bit-da / sepico Goto Github PK

Since we don't have a V100, how many hours did SePiCo use deelabv2 to train GTA5→Cityscapes on the V100 to reach the best score of 61mIoU?

Hi, i am intresting your paper very much! When release the code?
Looking forward very much......
Thanks

Hi,

Thanks for your impressive work. I am interested in reproducing the results you reported in paper. I tried Synthia to Cityscape dataset and adopts DistCL based on DAFormer, but my final results is 3 mIoU lower than your results, which is 64.3mIoU. I would like to ask if the parameters you used are the same as DAFormer? Or could you please upload a copy of your parameters used?

Thanks

Hi，could you tell me whether the final performance of the model or the highest performance in training is reported in the paper？Such as 61.0(gtav to cityscapes).

Hi, @BinhuiXie

Well done!

Hope to release soon.

Hi!

Thanks for your great work! The code is easy to follow.

I got a question when I try to reproduce the result.
I try to run the experiment based on GTAV → Cityscapes with DeepLab-v2 backbone, I got 59.1 mIoU following the same configs, which is lower than 61.0 mIoU. Could you please give any clue to fix this issue?

使用多卡的时候，是不是每张显卡都会有一个ProtoEstimator

When will you publish the code?
Thanks,
Shahaf

您好！
请问对应得代码什么时候会公开？
谢谢！

Hi, I would like to replicate the results represented in this paper for GTAV → Cityscapes (DAFormer based) experiment. I set the Experiment ID to 4, but I still could not get the same or close results. The results that I am getting is around 64 percent while it has been reported 70 percent in the paper. Could you please let me know if I should do any thing other than ID setting to replicate the results?

Hi,
Thank you for your wonderful work! I am more interested in the Cityscapes-->DarkZurich adaptation. However, in your run_experiments.py, there is no corresponding code implementation. Therefore, I build my own implementation of the Cityscapes-->DarkZurich code, but the results are poor, with only 28.69 MIoU on the validation set. Here is my training file, can you point out my mistakes. Thank you very much！
20230203_115623.log

Thanks for your great work, but a minor issue.
Are the results reported in your paper the average of the effects of multiple seeds? If not, is there an average for experiment 4 (gta->cityscapes and synthia->cityscapes based on SegFormer) that I can refer to? This result is preferably a fair comparison with DAFormer, that is, the test number is 1/4 of the number of times you publish code and the training resolution is 512x512. I hope to cite this result in my paper for comparison.

To compute proto_reg:
contrast_norm = num_classes * np.log(num_classes)
proto_sim = mean_feat.mm(mean.permute(1, 0).contiguous()) / contrast_temp
loss = torch.sum(torch.softmax(proto_sim, dim=1).log()) / contrast_norm

reg_weight * proto_reg(feat, mean, contrast_temp, contrast_norm=contrast_norm)

Why loss need < 0?
I don't understand it, Thanks!

Thank you for your great work, I'm really interested in it.

Due to my poor coding ability, there are some questions about the framework.
In your config file, we can know that you define uda part as a segmentor,

    def __getitem__(self, idx):
        if self.rcs_enabled:
            return self.get_rare_class_sample()
        else:
            s1 = self.source[idx // len(self.target)]
            s2 = self.target[idx % len(self.target)]
            return {
                **s1, 'target_img_metas': s2['img_metas'],
                'target_img': s2['img']
            }

为什么源域和目标域的索引，一个是//一个是%？

Hi,

Thanks for your great work! I would like to know if you could provide a checkpoint file of the synthia-cityscape dataset, since I can not reproduce the results in TABLE 4 (64.3) in the paper myself.

Thanks!

Hi,
thanks for your great work.

Table 9 states, that applying the contrastive loss to the deepest layer (4th layer) brings the largest improvement. However, applying it to layer 1 usually helped more than applying it to layer 3. That is a very interesting observation.

Further, you presented a t-SNE plot for the target domain, and I guess that the different colors refer to different classes. Did you also compute a t-SNE plot to examine whether the domains are sufficiently mixed? (I can't find a supplementary to the paper.)

Perhaps a t-SNE analysis could guide us to decide where to apply the contrastive loss? Perhaps it also depends on the progress of the training?

Best,
Marc

dear author：
如何安装mmcv包，用pip显示‘error: metadata-generation-failed’

contrastive_loss.py line:114，The mean parameter was not passed in。what is the mean parameter?

Hi, thanks for your great work of using contrastive learning to bridge the domain shift between source and target domains!

However, I am a littble bit confused of the scale of input images when training and testing.
In DAFormer, the input scale of training image is 512x512, and 1024x512 for testing, but in your setting, the scales are 640x640 and 1280x640 respectively.
Did you train the model with 512x512 and evaluate with 1024x512?
As ablated in SegFormer, training with 640x640 will be about 0.5 mIoU better than 512x512.

By the way, which one is the common practice in UDA? 640x640 or 512x512?

Looking forward to your reply.

请问老师，代码什么时候可以公开？