Module 'ModuleDict' has no attribute 'similarity'

[linzqin]

Hi @peikexin9

When I try to get the function similarity by running python command/inference/get_embedding.py, I got two errors.

1. The first error occurs in trex.py line 123

 @classmethod
    def setup_task(cls, args, **kwargs):
        # paths = utils.split_paths(args.data)

        paths = os.listdir(args.data)
        assert len(paths) > 0
        if len(paths) != len(configs.fields):
            print('ERROR: invalid paths:', paths)
            raise ValueError()

Code discovery: the length of paths at paths = os.listdir('data-bin/similarity') is required to be equal to configs.fields. But the paths under paths are input0, input1, label, and it is the subdirectory lengths of input0 and input1 that are equal to configs.fileds. data-bin/similarity related content generation is in command/finetune/preprocess preprocess.py, there is nothing wrong with the directory structure， So I chose to replace data-bin/similarity with data-bin/similarity/input0 in get_embedding.py and skipped this error. I don't know if this is the correct processing.

2. The second error is that when I continue to run to trex_script = torch.jit.script(trex.model), I get an error: Exception: RuntimeError Module 'ModuleDict' has no attribute 'similarity' in model.py line 238.

if classification_head_name is not None:
            x = {
                'features': self.classification_heads.similarity(x['features'])
            }
        return x, extra

I Check the ModuleDict manual. Indeed there is no such attribute.Also, have no attributes like similarity_pair.

def similarity_pair(self, concat_in: torch.Tensor):
        return self.classification_heads.similarity_pair(concat_in)

The Pytorch version is 1.13.1.

Thank you!

[BrainLyh]

@linzqin Hi! Have you solved this problem? I met the same problem yesterday，i do not sure why was happend(but I strongly suspect it's because I deleted some file in data-bin/similarity by mistake).Anyway, it works when i delete all content(In addition to the initial file) in data-* dir, and rerun the finetune.sh with the example data. Trace errors in source code is very tough...