from:ESPnet2 — ESPnet 202301 documentation
from :Change the configuration for training — ESPnet 202301 documentation
训练完之后微调的命令:
./run.sh --stage 11 --ngpu 1 --asr_args "--max_epoch 205 --optim_conf lr=0.1 --resume true" --asr_exp exp/asr_train_asr_transformer_raw_en_bpe30_sp/ --pretrained_model exp/asr_train_asr_transformer_raw_en_bpe30_sp/200epoch.pth
评估已训练model:
./run.sh --skip_data_prep true --skip_train true --asr_exp exp/asr_train_asr_transformer_raw_en_bpe30_sp --lm_exp exp/lm_train_lm_en_bpe30
使用预训练模型进行无训练评估:
./run.sh --download_model kamo-naoyuki/mini_an4_asr_train_raw_bpe_valid.acc.best --skip_train true --skip_data_prep true
预训练模型全名查找表:
espnet_model_zoo/table.csv at master · espnet/espnet_model_zoo · GitHub
GitHub - espnet/espnet_model_zoo: ESPnet Model Zoo