Keras Version

[nattari]

Thanks, Aniket for sharing the code. While using your implementation, I get an error at the following line in vgg_cbcnn.py: x = merge(compact_bilinear_arg_list, mode=compact_bilinear, name='compact_bilinear', output_shape=output_shape_cb). It gives a Typeerror (module object is not callable). By doing some search, I found that the new version of Keras has replaced merge with concatenate, but this did not help either. Do you have any insights?

[nattari]

On trying the implementation on a linux machine, I do not get version related error but I get another error. Please find attached the screenshot for the same.

[aniket03]

Hi @nattari Sorry for a late reply

1. Regarding point (1), yes layer merge has been replaced with concatenate and due to that you might be facing the error. For record I used the version Keras==2.1.4 while training the model, which allows the usage of merge layer.

2. For point 2, I am not too sure, why you got this error. I guess you are using a batch size of 64, and have 200 classes in your dataset. Could you pls check what is the shape of labels array that is being returned by train_generator or val_generator. Also if required (which IMO should not) you could try changing this line here to ensure that the target labels numpy array has the correct shape as required.

[liuqiang3]

I meet the same issue,when i using Lambda replaced merge ,the code can run.