Input and Output shape change after conversion to Tflite from OpenVINO

[letdivedeep]

Hi @PINTO0309

I have an onnx model which i need to convert to edge TPU. My onnx model takes the input as [1,3,288,352 ] as seen below

And the output of ONNX as [1x101x56x4]

When i convert the model to open vino i use this cmd

python /opt/intel/openvino_2021/deployment_tools/model_optimizer/mo.py \
  --input_model model.onnx \
  --input_shape [1,3,288,352] \
  --output_dir openvino \
  --data_type FP32

And then convert the openviono to edge tpu using this cmd

openvino2tensorflow \
  --model_path openvino/model.xml \
  --model_output_path saved_model \
  --output_saved_model \
  --output_h5 \
  --output_pb \
  --output_no_quant_float32_tflite \
--output_edgetpu \
  --output_weight_quant_tflite \
  --output_full_integer_quant_tflite \
 --output_integer_quant_type uint8 \
--output_onnx \
 --onnx_opset 11 \
--output_tftrt \
--output_integer_quant_tflite

I got this change in the tflite files, where the input and output shape is changed Input change

Output change

Can you point me to reason why its changing the shape ?

[PINTO0309]

https://github.com/PINTO0309/openvino2tensorflow#6-7-replace-weights-or-constant-values-in-const-op If you want an accurate response, please provide a model. It is natural behavior for the structure to collapse in Reshape.

[letdivedeep]

@PINTO0309 Thanks for the reply

I have attached the model in this gdrive link

moreover, the results are very much hampered in the converted tflite from the openvino framework

[PINTO0309]

Delete the layers of the OpenVINO XML file in order from the end and debug where the values are wrong. https://drive.google.com/drive/folders/10oXYlctN_1RXIZAzhSLzuBtiDdHwOS_d?usp=sharing

• 6-7. Replace weights or constant values in Const OP, and add Transpose or Reshape just before the operation specified by layer_id https://github.com/PINTO0309/openvino2tensorflow#6-7-replace-weights-or-constant-values-in-const-op-and-add-transpose-or-reshape-just-before-the-operation-specified-by-layer_id

[letdivedeep]

@PINTO0309 for the reply

I will try your suggestions

[PINTO0309]

Since there is no progress, I'm closing it.