Unsupported Slice operation using Tensorflow

[idata]

Hi,

I'm trying to deploy my own Tensorflow network on the Movidius NCS and I'm experiencing some issues with the Slice operation.I tried a simple slicing operation (e.g., extracting the first channel from an image)

output = tf.slice(input, [0, 0, 0, 0], [-1, -1, -1, 1])

but when I try to profile the network i get this error: [Error 5] Toolkit Error: Stage Details Not Supported: Slice type not supported

I'm running 1.11.00 release, according to https://github.com/movidius/ncsdk/releases/tag/v1.11.00.04 the slice operation seems to be supported.Thanks

[idata]

@mpoggi Thanks for bringing this to our attention. We should have support for the Tensorflow Slice operation. Can you provide the code you tested with the steps to reproduce the issue so that we can help debug the issue that you are facing? Thanks in advance.

[idata]

Thanks for yout help, you can find my code at https://pastebin.com/RV6ZvLtu (myNet.py script).

The code contains a flag, USING_SLICE: if set to True, the network uses a Slice op, otherwise it does not.I simply tried to profile the network, running

python myNet.py

mvNCProfile -s 12 output/myNet.meta -in=input -on=result/add

If USING_SLICE is set to False, the profile runs smooth, while I get the "Error 5" if I set it to True.

Thanks

[idata]

@Tome_at_Intel I found the problem, "-1" indexing is not supported and I need to pass the dimension of the tensor:

output = tf.slice(input, [0, 0, 0, 0], [output.get_shape().as_list()[0], output.get_shape().as_list()[1], output.get_shape().as_list()[2], 1])

Moreover, I tried to slice on dimensions 1 and 2 (i.e., reducing height and width), but it seems to be not supported according to /opt/movidius/NCSDK/ncsdk-x86_64/tk/Controllers/TensorFlowParser.py, line 1489. Any plan to make it supported?

Thanks

[idata]

@mpoggi With regards to the Slice support, we plan add additional layer support as needed by our developers. We can't make any promises, we will keep all feature requests in mind when building future releases. Thank you for your understanding and patience.