Hi,

Thanks for your contribution and the publication of your code. I was wondering if it is possible to run your method in a different setup, for example, let's say that I want to adapt Cityscapes ===> BDD.

I have tried to run this setup by making some minor modifications but it seems like your method is conditioned to these id2json class-wise objects, which I do not know how I can generate those files for the aforementioned setup Cityscapes ===> BDD.

Could you please help me out on this?

Thanks for sharing the code, it's really an awesome work!

But there seems to exist a mistake in "Table 2: Results of different domain adaptation methods for the SYNTHIA → Cityscapes task".

The mIoU* of 13 classes should be 59.9 but not 63.8 according to the IoU per class, can you check it?

hi, I want to ask is there any difference between the pretrained models in the baidu cloud and the model in http://vllab1.ucmerced.edu/~whung/adv-semi-seg/resnet101COCO-41f33a49.pth. And could you do me a favor to offer the model in http://vllab1.ucmerced.edu/~whung/adv-semi-seg/resnet101COCO-41f33a49.pth? I can't connect to that and download it.

RuntimeError: CUDA out of memory. Tried to allocate 18.00 MiB (GPU 0; 23.70 GiB total capacity; 1.38 GiB already allocated; 7.00 MiB free; 1.42 GiB reserved in total by PyTorch)

I have two 3090 GPU. But it CUDA out of memory

Hello, first of all thanks for the contribution and sharing the code. I have been using this implementation for a while and considering the class "rand_mixer" for mixing the long tail classes with source/target inputs, I notice that the long tail classes tend to not show up in the mixed result. To ilustrate the issue, I add some screenshots of plots for each variable.

Considering the code, I believe that the long tail class mixing in line 325 in train.py:

   data, target = transformsgpu.oneMix(MixMask, data=mixdata, target=mixtarget)

I introduced plotting the data and targets before and after this function to see the long tail mixing in action with the following code:

      plot_input(MixMask, 'mask_lt', 'Mask over the long tail class image', save=True)
      plot_input(img, 'image_lt', "Image where the long tail class is in", save=True)
      plot_input(lbl, 'label_lt', 'Label of the longtail class image', save=True)

      data, target = transformsgpu.oneMix(MixMask, data=mixdata, target=mixtarget)

      plot_input(mixdata[1], 'image_lt', 'Image where the long tail class comes in', save=True)
      plot_input(data[0], 'image_lt', 'Mixed image', save=True)

      plot_input(mixtarget[1], 'label_lt', 'Label where the long tail class comes in', save=True)
      plot_input(target[0], 'label_lt', 'Mixed label', save=True)

The plot input functions that is used above, is shown here:

   def plot_input(input, input_type, name=None, save=False):
    with torch.no_grad():

        if input_type == 'image_lt':  # Image while in long tail function
            input = input.cpu().detach().numpy()

            input = np.transpose(input, (1, 2, 0))

            image = input + IMG_MEAN
            image = image[:, :, ::-1]
 
            plt.imshow(image.astype(int))


        elif input_type == 'mask_lt': # mask while in long tail function

            input = input.cpu().detach().numpy()
            plt.imshow(input, cmap='binary')

        elif input_type == 'label_lt': # label while in long tail function

            input = input.cpu().detach().numpy()
            color_mask = cm_eval(input)
            # print(color_mask)
            plt.imshow(color_mask)

        else:
            print("Incorrect input for creating an image")

        if not (name is None):
            title_obj = plt.title(name)  # get the title property handler
            plt.getp(title_obj)

        if save:
            start_writeable = datetime.datetime.now().strftime('%m-%d_%H-%M')
            plt.savefig(os.path.join("./saved/images", start_writeable+input_type+name), bbox_inches='tight')
        plt.show()

So when I plot this. The following figures are displayed:

• The Mask over the long tail image:

• The Image with the long tail class object in it:

• The Label with the long tail class object in it:

• The Image where the long tail class object comes into:

• The Label where the long tail class object comes into:

• The Image with the long tail class object mixed in:

• The Label with the long tail class object mixed in:

As you can see, the long tail class is not mixed into the image. This happens in every occasion, but in many situations the long tail class is really small, so not visible in the image. Above a clear example is shown.

Countering the argument above, I have had two times a situation, where the image with the long tail class objects was overwritten in the image and label partly. However, this was done in a small strip in the vertical direction. An additional requirement for this to occur, was that the long tail class object was at the (x,0) position. I have established to save one of those examples. These are shown below:

• The Mask over the long tail image:

• The Image with the long tail class object in it:

• The Label with the long tail class object in it:

• The Image where the long tail class object comes into:

• The Label where the long tail class object comes into:

• The Image with the long tail class object mixed in:

• The Label with the long tail class object mixed in:

Could someone explain this behavior to me? And if possible suggest a solution to this problem.

Thanks for your sharing code, it's really an awesome work! I have 2 questions as follows:

1. Does there exist the warmup stage of adversarial training?
2. I notice that is only used in DSP, considering that the process is in a soft manner, why is it not used as a weight to compute loss?

If I try to double the current batch size of 2 to 4, I get the following error:

torch.Size([2, 19, 512, 512]) torch.Size([4, 512, 512]) Traceback (most recent call last): File "train.py", line 819, in <module> main() File "train.py", line 660, in main L_l2 = loss_calc(logits_t_s, labels) * (1-lam) + lam * loss_calc(logits_t_s, targets_t) File "train.py", line 112, in loss_calc return criterion(pred, label) File "~/miniconda3/envs/dsp_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, **kwargs) File "~/projects/DSP/utils/loss.py", line 30, in forward predict = predict[target_mask.view(n, h, w, 1).repeat(1, 1, 1, c)].view(-1, c) I have printed the shapes and it seems like logits_t_s still has a batch size of 2. Any particular reason for this?

Hi, I try to reproduce your code (gta5 -> cityscapes) and I found that it has a very slow training speed, which is nearly 4x slower than the DACS. I wonder why... And also, can I ask you the total training time for you to reproduce the result (miou =55) ? Another issue is I found that the training curve of DACS in your paper is mismatch with my experiment result as below, I wonder how you get the result in your paper ? My experiment on DACS

Yours