It is better for you to conduct your training program after the pre training model is loaded, which will hugely improve the validation accuracy.

It is better for you to conduct your training program after the pre training model is loaded, which will hugely improve the validation accuracy.

I understand what you said. I mean, under what circumstances did your model train? Did you train with or without the author's weight

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

I don't think that if you want to verify the performance of your reproduced code, you should not load the original author's weight during training
In particular, I expect you to validate your code and feed back your test results without loading the original author weight file.
Thanks.

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

I don't think that if you want to verify the performance of your reproduced code, you should not load the original author's weight during training
In particular, I expect you to validate your code and feed back your test results without loading the original author weight file.
Thanks.

yeah,you are excellent. I have changed my source code to verify the results. In this section, I have already loaded the CSPDarknet-53 weight file in the feature extraction network.

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

I don't think that if you want to verify the performance of your reproduced code, you should not load the original author's weight during training
In particular, I expect you to validate your code and feed back your test results without loading the original author weight file.
Thanks.

yeah,you are excellent. I have changed my source code to verify the results. In this section, I have already loaded the CSPDarknet-53 weight file in the feature extraction network.

哈哈,才知道你是中国人, 咨询下在不加载预训练模型情况下进行训练时,你的代码训练的效果是多少???

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

I don't think that if you want to verify the performance of your reproduced code, you should not load the original author's weight during training
In particular, I expect you to validate your code and feed back your test results without loading the original author weight file.
Thanks.

yeah,you are excellent. I have changed my source code to verify the results. In this section, I have already loaded the CSPDarknet-53 weight file in the feature extraction network.

你这个是最新的代码吗????

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

I don't think that if you want to verify the performance of your reproduced code, you should not load the original author's weight during training
In particular, I expect you to validate your code and feed back your test results without loading the original author weight file.
Thanks.

yeah,you are excellent. I have changed my source code to verify the results. In this section, I have already loaded the CSPDarknet-53 weight file in the feature extraction network.

哈哈,才知道你是中国人, 咨询下在不加载预训练模型情况下进行训练时,你的代码训练的效果是多少???

这个没有测试过，但结果应该不会很差。

yeah...I had used the author's weight (yolov4.weights) to get my results.For you, you can follow the instructions in read.me to conduct your experiment.

I don't think that if you want to verify the performance of your reproduced code, you should not load the original author's weight during training
In particular, I expect you to validate your code and feed back your test results without loading the original author weight file.
Thanks.

yeah,you are excellent. I have changed my source code to verify the results. In this section, I have already loaded the CSPDarknet-53 weight file in the feature extraction network.

你这个是最新的代码吗????

这个暂时是最新的，后面还有新的模块后续还会更新。

没有测试过，但结果应该不

这个我进行过测试, 训练不出来模型, 所以就请教你下

没有测试过，但结果应该不

这个我进行过测试, 训练不出来模型, 所以就请教你下

是用我的代码没有训练出模型？不加载预训练模型的话，还是可以训练的，不过训练会比较慢，很难拟合大型数据集并且精度会有所降低。

我从新训练下,不加载模型在VOC2007上我训练了51个epoch,map能达到64%

不过你的数据处理部分是没有mosaic的

Hi @argusswift ,
I think there is a bug in the forward function of CSPStage and CSPFirstStage:

        x0 = self.split_conv0(x)
        x1 = self.split_conv1(x)

        x1 = self.blocks_conv(x1)

        x = torch.cat([x0, x1], dim=1) # where [x1, x0] should be used as it is in the original implementation
        x = self.concat_conv(x)

Thanks.

Hi @argusswift ,
I think there is a bug in the forward function of CSPStage and CSPFirstStage:

        x0 = self.split_conv0(x)
        x1 = self.split_conv1(x)

        x1 = self.blocks_conv(x1)

        x = torch.cat([x0, x1], dim=1) # where [x1, x0] should be used as it is in the original implementation
        x = self.concat_conv(x)

Thanks.

Thanks, your are right.If possible, you can pull requests with the modified code.Thank you again!

这个提供的YOLOV4 DARKNET PRE-TRAINED WEIGHT是在COCO上训练的吗？为什么上来TOTAL LOSS非常的大。。。