Reproduce Caffe results

hgaiser commented 6 years ago

This PR includes the changes from https://github.com/broadinstitute/keras-resnet/pull/34. In addition, this PR:

Disables bias for convolutional layers, as per the original implementation.
Configures BatchNormalization to use epsilon=1e-5, as per Caffe's default.
Changes some padding="same" to ZeroPadding2D. The difference is that ZeroPadding2D pads symmetrically (same as padding=3 in Caffe) whereas padding="same" pads at the end of a blob until it can maintain the correct shape.
Adds tools for importing / exporting ResNet models from Caffe (first convert Caffe model to .h5 file, then import into Keras).

With these changes I got near identical (within reasonable floating point precision) results compared to Caffe output for a few test images.

Exported models from Caffe can be downloaded here.

hgaiser commented 6 years ago

0x00b1 commented 6 years ago

Fantastic! Thanks, @hgaiser.

broadinstitute / keras-resnet