diff --git a/official/audio/Tacotron2/README.md b/official/audio/Tacotron2/README.md index 53283ba9b019691a9454d9b9a7c8e51337f86611..fbf0ac81c7b5db5600dc4fe008f51e25842fc6bd 100644 --- a/official/audio/Tacotron2/README.md +++ b/official/audio/Tacotron2/README.md @@ -76,8 +76,8 @@ After installing MindSpore via the official website, you can start training and # example: bash run_standalone_train.sh /path/ljspeech.hdf5 0 # run distributed training - bash run_distributed_train.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] - # example: bash run_distributed_train.sh /path/ljspeech.h5 ../hccl_8p_01234567_127.0.0.1.json 8 0 + bash run_distribute_train_multi_core.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] + # example: bash run_distribute_train_multi_core.sh /path/ljspeech.h5 ../hccl_8p_01234567_127.0.0.1.json 8 0 # run evaluation bash run_eval.sh [OUTPUT_PATH] [MODEL_CKPT] [DEVICE_ID] text is set in config.py( can modify text of ljspeech_config.yaml) @@ -246,7 +246,7 @@ Parameters for both training and evaluation can be set in [DATASET]_config.yaml ```bash cd scripts - bash run_distributed_train.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] + bash run_distribute_train_multi_core.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] ``` Note: `DATASET_PATH` is the directory contains hdf5 file. diff --git a/official/audio/Tacotron2/README_CN.md b/official/audio/Tacotron2/README_CN.md index e104f6e4872f367bd2ab838ea118ee80fc74952c..d2a0e5d209eb7220c38459b7c2cf45380315fb09 100644 --- a/official/audio/Tacotron2/README_CN.md +++ b/official/audio/Tacotron2/README_CN.md @@ -76,8 +76,8 @@ Tacotron2实质上是一种包含编码器和解码器的序列到序列模型 # 示例:bash run_standalone_train.sh /path/ljspeech.hdf5 0 # 运行分布式训练 - bash run_distributed_train.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] - # 示例:bash run_distributed_train.sh /path/ljspeech.h5 ../hccl_8p_01234567_127.0.0.1.json 8 0 + bash run_distribute_train_multi_core.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] + # 示例:bash run_distribute_train_multi_core.sh /path/ljspeech.h5 ../hccl_8p_01234567_127.0.0.1.json 8 0 # 运行评估 bash run_eval.sh [OUTPUT_PATH] [MODEL_CKPT] [DEVICE_ID] text is set in config.py( can modify text of ljspeech_config.yaml) @@ -246,7 +246,7 @@ tacotron2/ ```bash cd scripts - bash run_distributed_train.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] + bash run_distribute_train_multi_core.sh [DATASET_PATH] [RANK_TABLE_PATH] [DATANAME] [RANK_SIZE] [DEVICE_BEGIN] ``` 注:`DATASET_PATH`是包含HDF5文件的目录。 diff --git a/official/audio/Tacotron2/src/tacotron2.py b/official/audio/Tacotron2/src/tacotron2.py index e685cf0f8d0ca516cc18a92674488e9a12b6ca44..a5b573f31a1615e71fb4faa2fd502f8cdfa4ef5e 100644 --- a/official/audio/Tacotron2/src/tacotron2.py +++ b/official/audio/Tacotron2/src/tacotron2.py @@ -1168,7 +1168,7 @@ class TrainStepWrap(nn.Cell): overflow = ops.logical_not(amp.all_finite(grads)) if self.reducer_flag: - overflow = self.allreduce(overflow.to(mstype.float32)) >= self.base + overflow = self.all_reduce(overflow.to(mstype.float32)) >= self.base overflow = self.loss_scaling_manager(self.loss_scale, overflow)