should use .pt or .weight ?

[yenai3726]

HI i get some problem need help

in README.md test use .weight and train is use ' ' or .pt i can't understand should use which one !

first i try to train and i get .pt
and this my order: python train.py --device 0 --batch-size 4 --data plate.yaml --cfg yolov4-csp-plate.cfg --weights 'weights/yolov4.conv.137' --name yolov4-csp

when i try to use test.py , i get this error

Model Summary: 327 layers, 6.41154e+07 parameters, 6.41154e+07 gradients Traceback (most recent call last): File "test.py", line 70, in test ckpt['model'] = {k: v for k, v in ckpt['model'].items() if model.state_dict()[k].numel() == v.numel()} File "test.py", line 70, in ckpt['model'] = {k: v for k, v in ckpt['model'].items() if model.state_dict()[k].numel() == v.numel()} KeyError: 'module_list.3.Conv2d.weight' During handling of the above exception, another exception occurred: Traceback (most recent call last): File "test.py", line 297, in opt.verbose) File "test.py", line 73, in test load_darknet_weights(model, weights[0]) File "F:\PyTorch_YOLOv4-master\models\models.py", line 441, in load_darknetweights conv.weight.data.copy(torch.from_numpy(weights[ptr:ptr + nw]).view_as(conv.weight)) RuntimeError: shape '[1024, 512, 3, 3]' is invalid for input of size 3955080

i make sure our cfg already change classes and filters
how can i fix this error?

[jaqub-manuel]

hey @yenai3726 , you trained wrong way. instead of yolov4.conv.137 (yolov4 weights), you should use yolov4-csp weights, download link below, it will be fine.https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-csp.conv.142

[yenai3726]

hey @jaqub-manuel Thank you for your attention and answer

oh my mistake I forgot this part about weight and i want to know if i use different yolov4-version weight What effect will it have ?

[Duarte-Nunes]

Hi, im not very tech savvy when it comes to Ai, i'm still learning but from my understanding the weights are tied to the structure of the network, if you use a different weight file on a Yolo version that has a different structure it won't work.

In sum, the weights need to match the yolo version.

[yenai3726]

I got it thank you for your's answer