I have a segmentation neural network using ResNet with some resize nearest neighbor layers to upsample intermediate images to the original resolution at the output. This network has been frozen into Tensorflow graph .pb file (TF 1.14.0). After that I am able to convert it to intermediate representation OpenVino using mo_tf.py with no errors. Now, I use benchmark_app (C++) to run inference using the exported network to obtain timing results. When I specify CPU device for inference, no errors are obtained and all good. However, when I switch to GPU, I obtain the following error:
[Step 7/11] Loading the model to the device
[ ERROR ] Unsupported scale in layer depth_prediction/ResizeNearestNeighbor
I attached the model definition in .xml format so that you can repeat the error.
ResizeNearestNeighbor is claimed to be supported in the table of layers, however this is not the case for me.
OpenVino version used: 2019.3.376. Ubuntu 18.04 64 bit. GPU Intel(R) Gen9 HD Graphics (Skylake GT2 [HD Graphics 520] (rev 07)). CPU Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz.
I'd like to try and reproduce this error - can you please give me the exact command you used to convert your model and the command you used to run your model through the Inference Engine.
to convert my frozen TF graph, I used the python tools from OpenVINO (folder openvino_2019.3.376/deployment_tools/model_optimizer) with the following command:
python3 mo_tf.py --input_model /mnt/data-local/openVino/dima/models/frozen_uncertainty_graph.pb
To run my model through the inference engine, I used the pre-built app benchmark_app from Intel samples, which I called with the following arguments:
./benchmark_app -d GPU -i /home/dima/intel/openvino_2019.3.376/deployment_tools/demo/car.png -m /home/dima/models/frozen_uncertainty_graph.xml -pc -niter 1000
When using -d CPU option instead, no error message is received.