Joel

Same issue here, about 5k extractions from raw htmls

Yes, it's obviously overfitting according to your log. You can put some regularization to prevent it.

It seems like you provide an empty batch to the network, therefore the shape is `()`.

You need to check the batch data provide by `next_batch`. If it returns empty batch, you may need to validate the path to image files.

Perhaps you can refer to my [repo](https://github.com/joelthchao/tensorflow-finetune-flickr-style). It is modified based on caffe-tensorflow.