GitLab

Add an example showing how to train the MNIST model with eager execution
enabled.

(This change requires changes to TensorFlow made after
the 1.6 release branch was cut, i.e., will require a build
from source or TensorFlow 1.7+)

The long path doesn't work for TensorFlow 1.6 because the .python package under cluster_resolver doesn't exist.

* Refactoring input function

* Updating name of variable

The `sparse` version is more efficient anyway.

I'm returning the labels shape [1] instead of []
because tf.accuracy fails otherwise.

Add multi-GPU option to MNIST

* Using nightly TF docker

* Explicitly pull image

* Update readme

* Respond to CR

The `sparse` version is more efficient anyway.

I'm returning the labels shape [1] instead of []
because tf.accuracy fails otherwise.

This reverts commit e8726907.

Add multi-GPU flag to MNIST and allow for setting of replicated optimizer and model_fn

Opening gzipped datasets in binary, read-only mode fixes the issue

This is a step towards merging the example in
https://github.com/tensorflow/tpu-demos/tree/master/cloud_tpu/models/mnist
with this repository, so we have a single model definition for
training across CPU/GPU/eager execution/TPU.

The change to dataset.py is so that the raw data can be read
from cloud storage systems (like GCS and S3).

- Prior to this change, the use of tf.data.Dataset essentially embedded
  the entire training/evaluation dataset into the graph as a constant,
  leading to unnecessarily humungous graphs (Fixes #3017)
- Also, use batching on the evaluation dataset to allow
  evaluation on GPUs that cannot fit the entire evaluation dataset in
  memory (Fixes #3046)

This will make it easier to share the model definition with
eager execution and TPU demos without any side effects of running
unnecessary code on module import.

https://github.com/tensorflow/models/pull/3024

- Use the object-oriented tf.layers API instead of the functional one.
  The object-oriented API is particularly useful when using the model
  with eager execution.
- Update unittest to train, evaluate, and predict using the model.
- Add a micro-benchmark for measuring step-time.
  The parameters (batch_size, num_steps etc.) have NOT been tuned,
  the purpose with this code is mostly to illustrate how model
  benchmarks may be written.

These changes are made as a step towards consolidating model definitions
for different TensorFlow features (like eager execution and support for
TPUs in
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/eager/python/examples/mnist
and
https://github.com/tensorflow/tpu-demos/tree/master/cloud_tpu/models/mnist

With examples, and updates to the README