Order of input channels switched on ONNX

hovnatan commented 2 years ago

Issue Type

Others

OS

Mac OS

OS architecture

aarch64

Programming Language

C++

Framework

TensorFlowLite

Download URL for tflite file

https://github.com/google/mediapipe/blob/master/mediapipe/modules/face_detection/face_detection_short_range.tflite

Convert Script

tflite2tensorflow --model_path face_detection_short_range.tflite --flatc_path ../flatc --schema_path ../schema.fbs --output_pb
tflite2tensorflow --model_path face_detection_short_range.tflite --flatc_path ../flatc --schema_path ../schema.fbs --output_onnx --onnx_opset 9

Description

The input to the TFlite model is 1x128x128x3 but it is switched to 1x3x128x128 on ONNX output.

Relevant Log Output

INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
inputs:
{'dtype': <class 'numpy.float32'>,
 'index': 0,
 'name': 'input',
 'quantization': (0.0, 0),
 'quantization_parameters': {'quantized_dimension': 0,
                             'scales': array([], dtype=float32),
                             'zero_points': array([], dtype=int32)},
 'shape': array([  1, 128, 128,   3], dtype=int32),
 'shape_signature': array([  1, 128, 128,   3], dtype=int32),
 'sparsity_parameters': {}}
outputs:
{'dtype': <class 'numpy.float32'>,
 'index': 175,
 'name': 'regressors',
 'quantization': (0.0, 0),
 'quantization_parameters': {'quantized_dimension': 0,
                             'scales': array([], dtype=float32),
                             'zero_points': array([], dtype=int32)},
 'shape': array([  1, 896,  16], dtype=int32),
 'shape_signature': array([  1, 896,  16], dtype=int32),
 'sparsity_parameters': {}}
{'dtype': <class 'numpy.float32'>,
 'index': 174,
 'name': 'classificators',
 'quantization': (0.0, 0),
 'quantization_parameters': {'quantized_dimension': 0,
                             'scales': array([], dtype=float32),
                             'zero_points': array([], dtype=int32)},
 'shape': array([  1, 896,   1], dtype=int32),
 'shape_signature': array([  1, 896,   1], dtype=int32),
 'sparsity_parameters': {}}
ONNX convertion started =============================================================

ONNX convertion complete! - saved_model/model_float32.onnx

Source code for simple inference testing code

No response

PINTO0309 commented 2 years ago

@hovnatan Is your desired configuration the one in the image below?