KeyError: 'rnn.1.embedding.weight'

[Amore-HDU]

First, I train a model use very large datasets. I want use this model to fine tuning my new datasets with the same keys.But when code load pre model, it occurs loading pretrained model from ./GenModel/netCRNN_447_10.pth Traceback (most recent call last): File "crnn_main.py", line 110, in if len(model_dict[weig1]) == len(pre_trainmodel[weig1]) and len(model_dict[bias1]) == len(pre_trainmodel[bias1]): KeyError: 'rnn.1.embedding.weight' @meijieru

[Amore-HDU]

@meijieru can you help me to solve this problem?Thank you.

[meijieru]

Occur when run demo?

[Amore-HDU]

Occur when run train @meijieru

[VenkteshV]

use .get on dictionary instead of indexing it will solve the error

[VenkteshV]

change this line text = [ self.dict[char.lower() if self._ignore_case else char] for char in text ]

in utils.py to

text = [ self.dict.get(char.lower() if self._ignore_case else char) for char in text ]

[zhengmzong]

@meijieru @Amore-HDU the error is caused by inconsistent name between defined model and saved mode.(*.pth) Name of params in saved mode is added prefix 'module.'. I have saved this problom below.

1. first, define a function:

def load_pretrained_model(model, pretrained_param): model_param = model.state_dict() assert(len(model_param) == len(pretrained_param)) for i in model_param.keys(): model_param[i] = pretrained_param['module.'+i] model.load_state_dict(model_param) 2. crnn_main.py change the line: crnn.load_state_dict(torch.load(opt.crnn)) to: load_pretrained_model(crnn, torch.load(opt.crnn))

[DLUTfangping]

I do not solve the problem using the method ,could you help me . @zhengmzong

[DLUTfangping]

Did you solve the problem? @Amore-HDU

[bianlongpeng]

I have solved it by the code: