Error 5 with Slice layer on a custom TF network

idata · ‎12-01-2017

Hi,

I am training a TF Net with TF v.1.3.0 and python2.7

When I try to compile the net with mvNCCompile -in img_n_cond -on gen/gen_op model_1.ckpt.meta

I get the following output

mvNCCompile v02.00, Copyright @ Movidius Ltd 2016

/usr/local/lib/python3.5/dist-packages/tensorflow/python/util/tf_inspect.py:45: DeprecationWarning: inspect.getargspec() is deprecated, use inspect.signature() instead
  if d.decorator_argspec is not None), _inspect.getargspec(target))
[Error 5] Toolkit Error: Stage Details Not Supported: Slice type not supported

I have the latest version of the ncsdk,

here's a snippet of the part of the net in question

        self.img_n_cond = tf.get_variable(dtype = tf.float32, shape = [1, conf.img_size, 2*conf.img_size, conf.img_channel], name = "img_n_cond")
        print(self.img_n_cond.name)
        self.image = tf.slice(self.img_n_cond, begin = [0, 0, 0, 0], size = [1, conf.img_size, conf.img_size, conf.img_channel])
        self.cond = tf.slice(self.img_n_cond, begin = [0, 0, conf.img_size, 0], size = [1, conf.img_size, conf.img_size, conf.img_channel])

I can see from the release notes that Slice is indeed supported, so I am not sure what the issue is.

idata · ‎12-21-2017

Hi,

I'm having the same issue, did you solve it?

idata · ‎04-26-2018

Can someone finally comment on whether tf.slice is supported or not? The release says it is, but I get "[Error 5] Toolkit Error: Stage Details Not Supported: Slice type not supported"

idata · ‎04-26-2018

Can someone finally comment on whether tf.slice is supported or not? The release note says it is, but I get "[Error 5] Toolkit Error: Stage Details Not Supported: Slice type not supported"

idata · ‎04-27-2018

@mumes Slicing is supported but it expects all of the arguments to be a 4 dimensional input shape. The slice start point has to be the start of the tensor and the slice size has to be smaller than the input tensor. You can view the slice details in TensorFlowParser.py in /opt/movidius/NCSDK/ncsdk-x86_64/tk/Controllers

idata · ‎04-27-2018

Thanks for the answer @Tome_at_Intel !

The release notes mention that a "Crop" operation is supported. Is that a Tensorflow operation?

idata · ‎05-18-2018

@Tome_at_Intel Now that Slicing is supported, there's a really easy trick to have multiple inputs to the graph. Let's say you want to input two images image_1 and image_2 (shape = [32, 32, 3]) to the graph.

This could in cases like style transfer or something like Cycle-GAN.

You can set -in to be a node formed by concating the two images and taking slices inside the graph.