GitLab

Better accuracy results using weight_decay = 2e-4 (90$ -> 92%), changing it to be the default value.

In reader.py, GFile returns an object of type bytes independently from the Python version. Therefore, the object must be decoded as string in any case. Removed the check of Python version

- 2 CPUs -> 2 GPUs
- is_cpu_ps=True

In-memory DataSet initiated without feed_dict are saved in TF graph def,
causing exported graph.pbtxt over-bloated.
Change to FixedLengthRecordDataSet and does a few refactorization.

- fixing typos
- reusing variables

A TODO was stated for adding LRN - Pending GPU support.
LRN was implemented by tensorflow/tensorflow@35df3ed43edabbc4ad1b2439bbc7de8917026d6e

Hyper-parameters taken from http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks