upload files

README.md

@@ -0,0 +1,5 @@
+Introduction
+See https://github.com/k2-fsa/icefall/pull/330
+and https://github.com/k2-fsa/icefall/pull/452
+
+It has random combiner inside.

data/lang_bpe_500/bpe.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c53433de083c4a6ad12d034550ef22de68cec62c4f58932a7b6b8b2f1e743fa5
+size 244865

decoding_results/fast_beam_search/log-decode-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model-2022-07-07-18-16-33

@@ -0,0 +1,27 @@
+2022-07-07 18:16:33,814 INFO [decode.py:645] Decoding started
+2022-07-07 18:16:33,814 INFO [decode.py:651] Device: cuda:0
+2022-07-07 18:16:33,816 INFO [decode.py:661] {'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampling_factor': 4, 'model_warm_step': 3000, 'env_info': {'k2-version': '1.15.1', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '350e449fe5d7b4231f77eb0add764782eed9f5d2', 'k2-git-date': 'Thu May 26 22:56:24 2022', 'lhotse-version': '1.4.0.dev+git.21b5b3f.dirty', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'rand_combine', 'icefall-git-sha1': '11f96e2-dirty', 'icefall-git-date': 'Thu Jun 30 12:11:09 2022', 'icefall-path': '/ceph-zw/workspace/experiments/icefall_rand_combine', 'k2-path': '/ceph-zw/workspace/share/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-jb/yaozengwei/workspace/rnnt/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-2-0602201035-5fb6d86964-mclm7', 'IP address': '10.177.74.202'}, 'epoch': 30, 'iter': 0, 'avg': 10, 'use_averaged_model': True, 'exp_dir': PosixPath('pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4'), 'bpe_model': 'data/lang_bpe_500/bpe.model', 'lang_dir': PosixPath('data/lang_bpe_500'), 'decoding_method': 'fast_beam_search', 'beam_size': 4, 'beam': 20.0, 'ngram_lm_scale': 0.01, 'max_contexts': 8, 'max_states': 64, 'context_size': 2, 'max_sym_per_frame': 1, 'num_paths': 200, 'nbest_scale': 0.5, 'num_encoder_layers': 18, 'dim_feedforward': 1024, 'nhead': 4, 'encoder_dim': 256, 'decoder_dim': 512, 'joiner_dim': 512, 'full_libri': True, 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 600, 'bucketing_sampler': True, 'num_buckets': 30, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'drop_last': True, 'return_cuts': True, 'num_workers': 2, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'input_strategy': 'PrecomputedFeatures', 'res_dir': PosixPath('pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/fast_beam_search'), 'suffix': 'epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model', 'blank_id': 0, 'unk_id': 2, 'vocab_size': 500}
+2022-07-07 18:16:33,817 INFO [decode.py:663] About to create model
+2022-07-07 18:16:33,990 INFO [decode.py:730] Calculating the averaged model over epoch range from 20 (excluded) to 30
+2022-07-07 18:16:39,275 INFO [decode.py:766] Number of model parameters: 30501996
+2022-07-07 18:16:39,275 INFO [asr_datamodule.py:444] About to get test-clean cuts
+2022-07-07 18:16:39,277 INFO [asr_datamodule.py:451] About to get test-other cuts
+2022-07-07 18:16:41,081 INFO [decode.py:548] batch 0/?, cuts processed until now is 47
+2022-07-07 18:17:01,503 INFO [decode.py:548] batch 20/?, cuts processed until now is 1656
+2022-07-07 18:17:18,296 INFO [decode.py:565] The transcripts are stored in pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/fast_beam_search/recogs-test-clean-beam_20.0_max_contexts_8_max_states_64-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2022-07-07 18:17:18,363 INFO [utils.py:420] [test-clean-beam_20.0_max_contexts_8_max_states_64] %WER 2.81% [1478 / 52576, 158 ins, 125 del, 1195 sub ]
+2022-07-07 18:17:18,525 INFO [decode.py:578] Wrote detailed error stats to pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/fast_beam_search/errs-test-clean-beam_20.0_max_contexts_8_max_states_64-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2022-07-07 18:17:18,526 INFO [decode.py:595] 
+For test-clean, WER of different settings are:
+beam_20.0_max_contexts_8_max_states_64	2.81	best for test-clean
+
+2022-07-07 18:17:19,990 INFO [decode.py:548] batch 0/?, cuts processed until now is 57
+2022-07-07 18:17:39,280 INFO [decode.py:548] batch 20/?, cuts processed until now is 1886
+2022-07-07 18:17:55,331 INFO [decode.py:565] The transcripts are stored in pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/fast_beam_search/recogs-test-other-beam_20.0_max_contexts_8_max_states_64-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2022-07-07 18:17:55,403 INFO [utils.py:420] [test-other-beam_20.0_max_contexts_8_max_states_64] %WER 6.76% [3539 / 52343, 354 ins, 327 del, 2858 sub ]
+2022-07-07 18:17:55,630 INFO [decode.py:578] Wrote detailed error stats to pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/fast_beam_search/errs-test-other-beam_20.0_max_contexts_8_max_states_64-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2022-07-07 18:17:55,630 INFO [decode.py:595] 
+For test-other, WER of different settings are:
+beam_20.0_max_contexts_8_max_states_64	6.76	best for test-other
+
+2022-07-07 18:17:55,630 INFO [decode.py:795] Done!

decoding_results/fast_beam_search/wer-summary-test-clean-beam_20.0_max_contexts_8_max_states_64-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+beam_20.0_max_contexts_8_max_states_64	2.81

decoding_results/fast_beam_search/wer-summary-test-other-beam_20.0_max_contexts_8_max_states_64-epoch-30-avg-10-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+beam_20.0_max_contexts_8_max_states_64	6.76

decoding_results/greedy_search/log-decode-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model-2022-07-07-19-14-17

@@ -0,0 +1,25 @@
+2022-07-07 19:14:17,821 INFO [decode.py:645] Decoding started
+2022-07-07 19:14:17,821 INFO [decode.py:651] Device: cuda:0
+2022-07-07 19:14:17,823 INFO [decode.py:661] {'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampling_factor': 4, 'model_warm_step': 3000, 'env_info': {'k2-version': '1.15.1', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '350e449fe5d7b4231f77eb0add764782eed9f5d2', 'k2-git-date': 'Thu May 26 22:56:24 2022', 'lhotse-version': '1.4.0.dev+git.21b5b3f.dirty', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'rand_combine', 'icefall-git-sha1': '11f96e2-dirty', 'icefall-git-date': 'Thu Jun 30 12:11:09 2022', 'icefall-path': '/ceph-zw/workspace/experiments/icefall_rand_combine', 'k2-path': '/ceph-zw/workspace/share/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-jb/yaozengwei/workspace/rnnt/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-2-0602201035-5fb6d86964-mclm7', 'IP address': '10.177.74.202'}, 'epoch': 30, 'iter': 0, 'avg': 10, 'use_averaged_model': True, 'exp_dir': PosixPath('pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4'), 'bpe_model': 'data/lang_bpe_500/bpe.model', 'lang_dir': PosixPath('data/lang_bpe_500'), 'decoding_method': 'greedy_search', 'beam_size': 4, 'beam': 20.0, 'ngram_lm_scale': 0.01, 'max_contexts': 8, 'max_states': 64, 'context_size': 2, 'max_sym_per_frame': 1, 'num_paths': 200, 'nbest_scale': 0.5, 'num_encoder_layers': 18, 'dim_feedforward': 1024, 'nhead': 4, 'encoder_dim': 256, 'decoder_dim': 512, 'joiner_dim': 512, 'full_libri': True, 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 600, 'bucketing_sampler': True, 'num_buckets': 30, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'drop_last': True, 'return_cuts': True, 'num_workers': 2, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'input_strategy': 'PrecomputedFeatures', 'res_dir': PosixPath('pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/greedy_search'), 'suffix': 'epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model', 'blank_id': 0, 'unk_id': 2, 'vocab_size': 500}
+2022-07-07 19:14:17,823 INFO [decode.py:663] About to create model
+2022-07-07 19:14:18,002 INFO [decode.py:730] Calculating the averaged model over epoch range from 20 (excluded) to 30
+2022-07-07 19:14:23,248 INFO [decode.py:766] Number of model parameters: 30501996
+2022-07-07 19:14:23,248 INFO [asr_datamodule.py:444] About to get test-clean cuts
+2022-07-07 19:14:23,249 INFO [asr_datamodule.py:451] About to get test-other cuts
+2022-07-07 19:14:24,324 INFO [decode.py:548] batch 0/?, cuts processed until now is 47
+2022-07-07 19:14:36,969 INFO [decode.py:565] The transcripts are stored in pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/greedy_search/recogs-test-clean-greedy_search-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2022-07-07 19:14:37,036 INFO [utils.py:420] [test-clean-greedy_search] %WER 2.87% [1507 / 52576, 154 ins, 140 del, 1213 sub ]
+2022-07-07 19:14:37,202 INFO [decode.py:578] Wrote detailed error stats to pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/greedy_search/errs-test-clean-greedy_search-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2022-07-07 19:14:37,202 INFO [decode.py:595] 
+For test-clean, WER of different settings are:
+greedy_search	2.87	best for test-clean
+
+2022-07-07 19:14:38,050 INFO [decode.py:548] batch 0/?, cuts processed until now is 57
+2022-07-07 19:14:50,096 INFO [decode.py:565] The transcripts are stored in pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/greedy_search/recogs-test-other-greedy_search-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2022-07-07 19:14:50,165 INFO [utils.py:420] [test-other-greedy_search] %WER 6.92% [3621 / 52343, 336 ins, 379 del, 2906 sub ]
+2022-07-07 19:14:50,395 INFO [decode.py:578] Wrote detailed error stats to pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/greedy_search/errs-test-other-greedy_search-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2022-07-07 19:14:50,395 INFO [decode.py:595] 
+For test-other, WER of different settings are:
+greedy_search	6.92	best for test-other
+
+2022-07-07 19:14:50,396 INFO [decode.py:795] Done!

decoding_results/greedy_search/wer-summary-test-clean-greedy_search-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+greedy_search	2.87

decoding_results/greedy_search/wer-summary-test-other-greedy_search-epoch-30-avg-10-context-2-max-sym-per-frame-1-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+greedy_search	6.92

decoding_results/modified_beam_search/log-decode-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model-2022-07-07-18-11-51

@@ -0,0 +1,27 @@
+2022-07-07 18:11:51,588 INFO [decode.py:645] Decoding started
+2022-07-07 18:11:51,588 INFO [decode.py:651] Device: cuda:0
+2022-07-07 18:11:51,590 INFO [decode.py:661] {'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampling_factor': 4, 'model_warm_step': 3000, 'env_info': {'k2-version': '1.15.1', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '350e449fe5d7b4231f77eb0add764782eed9f5d2', 'k2-git-date': 'Thu May 26 22:56:24 2022', 'lhotse-version': '1.4.0.dev+git.21b5b3f.dirty', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'rand_combine', 'icefall-git-sha1': '11f96e2-dirty', 'icefall-git-date': 'Thu Jun 30 12:11:09 2022', 'icefall-path': '/ceph-zw/workspace/experiments/icefall_rand_combine', 'k2-path': '/ceph-zw/workspace/share/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-jb/yaozengwei/workspace/rnnt/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-2-0602201035-5fb6d86964-mclm7', 'IP address': '10.177.74.202'}, 'epoch': 30, 'iter': 0, 'avg': 10, 'use_averaged_model': True, 'exp_dir': PosixPath('pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4'), 'bpe_model': 'data/lang_bpe_500/bpe.model', 'lang_dir': PosixPath('data/lang_bpe_500'), 'decoding_method': 'modified_beam_search', 'beam_size': 4, 'beam': 20.0, 'ngram_lm_scale': 0.01, 'max_contexts': 8, 'max_states': 64, 'context_size': 2, 'max_sym_per_frame': 1, 'num_paths': 200, 'nbest_scale': 0.5, 'num_encoder_layers': 18, 'dim_feedforward': 1024, 'nhead': 4, 'encoder_dim': 256, 'decoder_dim': 512, 'joiner_dim': 512, 'full_libri': True, 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 600, 'bucketing_sampler': True, 'num_buckets': 30, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'drop_last': True, 'return_cuts': True, 'num_workers': 2, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'input_strategy': 'PrecomputedFeatures', 'res_dir': PosixPath('pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/modified_beam_search'), 'suffix': 'epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model', 'blank_id': 0, 'unk_id': 2, 'vocab_size': 500}
+2022-07-07 18:11:51,590 INFO [decode.py:663] About to create model
+2022-07-07 18:11:51,765 INFO [decode.py:730] Calculating the averaged model over epoch range from 20 (excluded) to 30
+2022-07-07 18:12:02,077 INFO [decode.py:766] Number of model parameters: 30501996
+2022-07-07 18:12:02,077 INFO [asr_datamodule.py:444] About to get test-clean cuts
+2022-07-07 18:12:02,079 INFO [asr_datamodule.py:451] About to get test-other cuts
+2022-07-07 18:12:06,055 INFO [decode.py:548] batch 0/?, cuts processed until now is 47
+2022-07-07 18:13:10,238 INFO [decode.py:548] batch 20/?, cuts processed until now is 1656
+2022-07-07 18:13:50,070 INFO [decode.py:565] The transcripts are stored in pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/modified_beam_search/recogs-test-clean-beam_size_4-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model.txt
+2022-07-07 18:13:50,138 INFO [utils.py:420] [test-clean-beam_size_4] %WER 2.83% [1487 / 52576, 157 ins, 123 del, 1207 sub ]
+2022-07-07 18:13:50,303 INFO [decode.py:578] Wrote detailed error stats to pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/modified_beam_search/errs-test-clean-beam_size_4-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model.txt
+2022-07-07 18:13:50,303 INFO [decode.py:595] 
+For test-clean, WER of different settings are:
+beam_size_4	2.83	best for test-clean
+
+2022-07-07 18:13:54,012 INFO [decode.py:548] batch 0/?, cuts processed until now is 57
+2022-07-07 18:14:56,958 INFO [decode.py:548] batch 20/?, cuts processed until now is 1886
+2022-07-07 18:15:35,093 INFO [decode.py:565] The transcripts are stored in pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/modified_beam_search/recogs-test-other-beam_size_4-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model.txt
+2022-07-07 18:15:35,163 INFO [utils.py:420] [test-other-beam_size_4] %WER 6.75% [3533 / 52343, 361 ins, 314 del, 2858 sub ]
+2022-07-07 18:15:35,335 INFO [decode.py:578] Wrote detailed error stats to pruned_transducer_stateless5/exp-960-M-final-05-pure-0333-randomcombine-no-linear-from-layer-4/modified_beam_search/errs-test-other-beam_size_4-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model.txt
+2022-07-07 18:15:35,335 INFO [decode.py:595] 
+For test-other, WER of different settings are:
+beam_size_4	6.75	best for test-other
+
+2022-07-07 18:15:35,336 INFO [decode.py:795] Done!

decoding_results/modified_beam_search/wer-summary-test-clean-beam_size_4-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+beam_size_4	2.83

decoding_results/modified_beam_search/wer-summary-test-other-beam_size_4-epoch-30-avg-10-modified_beam_search-beam-size-4-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+beam_size_4	6.75

exp/decode-M.sh

@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+
+export CUDA_VISIBLE_DEVICES=5
+# greedy_search modified_beam_search fast_beam_search
+for epoch in 30; do
+for avg in 10; do
+ ./pruned_transducer_stateless5/decode.py \
+   --epoch $epoch \
+   --avg $avg \
+   --exp-dir ./pruned_transducer_stateless5/exp-M \
+   --max-duration 600 \
+   --decoding-method greedy_search \
+   --max-sym-per-frame 1 \
+   --num-encoder-layers 18 \
+   --dim-feedforward 1024 \
+   --nhead 4 \
+   --encoder-dim 256 \
+   --decoder-dim 512 \
+   --joiner-dim 512 \
+   --use-averaged-model True
+done
+done

exp/export-M.sh

@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+
+export CUDA_VISIBLE_DEVICES=5
+
+set -ex
+
+epoch=30
+avg=10
+
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp-M \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch $epoch \
+  --avg $avg \
+  --num-encoder-layers 18 \
+  --dim-feedforward 1024 \
+  --nhead 4 \
+  --encoder-dim 256 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --use-averaged-model True
+
+mv -v ./pruned_transducer_stateless5/exp-M/pretrained.pt ./pruned_transducer_stateless5/exp-M/pretrained-epoch-${epoch}-avg-${avg}.pt
+
+

exp/pretrained-epoch-30-avg-10.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b59423de5559aa5f6aff9743d66b3b027d351695d5a96911152e3efd752a45a
+size 122300963

exp/tensorboard/events.out.tfevents.1655275481.de-74279-k2-train-6-0603111709-76b8f59588-flzgf.177157.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74f226d7ce82a6b1677251b6fb28916afd10a00c31b209ac5f8f576f41502e1b
+size 1228527

exp/train-M.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --full-libri 1 \
+  --exp-dir pruned_transducer_stateless5/exp-M \
+  --max-duration 300 \
+  --use-fp16 0 \
+  --num-encoder-layers 18 \
+  --dim-feedforward 1024 \
+  --nhead 4 \
+  --encoder-dim 256 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --master-port 12346

test_wavs/trans.txt

@@ -0,0 +1,3 @@
+1089-134686-0001 AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+1221-135766-0001 GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+1221-135766-0002 YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION