Pruning: Training with Near 100% Target Sparsity Fails

[alanchiao]

Describe the bug
Pruning with high target sparsity (e.g. 0.99) causes a error.

System information

TensorFlow installed from (source or binary):

TensorFlow version: any

TensorFlow Model Optimization version: 0.2.1

Python version: any

Describe the expected behavior
Target sparsity of 0.99 should work.

Describe the current behavior
Training errors out with something like:

InvalidArgumentError: indices = -1 is not in [0, 40)
[[{{node prune_low_magnitude_dense_1/cond/cond/pruning_ops/GatherV2}}]]

Code to reproduce the issue
testPruneWithHighSparsity_Fails in prune_integration_test.py

Can search for "model-optimization/issues/215" in codebase to find unit test also.