tuyunbin / scorer Goto Github PK

Running the training script seems lead to an error.
The command: python train.py --cfg configs/dynamic/transformer.yaml
The error:

Traceback (most recent call last):
  File "train.py", line 223, in <module>
    torch.clamp_(change_detector.cross_layer[-1].self_attention.logit_scale.data, max=np.log(100))
  File "/hd1/mwt/miniconda3/envs/scorer/lib/python3.8/site-packages/torch/nn/modules/module.py", line 947, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'ChangeDetector' object has no attribute 'cross_layer'

The related code:

I'm trying to train the model with the given command python train.py --cfg configs/dynamic/transformer.yaml, but an error occurs. The error message is like: (last few lines)

  File "/hd1/mwt/SCORER/datasets/rcc_dataset_transformer.py", line 39, in __init__
    self.word_to_idx = json.load(open(self.vocab_json, 'r'))
FileNotFoundError: [Errno 2] No such file or directory: './data/transformer_vocab.json'

Seems a configuration error, I see vocab.json in the ./data directory, is that the correct file to use?

Hi, I found in the train.py, the generator is not being optimized, and this leads to non-decreasing cycle loss. Is this a bug or an intended design?

First thanks for this great work, it is very valuable to see your code and be able to run it,

I have reproduced your work on CLEVR-Change. Now I am trying to do it on a smaller dataset and as i see the results on Image Editing Request I think you could have some valuable insights if you are okay with sharing :)

1. Will you release the code for training on Image Editing Request ? Most of my questions could be answered with that I guess !
2. What was the training time with your GPU on Image Editing Request ?
3. What was the hyperparameters (learning rate, batch_size, other thing you can think of and you struggled with) ?

For now, i am training on a dataset of the size of IER. SCORER seems to struggle and I want to be sure I am not doing anything wrong to be able to compare my thing with it fairly.

Again, thanks for this work and for your time