Optimize graph & graph transform tools do not support NCHW

[khanrc]

I tried optimizing graph using both Graph transform tool and Optimize graph for inference. Both cases produced the same error because the fused batchnorm used not NCHW, but NHWC. I've got the error like this:

InvalidArgumentError (see above for traceback): Must provide as many biases as the channel dimension of the input tensor: [256] vs. 19 in [1,256,19,19]
	 [[Node: prefix/convblock/BatchNorm/FusedBatchNorm = BiasAdd[T=DT_FLOAT, data_format="NHWC", _device="/job:localhost/replica:0/task:0/device:GPU:0"](prefix/convblock/Conv2D, prefix/convblock/Conv2D_bn_offset)]

Although NCHW is faster than NHWC in GPU environment, why the tools do not support NCHW?

[aselle]

Could you provide a reproducible test case of what exactly you tried to do? Generally speaking, I think a lot of the tooling after training requires NHWC, as that was the only format when those were written. If you could provide a reproducible test case, we could work to improve it. @petewarden, do you have any other comments?

[khanrc]

I tried:

python tensorflow/python/tools/optimize_for_inference.py \
--input ./ckpt/frozen_model.pb \
--output ./ckpt/optimized_model.pb \
--frozen_graph true \
--input_names Placeholder \
--output_names policy_head/softmax,value_head/value/Tanh

and

tensorflow/bazel-bin/tensorflow/tools/graph_transforms/transform_graph \
--in_graph='./ckpt/frozen_model.pb' \
--out_graph='./ckpt/transformed_model.pb' \
--inputs='Placeholder' \
--outputs='policy_head/softmax,value_head/value/Tanh' \
--transforms='
fold_constants(ignore_errors=true)
fold_batch_norms
fold_old_batch_norms
fuse_pad_and_conv
fuse_resize_and_conv
fuse_resize_pad_and_conv
'

In both cases, the error occurred in fused batchnorm. The frozen model worked well, but the optimized model and transformed model emitted error.

[tensorflowbutler]

It has been 14 days with no activity and this issue has an assignee.Please update the label and/or status accordingly.

[tensorflowbutler]

Nagging Assigneee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[tensorflowbutler]

Nagging Assignee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.