search

zombie0117 / yolov3-tiny-onnx-TensorRT

convert your yolov3-tiny model to trt model
90 stars 34 forks source link

onnx_to_tensorrt.py get reshape error #12

Open dhuib2016 opened 4 years ago

dhuib2016 commented 4 years ago

when I run onnx_to_tensorrt.py, some error appeared like this:

Traceback (most recent call last): File "onnx_to_tensorrt.py", line 226, in main() File "onnx_to_tensorrt.py", line 204, in main trt_outputs = [output.reshape(shape) for output, shape in zip(trt_outputs, output_shapes)] ValueError: cannot reshape array of size 43095 into shape (1,18,13,13)

My running environment : python 2.7 onnx 1.4.1

Ndron commented 4 years ago

It looks like u use 80 clases , change number of filters 18->255 (filters = (5+clases)*3)

gzchenjiajun commented 4 years ago

@Ndron Where can I change this?

Ndron commented 4 years ago

@gzchenjiajun in your config file , before yolo layers (3 rows) . In my .cfg lines 603,689,776

git-manager commented 4 years ago

when I run onnx_to_tensorrt.py, some error appeared like this:

Traceback (most recent call last): File "onnx_to_tensorrt.py", line 226, in main() File "onnx_to_tensorrt.py", line 204, in main trt_outputs = [output.reshape(shape) for output, shape in zip(trt_outputs, output_shapes)] ValueError: cannot reshape array of size 43095 into shape (1,18,13,13)

My running environment : python 2.7 onnx 1.4.1

Did you solve this problem? i have a same question.

xiaoxiaowannote commented 4 years ago

@git-manager Did you solve this problem?

zhzhang747 commented 4 years ago

overriding the output_shapes on line 171 of onnx_to_tensorrt.py gave me valid results

output_shapes = output_shapes_dic[str(input_size)] <------- original

output_shapes = [(batch_size, 255, 13, 13), (batch_size, 255, 26, 26)] <------- Work Around

Originally posted by @SurionAndrew in https://github.com/zombie0117/yolov3-tiny-onnx-TensorRT/issues/2#issuecomment-540447480

But I don't know why.

Damon0626 commented 4 years ago

@zhzhang747 After tried on my devices, I observed the pattern. The output_shapes pattern maybe [(batch_size, filters, 13, 13), (batch_size, fileters, 26, 26)]

For example, if the classes of the model is 80, the size of trt_output[0] is 43095 ===>( (5+80) x 3 x 13 x 13) , the size of trt_output[1] is 172380 === >((5 + 80) x 3 x 26 x 26)

My own tiny model classes are 6, after runing the inference code, get the parms as follows: size of trt_output[0] is 5576, the same as above ((5+6) x 3 x 13 x13) size of trt_output[1] are 22308. ((5+6) x 3 x 26 x 26).