|
在最新的驱动以及版本中,将tflite的目标检测模型进行转化出现报错,错误log如下:
D RKNN output shape(concat): (1 1917 91)
D Process convert_scores_71 ...
D RKNN output shape(sigmoid): (1 1917 91)
D Process variable_77 ...
D RKNN output shape(variable): (1917 4)
D Process TFLite_Detection_PostProcess_72 ...
E Catch exception when loading tflite model: ./detect.tflite!
T Traceback (most recent call last):
T File "rknn/api/rknn_base.py", line 219, in rknn.api.rknn_base.RKNNBase.load_tflite
T File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 263, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build
T File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 281, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
T File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 289, in genexpr
T File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 289, in genexpr
T AttributeError: 'NoneType' object has no attribute 'to_string'
然后选择将版本降到1.0.0之前,重新进行测试,错误结果与上述相同。
在官方的dome中并未提供tflite检测模型的使用,只有pb模型的转化过程,请问是不是ssd的检测模型转化输入只能是pb格式。
如果只能是pb格式,那么量化图设置多少才能通过转化,因为当量化图设置少量时,转化后模型的精度极大损失与服务器端相差甚远,但是当量化图设置过多(200张时就会有这样的结果),会直接出现kill的情况。
谢谢
|
|