描述bug

我正在尝试使用crf-tagger进行命名实体识别(NER)任务。我使用bert-tokenizer对句子进行了预分词。句子和标签都是基于word-piece的，当发送到crf-tagger时。此外，我还从master安装了allennlp。
问题：训练过程中内存消耗不断增加。

重现问题

我的训练配置如下：
{ "dataset_reader": { "type": "wordpiece_bert_crf_reader", "tokenizers":{ "bert_tokenizer":{ "type": "whitespace" } }, "token_indexers": { "transformer": { "type": "pretrained_transformer_mismatched", "model_name": "/home/mhxia/whou/workspace/bert-base-uncased/", "max_length":510 } } }, "train_data_path": "/home/mhxia/whou/workspace/code_repo/space12-ner/data/processed_data/ensemble_data/ner_bert_tokenized/train_1.jsonl", "validation_data_path": "/home/mhxia/whou/workspace/code_repo/space12-ner/data/processed_data/ensemble_data/ner_bert_tokenized/dev_1.jsonl", "model": { "type": "crf_tagger", "include_start_end_transitions": false, "text_field_embedder": { "token_embedders": { "transformer": { "type": "pretrained_transformer_mismatched", "model_name": "/home/mhxia/whou/workspace/bert-base-uncased/", "max_length":510 } } }, "encoder": { "type": "pass_through", "input_dim": 768 } }, "data_loader": {"batch_size": 2}, "trainer": { "optimizer": "adam", "num_epochs": 5, "cuda_device":0 } }

预期行为

希望你能修复这个bug。

系统信息(请填写以下信息):

• 操作系统：Linux
• Python版本：3.6.9
• AllenNLP版本：从master安装的
• PyTorch版本：1.4.0