Did anyone have done inference with custom anchors?

[moromatt]

Hi guys, I'm using YOLOv5 3.0 and I've trained a yolov5L model generating 7 anchors for my custom dataset composed by 3x800x800 pixel images. The autoanchors function returns the array of generated anchors that I put in my yaml file, and they look like this:

# anchors
anchors:
- [23,24,  27,36,  38,29,  28,70,  53,49,  94,32,  187,26]
- [37,148,  72,95,  128,64,  38,286,  118,137,  403,76,  130,241]
- [217,159,  218,284,  191,548,  653,189,  391,316,  193,697,  697,717]

When I do some inference in Pytorch everything is ok.

Then I convert my model and the yaml to cfg file, and my cfg file look like this:

[net]
width=800
height=800
channels=3
batch=1
nc=6
depth_multiple=1.0
width_multiple=1.0
anchors=23, 24, 27, 34, 46, 32, 229, 66, 97, 29, 34, 146, 75, 88, 193, 40, 64, 210, 121, 129, 233, 161, 183, 248, 133, 556, 407, 234, 177, 765, 785, 269, 508, 599, 784, 813

Now I would like to use those anchors in this c++ solution, but for now I didn't have any success. I keep getting this error:

File does not exist : ../configs/yolov5-3.0/yolov5l_custom-kFLOAT-batch1.engine
Loading pre-trained weights...
Loading complete!
      layer               inp_size            out_size
(1)   Focus             3 x 800 x 800      64 x 400 x 400
(2)   Conv             64 x 400 x 400     128 x 200 x 200
(3)   BottleneckCSP   128 x 200 x 200     128 x 200 x 200
(4)   Conv            128 x 200 x 200     256 x 100 x 100
(5)   BottleneckCSP   256 x 100 x 100     256 x 100 x 100
(6)   Conv            256 x 100 x 100     512 x  50 x  50
(7)   BottleneckCSP   512 x  50 x  50     512 x  50 x  50
(8)   Conv            512 x  50 x  50    1024 x  25 x  25
(9)   SPP            1024 x  25 x  25    1024 x  25 x  25
(10)  BottleneckCSP  1024 x  25 x  25    1024 x  25 x  25
(11)  Conv           1024 x  25 x  25     512 x  25 x  25
(12)  Upsample        512 x  25 x  25     512 x  50 x  50
(13)  Concat          512 x  50 x  50    1024 x  50 x  50
(14)  BottleneckCSP  1024 x  50 x  50     512 x  50 x  50
(15)  Conv            512 x  50 x  50     256 x  50 x  50
(16)  Upsample        256 x  50 x  50     256 x 100 x 100
(17)  Concat          256 x 100 x 100     512 x 100 x 100
(18)  BottleneckCSP   512 x 100 x 100     256 x 100 x 100
(19)  Conv            256 x 100 x 100     256 x  50 x  50
(20)  Concat          256 x  50 x  50     512 x  50 x  50
(21)  BottleneckCSP   512 x  50 x  50     512 x  50 x  50
(22)  Conv            512 x  50 x  50     512 x  25 x  25
(23)  Concat          512 x  25 x  25    1024 x  25 x  25
(24)  BottleneckCSP  1024 x  25 x  25    1024 x  25 x  25
Assertion failed: size == (map_wts_[s_layer_name_ + ".weight"].size()), file .\yolo_tensorrt\modules\trt_utils.cpp, line 854

The layer that seems to not have the right size is s_layer_name = "model.24.m.0" = 19712 and int size = n_filters_ * chw[0] * n_kernel_size_ * n_kernel_size; = 8448 So for some reason size is exactly the half of s_layer_name.

Can someone help me out with this? I'm using TensorRT 7.1.3.4, OpenCV 4.4.0. Thanks in advance!