ResNet Fine-Tuning

[Simsso]

(sub-task of #23)

After training a ResNet model from scratch (#31) and retraining Inception (#29) have not yielded satisfying results, we will try to fine-tune a ResNet with the weights provided here.

Wiki docs

The overall goal is the extraction of the ResNet model from the provided code. Then it can be used for experiments such as the insertion of layers, addition of VQ (#25), etc. All of that with accuracies from 60 to 70%.

[Simsso]

Model Transfer

Transfer of the ALP model to our source code.

Links

• project template
• original ResNet model code

Findings / Minutes

• we managed to transfer the original ResNet code (see above) to a cleaner implementation using the template (also above)
• after a few fixes, we restored the pre-trained weights and achieved the 70% accuracy
• next, we added an operation with a custom variable to the graph, which will be required once we add experiments to make the model more robust
  • as expected, it initially didn't work, because the variable was declared inside of the same scope that the rest of the graph was initialized in - so the tf.train.Saver attempted to restore its value from the pre-trained weights, too
  • hence, we searched for a way to initialize our own variables in a different scope:

def init_model():
   with tf.variable_scope('custom_scope') as custom_scope:
      pass 
   with tf.variable_scope('resnet_v2_50', ...):   #  this scope is restored 
      #  resnet graph definition
      with tf.variable_scope(custom_scope, auxiliary_name_scope=False):
         # this will *not* be a child of the resnet_v2_50 scope
         net = some_custom_op(net)
      # rest of the resnet graph definition

[Simsso]

This issues does not have clear work items. Please go ahead and close it, if you agree, @FlorianPfisterer.

[FlorianPfisterer]

I agree, we track the progress on the overall task in other issues, like #42, #47, #49, #48.