Help needed to convert yolov4-tiny model to tensorRT engine (DS 5)

stas.lisetsky · February 24, 2021, 1:40am

hardware : x64, rtx 2060
cuda 10.2
deepstream 5.0.1
TRT: 7.0.0.11
driver: 450.102.04

Hello, I am using GitHub - Tianxiaomo/pytorch-YOLOv4: PyTorch ,ONNX and TensorRT implementation of YOLOv4
to make an engine file from cfg/weights

The problem is - the engine is producing nonsensical inference results (zero or infinite-sized bboxes, all confidences are 1)
A very similar-looking issue was described here:

And the this is seemingly happening even before I get ONNX model -
the demo.py in pytorch-YOLOv4 crashes because inference result has infinite-sized bboxes.

I have verified my model/cfg files with both darknet and a simple python opencv-dnn based script.
In both cases I get correct result (for the same image that I used in the demo.py script)

caveat: my opencv-dnn python script was also failing with the same infinity issue, and I fixed it by accident by trying a different opencv version (currently - 3.4.8.29)
I was hoping this would also fix the issue with pytorch-YOLOv4, but no.

Apparently, there’s some weird incompatibility going on and I’ve tried different pytorch versions (1.3, 1.4, 1.5, 1.7.1)

I would appreciate any ideas about what I can try.
thank you.

AastaLLL · February 25, 2021, 3:21am

Hi,

It’s also recommended to check this issue with the GitHub author:
https://github.com/Tianxiaomo/pytorch-YOLOv4

Do you run the inference with a custom YOLOv4 model?
If yes, could you check if there is any configuration need to be updated first?

For example:

github.com

Tianxiaomo/pytorch-YOLOv4/blob/master/models.py#L333


class Yolov4Head(nn.Module):
    def __init__(self, output_ch, n_classes, inference=False):
        super().__init__()
        self.inference = inference
        self.conv1 = Conv_Bn_Activation(128, 256, 3, 1, 'leaky')
        self.conv2 = Conv_Bn_Activation(256, output_ch, 1, 1, 'linear', bn=False, bias=True)
        self.yolo1 = YoloLayer(
                                anchor_mask=[0, 1, 2], num_classes=n_classes,
                                anchors=[12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401],
                                num_anchors=9, stride=8)
        # R -4
        self.conv3 = Conv_Bn_Activation(128, 256, 3, 2, 'leaky')
        # R -1 -16
        self.conv4 = Conv_Bn_Activation(512, 256, 1, 1, 'leaky')
        self.conv5 = Conv_Bn_Activation(256, 512, 3, 1, 'leaky')
        self.conv6 = Conv_Bn_Activation(512, 256, 1, 1, 'leaky')
        self.conv7 = Conv_Bn_Activation(256, 512, 3, 1, 'leaky')

Thanks.

TomK · October 12, 2021, 4:15pm