Reproduce results

[liheyoung]

Reproduce results

[LiheYoung]

The ignore_mask is to mark:
(1) ignored region provided in Pascal GT mask,
(2) padded region during pre-processing.

The mask == 254 is indeed our padded region (see here). It does not use value 255 for unlabeled images in the first place because we want to tell it from the officially provided 255 values in GT masks, which we should not leverage on unlabeled images.

[LiheYoung]

• Previously, we have re-run our Pascal experiments three times under each data partition. Below are the results and corresponding training logs. Not that, all results are obtained with ResNet-101 and training size 321. The reported results in our paper are listed in the last row for clear comparison. I think it is easy to reproduce them. Any PyTorch version >= 1.8 should be okay.

Pascal	92	183	366	732	1464	1/16	1/8	1/4
Run 1	75.46	76.75	78.51	79.98	80.78	76.91	77.46	77.54
Run 2	74.78	78.17	78.69	80.21	80.71	76.98	77.74	77.21
Run 3	74.74	77.69	78.42	79.48	81.01	77.15	77.05	77.44
Mean (std)	74.99 (0.40)	77.54 (0.72)	78.54 (0.14)	79.89 (0.37)	80.83 (0.16)	77.01 (0.12)	77.42 (0.35)	77.40 (0.17)
Paper	75.2	77.2	78.8	79.9	81.2	76.5	77.0	77.2

• Indeed, we have clarified the online hard example mining loss (OHEM) in the Cityscapes paragraph of Section 4.2 in our main paper. To use it, you can follow the Cityscapes config file.

[LiheYoung]

I will close this issue. Feel free to re-open it if you have any further questions.

[zhibotian]

Hi, as your method in the paper, you also do experiments to use Dual-Stream Feature-Level Perturbations, can you tell me the details about it, using two different weak perturbations feeding to the network to get different Feature-Level perturbations results or just use one weak perturbations but return two after the encoder g. Thanks.

[LiheYoung]

It is the latter one as you described. We only use a single weak view, but perform twice random feature-level perturbations (Dropout) on it.

Briefly, the code should be changed from:

UniMatch/model/semseg/deeplabv3plus.py

Lines 45 to 51 in 7292bac

	if need_fp:
	outs = self._decode(torch.cat((c1, nn.Dropout2d(0.5)(c1))),
	torch.cat((c4, nn.Dropout2d(0.5)(c4))))
	outs = F.interpolate(outs, size=(h, w), mode="bilinear", align_corners=True)
	out, out_fp = outs.chunk(2)
	return out, out_fp

to:

if need_fp:
    outs = self._decode(torch.cat((c1, nn.Dropout2d(0.5)(c1), nn.Dropout2d(0.5)(c1))),
                        torch.cat((c4, nn.Dropout2d(0.5)(c4), nn.Dropout2d(0.5)(c4))))
    outs = F.interpolate(outs, size=(h, w), mode="bilinear", align_corners=True)
    out, out_fp1, out_fp2 = outs.chunk(3)

    return out, out_fp1, out_fp2

Besides, the original loss function loss_u_w_fp should also be averaged on the dual views of out_fp1 and out_fp2.

[zhibotian]

Okay, get it, thank you very much for your patient reply.

[zhibotian]

Hi, Sorry to bother you again, Can you tell me the purpose of this code ignore_mask[mask == 254] = 255 and the role of the variable ignore_mask in the code in semi.py, thanks a lot.

[zhibotian]

Hello, I noticed that the findings of Resnet101 spilt 1/6, 1/8, 1/4, for U2PL in your work deviate from the original paper; have you recreated this result? (The first image is the outcome of your paper.)

[LiheYoung]

Please refer to Haochen-Wang409/U2PL#3 for detail.

[zhibotian]

Soga! Thanks.