upload files

decoding_result/fast_beam_search/log-decode-epoch-50-avg-23-beam-20.0-max-contexts-8-max-states-64-use-averaged-model-2023-10-08-19-03-08

@@ -0,0 +1,31 @@
+2023-10-08 19:03:08,979 INFO [decode.py:670] Decoding started
+2023-10-08 19:03:08,979 INFO [decode.py:676] Device: cuda:0
+2023-10-08 19:03:09,013 INFO [decode.py:686] {'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, 'warm_step': 2000, 'env_info': {'k2-version': '1.24.1', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'c51a0b9684442a88ee37f3ce0af686a04b66855b', 'k2-git-date': 'Mon May 1 21:38:03 2023', 'lhotse-version': '1.12.0.dev+git.891bad1.clean', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'new-zipformer-onnx-decode', 'icefall-git-sha1': 'a53d7102-dirty', 'icefall-git-date': 'Fri Jun 30 11:38:26 2023', 'icefall-path': '/ceph-zw/workspace/zipformer/icefall_zipformer', 'k2-path': '/ceph-zw/workspace/k2/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-zw/workspace/share/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-9-0208143539-7dbf569d4f-r7nrb', 'IP address': '10.177.13.150'}, 'epoch': 50, 'iter': 0, 'avg': 23, 'use_averaged_model': True, 'exp_dir': PosixPath('zipformer/exp-small'), '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': '2,2,2,2,2,2', 'downsampling_factor': '1,2,4,8,4,2', 'feedforward_dim': '512,768,768,768,768,768', 'num_heads': '4,4,4,8,4,4', 'encoder_dim': '192,256,256,256,256,256', 'query_head_dim': '32', 'value_head_dim': '12', 'pos_head_dim': '4', 'pos_dim': 48, 'encoder_unmasked_dim': '192,192,192,192,192,192', 'cnn_module_kernel': '31,31,15,15,15,31', 'decoder_dim': 512, 'joiner_dim': 512, 'causal': False, 'chunk_size': '16,32,64,-1', 'left_context_frames': '64,128,256,-1', 'use_transducer': True, 'use_ctc': False, 'full_libri': True, 'mini_libri': False, '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('zipformer/exp-small/fast_beam_search'), 'suffix': 'epoch-50-avg-23-beam-20.0-max-contexts-8-max-states-64-use-averaged-model', 'blank_id': 0, 'unk_id': 2, 'vocab_size': 500}
+2023-10-08 19:03:09,014 INFO [decode.py:688] About to create model
+2023-10-08 19:03:09,563 INFO [decode.py:755] Calculating the averaged model over epoch range from 27 (excluded) to 50
+2023-10-08 19:03:18,234 INFO [decode.py:791] Number of model parameters: 23285615
+2023-10-08 19:03:18,234 INFO [asr_datamodule.py:465] About to get test-clean cuts
+2023-10-08 19:03:18,240 INFO [asr_datamodule.py:472] About to get test-other cuts
+2023-10-08 19:03:21,485 INFO [decode.py:562] batch 0/?, cuts processed until now is 43
+2023-10-08 19:03:49,789 INFO [zipformer.py:1728] name=None, attn_weights_entropy = tensor([2.0788, 2.8359, 2.6445, 2.9472, 1.7865, 2.8430, 3.0297, 2.9063],
+       device='cuda:0')
+2023-10-08 19:03:59,044 INFO [decode.py:562] batch 20/?, cuts processed until now is 1430
+2023-10-08 19:04:39,332 INFO [decode.py:562] batch 40/?, cuts processed until now is 2561
+2023-10-08 19:04:40,532 INFO [decode.py:580] The transcripts are stored in zipformer/exp-small/fast_beam_search/recogs-test-clean-beam_20.0_max_contexts_8_max_states_64-epoch-50-avg-23-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2023-10-08 19:04:40,648 INFO [utils.py:562] [test-clean-beam_20.0_max_contexts_8_max_states_64] %WER 2.43% [1280 / 52576, 136 ins, 95 del, 1049 sub ]
+2023-10-08 19:04:40,950 INFO [decode.py:593] Wrote detailed error stats to zipformer/exp-small/fast_beam_search/errs-test-clean-beam_20.0_max_contexts_8_max_states_64-epoch-50-avg-23-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2023-10-08 19:04:40,954 INFO [decode.py:610] 
+For test-clean, WER of different settings are:
+beam_20.0_max_contexts_8_max_states_64	2.43	best for test-clean
+
+2023-10-08 19:04:43,548 INFO [decode.py:562] batch 0/?, cuts processed until now is 52
+2023-10-08 19:05:19,615 INFO [decode.py:562] batch 20/?, cuts processed until now is 1647
+2023-10-08 19:05:58,991 INFO [decode.py:562] batch 40/?, cuts processed until now is 2870
+2023-10-08 19:06:00,126 INFO [decode.py:580] The transcripts are stored in zipformer/exp-small/fast_beam_search/recogs-test-other-beam_20.0_max_contexts_8_max_states_64-epoch-50-avg-23-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2023-10-08 19:06:00,247 INFO [utils.py:562] [test-other-beam_20.0_max_contexts_8_max_states_64] %WER 5.78% [3028 / 52343, 312 ins, 253 del, 2463 sub ]
+2023-10-08 19:06:00,560 INFO [decode.py:593] Wrote detailed error stats to zipformer/exp-small/fast_beam_search/errs-test-other-beam_20.0_max_contexts_8_max_states_64-epoch-50-avg-23-beam-20.0-max-contexts-8-max-states-64-use-averaged-model.txt
+2023-10-08 19:06:00,563 INFO [decode.py:610] 
+For test-other, WER of different settings are:
+beam_20.0_max_contexts_8_max_states_64	5.78	best for test-other
+
+2023-10-08 19:06:00,564 INFO [decode.py:822] Done!

decoding_result/fast_beam_search/wer-summary-test-clean-beam_20.0_max_contexts_8_max_states_64-epoch-50-avg-23-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	2.43

decoding_result/fast_beam_search/wer-summary-test-other-beam_20.0_max_contexts_8_max_states_64-epoch-50-avg-23-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	5.78

decoding_result/greedy_search/log-decode-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model-2023-10-08-18-55-30

@@ -0,0 +1,25 @@
+2023-10-08 18:55:30,506 INFO [decode.py:670] Decoding started
+2023-10-08 18:55:30,507 INFO [decode.py:676] Device: cuda:0
+2023-10-08 18:55:30,512 INFO [decode.py:686] {'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, 'warm_step': 2000, 'env_info': {'k2-version': '1.24.1', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'c51a0b9684442a88ee37f3ce0af686a04b66855b', 'k2-git-date': 'Mon May 1 21:38:03 2023', 'lhotse-version': '1.12.0.dev+git.891bad1.clean', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'new-zipformer-onnx-decode', 'icefall-git-sha1': 'a53d7102-clean', 'icefall-git-date': 'Fri Jun 30 11:38:26 2023', 'icefall-path': '/ceph-zw/workspace/zipformer/icefall_zipformer', 'k2-path': '/ceph-zw/workspace/k2/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-zw/workspace/share/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-9-0208143539-7dbf569d4f-r7nrb', 'IP address': '10.177.13.150'}, 'epoch': 50, 'iter': 0, 'avg': 23, 'use_averaged_model': True, 'exp_dir': PosixPath('zipformer/exp-small'), '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': '2,2,2,2,2,2', 'downsampling_factor': '1,2,4,8,4,2', 'feedforward_dim': '512,768,768,768,768,768', 'num_heads': '4,4,4,8,4,4', 'encoder_dim': '192,256,256,256,256,256', 'query_head_dim': '32', 'value_head_dim': '12', 'pos_head_dim': '4', 'pos_dim': 48, 'encoder_unmasked_dim': '192,192,192,192,192,192', 'cnn_module_kernel': '31,31,15,15,15,31', 'decoder_dim': 512, 'joiner_dim': 512, 'causal': False, 'chunk_size': '16,32,64,-1', 'left_context_frames': '64,128,256,-1', 'use_transducer': True, 'use_ctc': False, 'full_libri': True, 'mini_libri': False, '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('zipformer/exp-small/greedy_search'), 'suffix': 'epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model', 'blank_id': 0, 'unk_id': 2, 'vocab_size': 500}
+2023-10-08 18:55:30,512 INFO [decode.py:688] About to create model
+2023-10-08 18:55:31,099 INFO [decode.py:755] Calculating the averaged model over epoch range from 27 (excluded) to 50
+2023-10-08 18:55:43,284 INFO [decode.py:791] Number of model parameters: 23285615
+2023-10-08 18:55:43,284 INFO [asr_datamodule.py:465] About to get test-clean cuts
+2023-10-08 18:55:43,291 INFO [asr_datamodule.py:472] About to get test-other cuts
+2023-10-08 18:55:45,711 INFO [decode.py:562] batch 0/?, cuts processed until now is 43
+2023-10-08 18:56:26,034 INFO [decode.py:580] The transcripts are stored in zipformer/exp-small/greedy_search/recogs-test-clean-greedy_search-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2023-10-08 18:56:26,156 INFO [utils.py:562] [test-clean-greedy_search] %WER 2.46% [1293 / 52576, 135 ins, 107 del, 1051 sub ]
+2023-10-08 18:56:26,463 INFO [decode.py:593] Wrote detailed error stats to zipformer/exp-small/greedy_search/errs-test-clean-greedy_search-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2023-10-08 18:56:26,467 INFO [decode.py:610] 
+For test-clean, WER of different settings are:
+greedy_search	2.46	best for test-clean
+
+2023-10-08 18:56:28,204 INFO [decode.py:562] batch 0/?, cuts processed until now is 52
+2023-10-08 18:57:08,107 INFO [decode.py:580] The transcripts are stored in zipformer/exp-small/greedy_search/recogs-test-other-greedy_search-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2023-10-08 18:57:08,236 INFO [utils.py:562] [test-other-greedy_search] %WER 5.86% [3066 / 52343, 302 ins, 259 del, 2505 sub ]
+2023-10-08 18:57:08,556 INFO [decode.py:593] Wrote detailed error stats to zipformer/exp-small/greedy_search/errs-test-other-greedy_search-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model.txt
+2023-10-08 18:57:08,559 INFO [decode.py:610] 
+For test-other, WER of different settings are:
+greedy_search	5.86	best for test-other
+
+2023-10-08 18:57:08,559 INFO [decode.py:822] Done!

decoding_result/greedy_search/wer-summary-test-clean-greedy_search-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+greedy_search	2.46

decoding_result/greedy_search/wer-summary-test-other-greedy_search-epoch-50-avg-23-context-2-max-sym-per-frame-1-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+greedy_search	5.86

decoding_result/modified_beam_search/log-decode-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model-2023-09-11-14-16-47

@@ -0,0 +1,31 @@
+2023-09-11 14:16:47,669 INFO [decode.py:670] Decoding started
+2023-09-11 14:16:47,669 INFO [decode.py:676] Device: cuda:0
+2023-09-11 14:16:47,671 INFO [decode.py:686] {'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, 'warm_step': 2000, 'env_info': {'k2-version': '1.24.1', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'c51a0b9684442a88ee37f3ce0af686a04b66855b', 'k2-git-date': 'Mon May 1 21:38:03 2023', 'lhotse-version': '1.12.0.dev+git.891bad1.clean', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'new-zipformer-onnx-decode', 'icefall-git-sha1': 'a53d7102-clean', 'icefall-git-date': 'Fri Jun 30 11:38:26 2023', 'icefall-path': '/ceph-zw/workspace/zipformer/icefall_zipformer', 'k2-path': '/ceph-zw/workspace/k2/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-zw/workspace/share/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-7-1218101249-5d97868c7c-98fdf', 'IP address': '10.177.22.20'}, 'epoch': 50, 'iter': 0, 'avg': 23, 'use_averaged_model': True, 'exp_dir': PosixPath('zipformer/exp-small'), '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': '2,2,2,2,2,2', 'downsampling_factor': '1,2,4,8,4,2', 'feedforward_dim': '512,768,768,768,768,768', 'num_heads': '4,4,4,8,4,4', 'encoder_dim': '192,256,256,256,256,256', 'query_head_dim': '32', 'value_head_dim': '12', 'pos_head_dim': '4', 'pos_dim': 48, 'encoder_unmasked_dim': '192,192,192,192,192,192', 'cnn_module_kernel': '31,31,15,15,15,31', 'decoder_dim': 512, 'joiner_dim': 512, 'causal': False, 'chunk_size': '16,32,64,-1', 'left_context_frames': '64,128,256,-1', 'use_transducer': True, 'use_ctc': False, 'full_libri': True, 'mini_libri': False, '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('zipformer/exp-small/modified_beam_search'), 'suffix': 'epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model', 'blank_id': 0, 'unk_id': 2, 'vocab_size': 500}
+2023-09-11 14:16:47,672 INFO [decode.py:688] About to create model
+2023-09-11 14:16:48,034 INFO [decode.py:755] Calculating the averaged model over epoch range from 27 (excluded) to 50
+2023-09-11 14:16:55,204 INFO [decode.py:791] Number of model parameters: 23285615
+2023-09-11 14:16:55,204 INFO [asr_datamodule.py:465] About to get test-clean cuts
+2023-09-11 14:16:55,207 INFO [asr_datamodule.py:472] About to get test-other cuts
+2023-09-11 14:17:00,230 INFO [decode.py:562] batch 0/?, cuts processed until now is 43
+2023-09-11 14:18:22,246 INFO [decode.py:562] batch 20/?, cuts processed until now is 1430
+2023-09-11 14:19:13,020 INFO [zipformer.py:1728] name=None, attn_weights_entropy = tensor([5.0861, 4.5347, 4.5748, 4.8806], device='cuda:0')
+2023-09-11 14:19:29,593 INFO [decode.py:562] batch 40/?, cuts processed until now is 2574
+2023-09-11 14:19:32,196 INFO [decode.py:580] The transcripts are stored in zipformer/exp-small/modified_beam_search/recogs-test-clean-beam_size_4-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model.txt
+2023-09-11 14:19:32,269 INFO [utils.py:562] [test-clean-beam_size_4] %WER 2.42% [1273 / 52576, 139 ins, 95 del, 1039 sub ]
+2023-09-11 14:19:32,442 INFO [decode.py:593] Wrote detailed error stats to zipformer/exp-small/modified_beam_search/errs-test-clean-beam_size_4-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model.txt
+2023-09-11 14:19:32,443 INFO [decode.py:610] 
+For test-clean, WER of different settings are:
+beam_size_4	2.42	best for test-clean
+
+2023-09-11 14:19:37,300 INFO [decode.py:562] batch 0/?, cuts processed until now is 52
+2023-09-11 14:20:56,041 INFO [decode.py:562] batch 20/?, cuts processed until now is 1646
+2023-09-11 14:21:39,690 INFO [zipformer.py:1728] name=None, attn_weights_entropy = tensor([6.0102, 5.4262, 5.4479, 5.3197], device='cuda:0')
+2023-09-11 14:22:01,544 INFO [decode.py:562] batch 40/?, cuts processed until now is 2870
+2023-09-11 14:22:04,095 INFO [decode.py:580] The transcripts are stored in zipformer/exp-small/modified_beam_search/recogs-test-other-beam_size_4-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model.txt
+2023-09-11 14:22:04,168 INFO [utils.py:562] [test-other-beam_size_4] %WER 5.73% [3001 / 52343, 308 ins, 244 del, 2449 sub ]
+2023-09-11 14:22:04,414 INFO [decode.py:593] Wrote detailed error stats to zipformer/exp-small/modified_beam_search/errs-test-other-beam_size_4-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model.txt
+2023-09-11 14:22:04,415 INFO [decode.py:610] 
+For test-other, WER of different settings are:
+beam_size_4	5.73	best for test-other
+
+2023-09-11 14:22:04,415 INFO [decode.py:822] Done!

decoding_result/modified_beam_search/wer-summary-test-clean-beam_size_4-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+beam_size_4	2.42

decoding_result/modified_beam_search/wer-summary-test-other-beam_size_4-epoch-50-avg-23-modified_beam_search-beam-size-4-use-averaged-model.txt

@@ -0,0 +1,2 @@
+settings	WER
+beam_size_4	5.73

exp/decode.sh

@@ -1,8 +1,8 @@
 export CUDA_VISIBLE_DEVICES="0"
 for m in greedy_search modified_beam_search fast_beam_search; do
   ./zipformer/decode.py \
-    --epoch 40 \
-    --avg 13 \
+    --epoch 50 \
+    --avg 23 \
     --exp-dir zipformer/exp-small \
     --max-duration 600 \
     --causal 0 \

exp/epoch-50.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99af26d0a84261c9081172d45c0b312eae30c2f7db216756b00c59b473988707
+size 373308527

exp/export.sh

@@ -5,8 +5,8 @@
   --feedforward-dim 512,768,768,768,768,768 \
   --encoder-dim 192,256,256,256,256,256 \
   --encoder-unmasked-dim 192,192,192,192,192,192 \
-  --epoch 40 \
-  --avg 13 \
+  --epoch 50 \
+  --avg 23 \
   --jit 1
 
 

exp/jit_script.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7aa3bc63719d5d23a94a166d072eea86fc84921171c4dbdf8a63a8cb4020a858
-size 95780838
+oid sha256:3f325d3ebac1b42613fc64200cee3bd169a62d0a41c2d2583b53510de8b8fed7
+size 95783654

exp/pretrained.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cef695fa62fa0e4e2d01fef5fcacb66ddb89f590fbf8a7381d428155de070df7
+oid sha256:2c0338660c6a31e53224646149f560ab14439bbc4f73e2eef59055a1c4466ebe
 size 93451027

exp/train.sh

@@ -1,7 +1,7 @@
 export CUDA_VISIBLE_DEVICES="0,1"
 ./zipformer/train.py \
   --world-size 2 \
-  --num-epochs 40 \
+  --num-epochs 50 \
   --start-epoch 1 \
   --use-fp16 1 \
   --exp-dir zipformer/exp-small \