Unable to convert a custom yolov3-spp model

hunglc007 / tensorflow-yolov4-tflite

YOLOv4, YOLOv4-tiny, YOLOv3, YOLOv3-tiny Implemented in Tensorflow 2.0, Android. Convert YOLO v4 .weights tensorflow, tensorrt and tflite

MIT License

2.23k stars 1.24k forks source link

[{'name': 'input_1', 'index': 0, 'shape': array([ 1, 416, 416, 3], dtype=int32), 'shape_signature': array([ -1, 416, 416, 3], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}] [{'name': 'Identity', 'index': 523, 'shape': array([ 1, 10647, 4], dtype=int32), 'shape_signature': array([ 1, -1, 4], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'Identity_1', 'index': 502, 'shape': array([ 1, 10647, 80], dtype=int32), 'shape_signature': array([ 1, -1, 80], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}] [array([[[nan, nan, nan, nan], [nan, nan, nan, nan], [nan, nan, nan, nan], ..., [nan, nan, nan, nan], [nan, nan, nan, nan], [nan, nan, nan, nan]]], dtype=float32), array([[[nan, nan, nan, ..., nan, nan, nan], [nan, nan, nan, ..., nan, nan, nan], [nan, nan, nan, ..., nan, nan, nan], ..., [nan, nan, nan, ..., nan, nan, nan], [nan, nan, nan, ..., nan, nan, nan], [nan, nan, nan, ..., nan, nan, nan]]], dtype=float32)]

When I try to convert a custom model to tflite model, I am getting all nan


[{'name': 'input_1', 'index': 0, 'shape': array([  1, 416, 416,   3], dtype=int32), 'shape_signature': array([ -1, 416, 416,   3], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
[{'name': 'Identity', 'index': 523, 'shape': array([    1, 10647,     4], dtype=int32), 'shape_signature': array([ 1, -1,  4], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'Identity_1', 'index': 502, 'shape': array([    1, 10647,    80], dtype=int32), 'shape_signature': array([ 1, -1, 80], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
[array([[[nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        ...,
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan]]], dtype=float32), array([[[nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        ...,
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan]]], dtype=float32)]

When I try to run detect on the sample image, there are no inferences. This custom model is trained on ultralytics/yolov3 PS: This works for the yolov4 model though.

Have you sloved this problem and would mind giving some suggestions for me, please!

hunglc007 / tensorflow-yolov4-tflite

Unable to convert a custom yolov3-spp model #189