Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Feldman__Daniel
Beginner
87 Views

Tensorflow slicing operation causes IR generation failure

I was trying to convert a model which uses tf.tile and tf.slice operations after a convolution layer, and got an error while generating the IR:

python3 mo_tf.py --input_model frozen_out/frozen_graph.pb --output_dir out_IR

Model Optimizer arguments

    Batch:     1
    Precision of IR:     FP32
    Enable fusing:     True
    Enable gfusing:     True
    Names of input layers:     inherited from the model
    Path to the Input Model:     frozen_out/frozen_graph.pb
    Input shapes:     inherited from the model
    Log level:     ERROR
    Mean values:     ()
    IR output name:     inherited from the model
    Names of output layers:     inherited from the model
    Path for generated IR:     out_IR
    Reverse input channels:     False
    Scale factor:     None
    Scale values:     ()
    Version:     0.3.61.37271eb9
    Input model in text protobuf format:     False
    Offload unsupported operations:     False
    Path to model dump for TensorBoard:     None
    Update the configuration file with input/output node names:     None
    Operations to offload:     None
    Patterns to offload:     None
    Use the config file:     None

[ ERROR ]  Shape is not defined for output 0 of "output".
[ ERROR ]  Cannot infer shapes or values for node "output".
[ ERROR ]  Not all output shapes were inferred or fully defined for node "output". For more information please refer to Model Optimizer FAQ, question #40.

I looked at FAQ#40, it suggests --input_shape option as a fix, but all the input shapes of the tensor “output” are valid.
 

My environment is Ubuntu 17.10 x64 with python3.6, tensorflow 1.8, CV-SDK 2018.0.234
 

I attached both the frozen graph and the tensorflow code which was used for creating the graph. In addition, I added full debug log of the error.
 

According to the Deployment Toolkit Developer Guide all my model operations should be supported.
 

Any help would be appreciated.

 
0 Kudos
5 Replies
Zhen_Z_Intel
Employee
87 Views

Hi Daniel,

It seems your shape is not defined, please try to use this option --input_shape to override model input shape.

--input_shape [N,H,W,C] for TF, please follow developer guide to learn more. Thank you.

Best regards,
Fiona

Feldman__Daniel
Beginner
87 Views

Thanks Fiona.

But as you can see in the full log I attached, the input shape is well defined:

[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:71 ]  Partial infer for input
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:72 ]  Op: Placeholder
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:77 ]  Inputs:
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:79 ]  Outputs:
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:39 ]  output[0]: shape = [ 1 4 4 20], value = <UNKNOWN>

As you suggested, I tried to set it with the --input_shape options but the error is the same.
(command: python3 mo_tf.py --input_model frozen_out/frozen_graph.pb --output_dir out_IR --input_shape=[1,4,4,20])

Furthermore, as I mentioned before, all the input shapes of the tensor “output” are valid:

[ 2018-05-29 09:41:18,178 ] [ DEBUG ] [ infer:72 ]  Op: Slice
[ 2018-05-29 09:41:18,178 ] [ DEBUG ] [ infer:77 ]  Inputs:
[ 2018-05-29 09:41:18,178 ] [ DEBUG ] [ infer:39 ]  input[0]: shape = [ 1 4 4 20], value = <UNKNOWN>
[ 2018-05-29 09:41:18,178 ] [ DEBUG ] [ infer:39 ]  input[1]: shape = [4], value = [0 0 0 0]
[ 2018-05-29 09:41:18,179 ] [ DEBUG ] [ infer:39 ]  input[2]: shape = [4], value = [ 1 4 4 10]
[ 2018-05-29 09:41:18,179 ] [ DEBUG ] [ infer:79 ]  Outputs:
[ 2018-05-29 09:41:18,179 ] [ DEBUG ] [ infer:39 ]  output[0]: shape = <UNKNOWN>, value = <UNKNOWN>

Monique_J_Intel
Employee
87 Views

Hi Daniel,

It seems that you may want to supply the command line parameter --input <node name> so that Model Optimizer will know what node you are providing the shape for.

Kind Regards,

Monique Jones

Feldman__Daniel
Beginner
87 Views

Hi Monique,


The --input option is not relevant in this case. The input node for the optimizer is the first node in the tensorflow graph, so there is no need to set --input.


In addition, you can see in my previous post, the input placeholder is well defined:

[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:71 ]  Partial infer for input
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:72 ]  Op: Placeholder
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:77 ]  Inputs:
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:79 ]  Outputs:
[ 2018-05-29 09:41:18,172 ] [ DEBUG ] [ infer:39 ]  output[0]: shape = [ 1 4 4 20], value = <UNKNOWN>

Nevertheless, I added the input node explicitly -  as expected the error is the same.
(python3 mo_tf.py --input_model frozen_out/frozen_graph.pb --input=input --input_shape=[1,4,4,20] --output_dir out_IR)

Zhen_Z_Intel
Employee
87 Views

Thank you, we will investigate to see if the problem of slide layer.

Fiona