I want to deploy tensorflow version of yolov2 on compute stick 2.I can successfully generate IR file.But when inference ,I got an error as following:
RuntimeError: [VPU] Unsupported precision I32for data strided_slice_2/stack_1/Output_0/Data__const
Then I open the .xml file and modify I32 to FP16,I got an error as following now:
RuntimeError: [VPU] Cannot convert layer "strided_slice_3" due to unsupported layer type "StridedSlice"
I successfully inferred on compute stick 2 based on darknet's yolov2. Now I want to deploy tensorflow version of yolov2 on compute stick 2. Then I got an above error.
What command did you use to convert your model?
When converting your model in the Model Optimizer, did you specify the command --data_type FP16 to run on the NCS2? You would need to convert the model in FP16 format before running inference.
Hi Sahira ,
I use the command to convert my model as follows:
python mo_tf.py --input_model E:\inference_graph_v2.pb -b 1 --data_type FP16
But still will report above error.
I have read the files about yolov2 instructions.The document is converted to tensorflow by darknet and then IR is generated. My yolov2 model is a parameter that is saved directly using tensorflow training, and then generates an IR file. The IR of my model has the above error. I have some questions I would like to ask you.
1、Can NCS2 support the tensorflow version of yolov2 deployment?
2、The above error is because NCS2 does not support the StridedSlice operation for the time being?
It looks like you're missing the Yolo v2 config files. Add the parameter --tensorflow_use_custom_operations_config with the corresponding configuration file (config files are found in the model optimizer directory under extensions > front > tf). For more information about converting your model, check this article out.
StridedSlice is currently not supported by the NCS2. Visit this page for more information about supported operations.
Please let me know if you have any further questions!