update conformer

conformer/34e9cabc2c29fd0e3a2917ffa525d98b/exp/lm_train_lm_transformer_char_batch_bins2000000/perplexity_test/ppl

+51.1541598159927

conformer/34e9cabc2c29fd0e3a2917ffa525d98b/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/asr_inference_onnx.sh

+python asr_inference_onnx.py \
+    --onnx_exp /home/sunzhq/workspace/yidong-infer/conformer/onnx_models/transformer_lm \
+    --test_sets "test" \
+    --data_dir /data/datasets/0/data_aishell/wav \
+    # --batch_size 4
+    # --tag_name transformer_lm
+
+
+
+
+# # 使用量化模型
+# python asr_inference_onnx.py --onnx_exp exp/conformer_onnx --use_quantized --batch_size 4
+
+# # 使用GPU加速
+# python asr_inference_onnx.py --onnx_exp exp/conformer_onnx --device gpu --inference_nj 8
+
+# - --onnx_exp : ONNX实验目录（必需）
+# - --data_dir : 数据目录（默认：./data）
+# - --test_sets : 测试集名称（默认：test）
+# - --batch_size : 批量大小（默认：1）
+# - --device : 推理设备（cpu/gpu，默认：cpu）
+# - --inference_nj : 并行任务数（默认：4）
+# - --use_quantized : 使用量化模型
\ No newline at end of file

conformer/convert_onnx_batch_size.sh

+# 批量转换所有模型
+python convert_onnx_batch_size.py \
+  --input /home/sunzhq/workspace/yidong-infer/conformer/onnx_models/transformer_lm/full \
+  --output /home/sunzhq/workspace/yidong-infer/conformer/onnx_models_batch24 \
+  --batch_size 24 \
+  --batch_mode
\ No newline at end of file

conformer/data.py

+import librosa
+import os
+
+sr=16000
+audio_dir = "/data/datasets/1/data_aishell/wav/test"
+dir_list = os.listdir(audio_dir)
+tmp = []
+# print(dir_list)
+for index in dir_list:
+    audio_paths = os.listdir(os.path.join(audio_dir,index))
+    for audio_path in audio_paths:
+        y, sr = librosa.load(os.path.join(audio_dir,index,audio_path), sr=sr)
+        if len(y)/sr == 14.6999375:
+            print(os.path.join(audio_dir,index,audio_path))
+        
+        tmp.append(len(y)/sr)
+# print(sorted(tmp))    
+    # print(audio_paths)
+
+
+
+# y, sr = librosa.load(audio_path, sr=16000)
+# print(f"音频总长: {len(y)/sr:.2f}秒 ({len(y)}采样点)")
\ No newline at end of file