./yolov9-tensorrt.exe ./model.engine ./1096.jpg Error Code 1: Cuda Runtime (an illegal memory access was encountered)

[molyswu]

Hi,
./yolov9-tensorrt.exe ./model.engine ./1096.jpg error as follower:
[04/30/2024-10:14:29] [E] [TRT] 1: [weightConverters.cu::nvinfer1::rt::cuda::padCaskConvWeight::166] Error Code 1: Cuda Runtime (an illegal memory access was encountered)
CUDA error 700 at F:\TensorRT-YOLOv9\src\yolov9.cpp:97Assertion failed: 0, file F:\TensorRT-YOLOv9\src\yolov9.cpp, line 97

1096.jpg :409*962

[spacewalk01]

Hi, width=409 and height=962 image right? I will test it. Thank you for informing

[spacewalk01]

I have tested it with 427x640 image and it worked:

[spacewalk01]

Will you send me your onnx model?

[molyswu]

The weights. zip file contains onnx and jpg images of the target copper surface

[molyswu]

onnx

[molyswu]

156M File size too big: 25 MB

[molyswu]

Do you have a cloud drive that allows you to upload ONNX files？

[spacewalk01]

Is it a costumized model? what is the number of classes of your model?

[molyswu]

Locate the incorrect statement：
CUDA_CHECK(cudaMemcpyAsync(cpu_output_buffer, gpu_buffers[1], num_detections * detection_attribute_size * sizeof(float), cudaMemcpyDeviceToHost, cuda_stream));
CUDA_CHECK(cudaStreamSynchronize(cuda_stream));

[molyswu]

It is costumized mode

[molyswu]

2 classes

[molyswu]

[spacewalk01]

detection_attribute_size * num_detections == 84 x 8400 for default 80 classes. So in your case, when the number of classes is 2, it should be 6 x 8400. Will you print out result? You can also analyze onnx file using https://netron.app application. It shows something like this:

[molyswu]

I can't find this line in your code：“detection_attribute_size * num_detections == 84 x 8400”，I don't know where to modify it

[molyswu]

CONF_THRESH = 0.3
IOU_THRESHOLD = 0.2
NUM_DETECTIONS = 8400
LEN_ONE_RESULT = 84

[molyswu]

python

[spacewalk01]

You trained your model with 2 classes but your onnx file has 80 classes. I think that is the problem. You need to set the number of classes of your model into 2 in reparameterize.py.

[spacewalk01]

https://github.com/spacewalk01/TensorRT-YOLOv9/blob/main/reparameterize.py

[spacewalk01]

In line 8, fix it as follows: model = Model(cfg, ch=3, nc=2, anchors=3)

[molyswu]

python yolov9_trt.py --engine yolov9-c.engine --data images --outdir output error:
Traceback (most recent call last):
File "C:\ProgramData\anaconda3\envs\yolov9\lib\threading.py", line 1016, in _bootstrap_inner
self.run()
File "F:\TensorRT-YOLOv9\python\yolov9_trt.py", line 394, in run
batch_image_raw, use_time = self.yolov9_wrapper.infer(self.yolov9_wrapper.get_raw_image(self.image_path_batch))
File "\TensorRT-YOLOv9\python\yolov9_trt.py", line 170, in infer
result_boxes, result_scores, result_classid = self.post_process(
File "F:\TensorRT-YOLOv9\python\yolov9_trt.py", line 293, in post_process
pred = np.reshape(output, (-1, NUM_DETECTIONS)).T
File "C:\ProgramData\anaconda3\envs\yolov9\lib\site-packages\numpy\core\fromnumeric.py", line 285, in reshape
return _wrapfunc(a, 'reshape', newshape, order=order)
File "C:\ProgramData\anaconda3\envs\yolov9\lib\site-packages\numpy\core\fromnumeric.py", line 59, in _wrapfunc
return bound(*args, **kwds)
ValueError: cannot reshape array of size 105966 into shape (8400)

[molyswu]

I modified the program and there is the same issue：In line 8, fix it as follows: model = Model(cfg, ch=3, nc=2, anchors=3)

[molyswu]

File "C:\ProgramData\anaconda3\envs\yolov9\lib\site-packages\numpy\core\fromnumeric.py", line 285, in reshape
return _wrapfunc(a, 'reshape', newshape, order=order)
File "C:\ProgramData\anaconda3\envs\yolov9\lib\site-packages\numpy\core\fromnumeric.py", line 59, in _wrapfunc
return bound(*args, **kwds)
ValueError: cannot reshape array of size 141288 into shape (8400)