[Intel MKL] Adding Quantized Convolution (Int8) - Part 3

[mahmoud-abuzaina]

This PR is part 3 of four PRs that add quantized version of convolution using MKL-DNN. This part adds changes that are required to quantize a graph and it enables fusions of MKL quantized convolution.

Please note that the below PRs replace this one #19425
Part 1: #21483
Part 2: #21456
Part 3: #21465
Part 4: #21466

[dspinxd]

在 2018年8月8日，08:28，Mahmoud Abuzaina <notifications@github.com<mailto:notifications@github.com>> 写道： This PR is part 3 of four PRs that add quantized version of convolution using MKL-DNN. This part adds changes that are required to quantize a graph and it enables fusions of MKL quantized convolution.

…

________________________________ You can view, comment on, or merge this pull request online at: #21465 Commit Summary * Adding int8 conv part3 File Changes * M tensorflow/tools/graph_transforms/BUILD<https://github.com/tensorflow/tensorflow/pull/21465/files#diff-0> (1) * A tensorflow/tools/graph_transforms/fuse_quantized_convolution.cc<https://github.com/tensorflow/tensorflow/pull/21465/files#diff-1> (220) * M tensorflow/tools/quantization/quantize_graph.py<https://github.com/tensorflow/tensorflow/pull/21465/files#diff-2> (348) Patch Links: * https://github.com/tensorflow/tensorflow/pull/21465.patch * https://github.com/tensorflow/tensorflow/pull/21465.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#21465>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ARHDlNRwr6aBck0dOv5vV2agM9YqImu7ks5uOjCzgaJpZM4VzFan>.

[tensorflowbutler]

Nagging Reviewer @tatianashp, @raghuraman-k: You have been added as a reviewer to this pull request. Please add your review or reassign. It has been 50 days with no activity and the awaiting review label has been applied.

[raghuraman-k]

Please remove quantize_graph from this PR. It will be best to host it separately.

[mdfaijul]

@raghuraman-k @drpngx removed quantize_graph.py

[mdfaijul]

@drpngx Tests looks fine. Ready to go?

[tatianashp]

@suharshs Can you take a look? Is this PR still relevant?

[drpngx]

@mahmoud-abuzaina please return an error (invalid argument) for when the tensor parsing fails and not implemented if the dtype is wrong:

CHECK(float_tensor.FromProto(float_tensor_proto));
CHECK_EQ(float_tensor.dtype(), DT_FLOAT);

We do not allow CHECK in new code.

[mdfaijul]

@drpngx Removed CHECK and CHECK_EQ, added errors::InvalidArgument and errors::Unimplemented. Could you please check?