Issue description

The generated 32x32 resnet model in caffe2/python/models/resnet.py does not reflect the architecture of He et al (2015). For example, if the num_groups parameter (supposedly n in He et al. 2015) is set to 3, it should generate ResNet18 having 6*n (+ 2) layers. However, it generates a network having 36 (+2) layers (thus ResNet36).

The bug is in line 256 of https://github.com/pytorch/pytorch/blob/master/caffe2/python/models/resnet.py
The line for blockidx in range(0, 2 * num_groups): should be for blockidx in range(0, num_groups): as each simple block already contains 2 conv layers.