cnn-fixations

cnn-fixations copied to clipboard

I added additional notebook for demo of tf-resnet support. Please review the result, and merge into this repo.

Tensorflow support will definitely help deep learning community.

Thank you for sharing your base code.

Hi Insikk,

I am trying to add support for Inception-v3 through tf-slim. Could you please tell me where can I get the layer mappings similar to "resnet_v1_101_caffe_to_tf" for Inception-v3?

Thanks, Tarang

@tchugh The resnet_v1_101 weight mapping is available because the resnet v1 tensorflow implementation is borrowing caffe's weight by design. What I did was comparing two model architecture of resnet in both tensorflow and caffe, then mapping layer by layer. Of course, there was minor modification that was handled by some tinkering.

However, inception networks are mainly built on tensorflow directly. So I don't think there is an easy way. You might able to find a caffe implementation of inception-v3 which uses its weight trained on tensorflow. Then, you can follow what I did for resnet cnn fixation.