EasyNLP

EasyNLP copied to clipboard

https://zhuanlan.zhihu.com/c_1500207477181960192

chywang

minghui

add UPT code to EasyNLPexamples/upt

chywang

ERROR:torch.distributed.elastic.multiprocessing.api:failed

1

.

MagdaEhab

Predictor not support multi-gpu

1

The predictor seems only support single gpu inference: https://github.com/alibaba/EasyNLP/blob/cda9b42358156afed3a0e27173d5e48ef32f0686/easynlp/core/predictor.py#L130 but the trainer supports multi-gpu training: https://github.com/alibaba/EasyNLP/blob/01fef50c3cda57971de2764b9095a152b1300a9e/easynlp/core/trainer.py#L96 Is there any plan for supporting multi-gpu inference?

yslin1995

_common_io.UserException: table/table_buffer.cpp(93): UserException: Read table time out!

2

参考脚本quick_start_user_defined/run_user_defined_pai.sh 输入是odps表，输出模型到oss 如果在main.py 中将 evaluator = None替换为 evaluator = get_application_evaluator(app_name=args.app_name,valid_dataset=valid_dataset,user_defined_parameters=user_defined_parameters,eval_batch_size=args.micro_batch_size) 时会出现_common_io.UserException: table/table_buffer.cpp(93): UserException: Read table time out!错误，反复运行多次，大概都在正常运行20分钟左右会出现。 如果设置evaluator=None则正常训练。

shumuha

minghui

args.tables

1

改成args.inputs

minghui

代码位置`easynlp.appzoo.sequence_labeling.predictor.py` `SequenceLabelingPredictor.preprocess` ```python text_a = record[self.first_sequence] if self.tokenized else " ".join(record[self.first_sequence]) # record[self.first_sequence] = '中 国 最 棒' # text_a = '中 国 最 棒' (多加了一个空格) ``` 除非指明要预测的序列不包含空格间隔，不然会多一个空格。

LucasGY

FreePromptEditing project and code is missing

1

hangzeli