macbert训练后如何转化为onnx模型呢

[lazywangyuan]

if ckpt_path and os.path.exists(ckpt_path):
    model.load_state_dict(torch.load(ckpt_path)['state_dict'])
    # 先保存原始transformer bert model
    tokenizer.save_pretrained(cfg.OUTPUT_DIR)
    bert = BertForMaskedLM.from_pretrained(cfg.MODEL.BERT_CKPT)
    bert.save_pretrained(cfg.OUTPUT_DIR)
    state_dict = torch.load(ckpt_path)['state_dict']
    new_state_dict = OrderedDict()
    if cfg.MODEL.NAME in ['macbert4csc']:
        for k, v in state_dict.items():
            if k.startswith('bert.'):
                new_state_dict[k[5:]] = v
    else:
        new_state_dict = state_dict
    # 再保存finetune训练后的模型文件，替换原始的pytorch_model.bin
    torch.save(new_state_dict, os.path.join(cfg.OUTPUT_DIR, 'pytorch_model.bin'))

[shibing624]

还没做，稍后我补充下。

[Yazooliu]

还没做，稍后我补充下。

Could I update code and PR to support this exchange model to onnx format ? thanks.

BestRegards Yazhou

[shibing624]

可以

[Yazooliu]

可以

Thanks. I will update it as soon as possible

[ChenchenJT]

你好，想问一下这个issue实现了吗，在项目没找到相应的代码或者PR

[shibing624]

后续我不添加该功能，官方教程就可以转换为onnx。