Kai Han

> Also, I am getting this error, don't know why? > > ``` > ValueError Traceback (most recent call last) > in () > 11 model_raw.summary() > 12 share_fea_map =...

The categorical_crossentropy loss in Keras may changed its API. you can try to set `from_logits=True` in `keras.backend.categorical_crossentropy(target, output, from_logits=False, axis=-1)`.

The `region_score_map` in https://github.com/iamhankai/attribute-aware-attention/blob/master/cub_demo.py#L112 is the attention values. You can visualize it.

Thanks for you attention. Please refer to https://github.com/huawei-noah/CV-Backbones/tree/master/tnt_pytorch for training.

If you need pretrained weights, you can use this: https://www.github.com/rwightman/pytorch-image-models/tree/master/timm/models/tnt.py

It seems that `torch::CPU(torch::kFloat32)` has been deprecated. Use the updated class or lower version of pytorch.

Thanks for your attention. The input of A3M attention module is feature maps from CNN, you may change the code to input features of transformer or LSTM, refer to https://github.com/iamhankai/attribute-aware-attention/blob/4b4cf873d6e398f1e64891dbc34ccb8fbd891f30/cub_demo.py#L109

refer to https://github.com/iamhankai/attribute-aware-attention/issues/12#issuecomment-494344421

Yes

Try different `image_data_format` and `backend`.