.yml @@ -39,6 +39,18 @@ pull_request_rules: actions: label: remove: ["conflicts"] + - name: "auto add label=enhancement" + conditions: + - files~=^deepspeech/ + actions: + label: + add: ["enhancement"] + - name: "auto add label=Example" + conditions: + - files~=^examples/ + actions: + label: + add: ["Example"] - name: "auto add label=README" conditions: - files~=README.md diff --git a/examples/aishell/s1/conf/conformer.yaml b/examples/aishell/s1/conf/conformer.yaml index 40b04ed7b..10c3a2822 100644 --- a/examples/aishell/s1/conf/conformer.yaml +++ b/examples/aishell/s1/conf/conformer.yaml @@ -74,13 +74,13 @@ model: training: - n_epoch: 300 + n_epoch: 240 accum_grad: 2 global_grad_clip: 5.0 optim: adam optim_conf: lr: 0.002 - weight_decay: 1e-06 + weight_decay: 1e-6 scheduler: warmuplr # pytorch v1.1.0+ required scheduler_conf: warmup_steps: 25000 @@ -99,7 +99,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/librispeech/s1/conf/chunk_confermer.yaml b/examples/librispeech/s1/conf/chunk_confermer.yaml index 3ee31e1b2..0d07d0a7e 100644 --- a/examples/librispeech/s1/conf/chunk_confermer.yaml +++ b/examples/librispeech/s1/conf/chunk_confermer.yaml @@ -104,7 +104,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/librispeech/s1/conf/chunk_transformer.yaml b/examples/librispeech/s1/conf/chunk_transformer.yaml index 265e6e0b6..3939ffc68 100644 --- a/examples/librispeech/s1/conf/chunk_transformer.yaml +++ b/examples/librispeech/s1/conf/chunk_transformer.yaml @@ -97,7 +97,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/librispeech/s1/conf/conformer.yaml b/examples/librispeech/s1/conf/conformer.yaml index 1981b946f..f89f82655 100644 --- a/examples/librispeech/s1/conf/conformer.yaml +++ b/examples/librispeech/s1/conf/conformer.yaml @@ -8,7 +8,7 @@ data: spm_model_prefix: 'data/bpe_unigram_5000' mean_std_filepath: "" augmentation_config: conf/augmentation.json - batch_size: 64 + batch_size: 16 min_input_len: 0.5 # seconds max_input_len: 20.0 # seconds min_output_len: 0.0 # tokens @@ -76,7 +76,7 @@ model: training: n_epoch: 120 - accum_grad: 2 + accum_grad: 8 global_grad_clip: 5.0 optim: adam optim_conf: @@ -100,7 +100,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/librispeech/s1/conf/transformer.yaml b/examples/librispeech/s1/conf/transformer.yaml index 8ef9e12f1..9014e5b84 100644 --- a/examples/librispeech/s1/conf/transformer.yaml +++ b/examples/librispeech/s1/conf/transformer.yaml @@ -95,7 +95,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/tiny/s1/conf/chunk_confermer.yaml b/examples/tiny/s1/conf/chunk_confermer.yaml index bd4279e2b..790066264 100644 --- a/examples/tiny/s1/conf/chunk_confermer.yaml +++ b/examples/tiny/s1/conf/chunk_confermer.yaml @@ -104,7 +104,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/tiny/s1/conf/chunk_transformer.yaml b/examples/tiny/s1/conf/chunk_transformer.yaml index ba60c2735..aa2b145a6 100644 --- a/examples/tiny/s1/conf/chunk_transformer.yaml +++ b/examples/tiny/s1/conf/chunk_transformer.yaml @@ -97,7 +97,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/tiny/s1/conf/conformer.yaml b/examples/tiny/s1/conf/conformer.yaml index 83f4f5af4..3813daa04 100644 --- a/examples/tiny/s1/conf/conformer.yaml +++ b/examples/tiny/s1/conf/conformer.yaml @@ -100,7 +100,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. diff --git a/examples/tiny/s1/conf/transformer.yaml b/examples/tiny/s1/conf/transformer.yaml index 3f3170bdf..35c11731c 100644 --- a/examples/tiny/s1/conf/transformer.yaml +++ b/examples/tiny/s1/conf/transformer.yaml @@ -95,7 +95,7 @@ decoding: cutoff_prob: 1.0 cutoff_top_n: 0 num_proc_bsearch: 8 - ctc_weight: 0.0 # ctc weight for attention rescoring decode mode. + ctc_weight: 0.5 # ctc weight for attention rescoring decode mode. decoding_chunk_size: -1 # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set.