Is it trained after loading the pre training model or not about yolov4-pytorch HOT 16 OPEN

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

from yolov4-pytorch.

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

from yolov4-pytorch.

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.

from yolov4-pytorch.

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.

from yolov4-pytorch.

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.

from yolov4-pytorch.

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**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???

from yolov4-pytorch.

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????

from yolov4-pytorch.

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**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。

from yolov4-pytorch.

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。

from yolov4-pytorch.

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

from yolov4-pytorch.

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。

from yolov4-pytorch.

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%

from yolov4-pytorch.

不过你的数据处理部分是没有mosaic的

from yolov4-pytorch.

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.

from yolov4-pytorch.

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!

from yolov4-pytorch.

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。。。

from yolov4-pytorch.