ResNet Model Function

[Simsso]

171 additions and 247 deletions.

Well, that's a word! This PR contains a refactoring of the ResNet code which we've copied and pasted from the TF code base. Only one file has been modified.

The new ResNet code has the following properties:

• Only contains code that's need in our application
• ALP weights can be imported
• It's easy to read in a sequential manner; the core function looks like that:

  x = ResNet.first_conv(x)
  x = ResNet.v2_block(x, 'block1', base_depth=64, num_units=3, stride=2)
  x = ResNet.v2_block(x, 'block2', base_depth=128, num_units=4, stride=2)
  x = ResNet.v2_block(x, 'block3', base_depth=256, num_units=6, stride=2)
  x = ResNet.v2_block(x, 'block4', base_depth=512, num_units=3, stride=1)
  x = ResNet.batch_norm(x)
  return self.global_avg_pooling(x)

• New layers can be added with ease (not contained in the PR, but tested)

A few method comments are not complete yet.

Merging this PR resolves #48.