conformer add post and ana

conformer/torch-infer/exp/lm_train_lm_transformer_char_batch_bins2000000/perplexity_test/ppl

+51.1541598159927

conformer/torch-infer/1.py → conformer/torch-infer/infer-torch.py

@@ -158,8 +158,14 @@ if __name__ == '__main__':
 
             # b. Forward Encoder
             # enc: [N, T, C]
+            
+            feats, feats_lengths = speech2text.asr_model.pre_data(**batch)
+            feats_lengths_1 = torch.ceil(feats_lengths.float() / 4).long()
+            print("feats_lengths_1:",feats_lengths_1)
+            # print("feats_lengths:",feats_lengths)
             ll_time = time.time()
-            encoder_out, encoder_out_lens = speech2text.asr_model.encode(**batch)
+            encoder_out, encoder_out_lens = speech2text.asr_model.encode(feats, feats_lengths)
+            print("encoder_out_lens:",encoder_out_lens)
             # ctc_log_probs: [N, T, C]
             ctc_log_probs = torch.nn.functional.log_softmax(
                 speech2text.asr_model.ctc.ctc_lo(encoder_out), dim=2

conformer/torch-infer/infer.sh

+#!/usr/bin/bash
+# if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+# asr_train_config="/home/sunzhq/workspace/yidong-infer/conformer/34e9cabc2c29fd0e3a2917ffa525d98b/exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/config.yaml"
+# asr_model_file="/home/sunzhq/workspace/yidong-infer/conformer/34e9cabc2c29fd0e3a2917ffa525d98b/exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/valid.acc.ave_10best.pth"
+# lm_train_config=/home/sunzhq/workspace/yidong-infer/conformer/34e9cabc2c29fd0e3a2917ffa525d98b/exp/lm_train_lm_transformer_char_batch_bins2000000/config.yaml
+# lm_path=/home/sunzhq/workspace/yidong-infer/conformer/34e9cabc2c29fd0e3a2917ffa525d98b/exp/lm_train_lm_transformer_char_batch_bins2000000/valid.loss.ave_10best.pth
+# manifest="/home/sunzhq/workspace/yidong-infer/conformer/torch-infer/test"
+
+asr_train_config="/home/sunzhq/workspace/yidong-infer/conformer/torch-infer/exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/config.yaml"
+asr_model_file="/home/sunzhq/workspace/yidong-infer/conformer/torch-infer/exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/valid.acc.ave_10best.pth"
+lm_train_config=/home/sunzhq/workspace/yidong-infer/conformer/torch-infer/exp/lm_train_lm_transformer_char_batch_bins2000000/config.yaml
+lm_path=/home/sunzhq/workspace/yidong-infer/conformer/torch-infer/exp/lm_train_lm_transformer_char_batch_bins2000000/valid.loss.ave_10best.pth
+manifest="/home/sunzhq/workspace/yidong-infer/conformer/torch-infer/test"
+
+
+mkdir -p logs
+# mode='attention_rescoring'
+mode='lm_rescoring'
+# num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+export HIP_VISIBLE_DEVICES=0
+nohup numactl -N 0 -m 0 python3 infer.py \
+        --config $asr_train_config \
+        --model_path $asr_model_file \
+        --lm_config $lm_train_config \
+        --lm_path $lm_path \
+        --gpu 0 \
+        --wav_scp $manifest/wav.scp --text $manifest/text \
+        --result_file ./logs/predictions_${mode}_$gpu_id.txt \
+        --log_file ./logs/log_${mode}_$gpu_id.txt \
+        --batch_size 24 --beam_size 10 \
+        --mode $mode 2>&1 | tee result_0.log &
+
+
+export HIP_VISIBLE_DEVICES=1
+nohup numactl -N 1 -m 1 python3 infer.py \
+        --config $asr_train_config \
+        --model_path $asr_model_file \
+        --lm_config $lm_train_config \
+        --lm_path $lm_path \
+        --gpu 0 \
+        --wav_scp $manifest/wav.scp --text $manifest/text \
+        --result_file ./logs/predictions_${mode}_$gpu_id.txt \
+        --log_file ./logs/log_${mode}_$gpu_id.txt \
+        --batch_size 24 --beam_size 10 \
+        --mode $mode 2>&1 | tee result_1.log &
+
+export HIP_VISIBLE_DEVICES=2
+nohup numactl -N 2 -m 2 python3 infer.py \
+        --config $asr_train_config \
+        --model_path $asr_model_file \
+        --lm_config $lm_train_config \
+        --lm_path $lm_path \
+        --gpu 0 \
+        --wav_scp $manifest/wav.scp --text $manifest/text \
+        --result_file ./logs/predictions_${mode}_$gpu_id.txt \
+        --log_file ./logs/log_${mode}_$gpu_id.txt \
+        --batch_size 24 --beam_size 10 \
+        --mode $mode 2>&1 | tee result_2.log &
+
+export HIP_VISIBLE_DEVICES=3
+nohup numactl -N 3 -m 3 python3 infer.py \
+        --config $asr_train_config \
+        --model_path $asr_model_file \
+        --lm_config $lm_train_config \
+        --lm_path $lm_path \
+        --gpu 0 \
+        --wav_scp $manifest/wav.scp --text $manifest/text \
+        --result_file ./logs/predictions_${mode}_$gpu_id.txt \
+        --log_file ./logs/log_${mode}_$gpu_id.txt \
+        --batch_size 24 --beam_size 10 \
+        --mode $mode 2>&1 | tee result_3.log &
+

conformer/torch-infer/meta.yaml

+espnet: 0.9.0
+files:
+  asr_model_file: exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/valid.acc.ave_10best.pth
+  lm_file: exp/lm_train_lm_transformer_char_batch_bins2000000/valid.loss.ave_10best.pth
+python: "3.7.3 (default, Mar 27 2019, 22:11:17) \n[GCC 7.3.0]"
+timestamp: 1603088092.704853
+torch: 1.6.0
+yaml_files:
+  asr_train_config: exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/config.yaml
+  lm_train_config: exp/lm_train_lm_transformer_char_batch_bins2000000/config.yaml

conformer/torch-infer/post.sh

+python3 conformer-compute-wer.py ./logs/ref.trn ./logs/hyp.trn
\ No newline at end of file

meta.yaml

+espnet: 0.9.0
+files:
+  asr_model_file: exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/valid.acc.ave_10best.pth
+  lm_file: exp/lm_train_lm_transformer_char_batch_bins2000000/valid.loss.ave_10best.pth
+python: "3.7.3 (default, Mar 27 2019, 22:11:17) \n[GCC 7.3.0]"
+timestamp: 1603088092.704853
+torch: 1.6.0
+yaml_files:
+  asr_train_config: exp/asr_train_asr_conformer3_raw_char_batch_bins4000000_accum_grad4_sp/config.yaml
+  lm_train_config: exp/lm_train_lm_transformer_char_batch_bins2000000/config.yaml