hello, we re-run the cbst_train with the default settings you provide, but get bad results as shown in the fig, even worse than the source only method. i wonder the stability of the training of cbst, and i will appreciate that if you can provide the training log of the cbst. THANK YOU VERY MUCH!

Why is the domain adaptation MIOU on the CodaLab site so high? Shouldn't the "Oracle" MIOU provided in the paper be the highest MIOU for this domain adaptation task?

The results submitted to the CodaLab get zero score and zero ExecutionTime. I wonder is it any wrong with the CodaLab or it is just my own mistake. The output class index is 0~6 with 1024*1024 pixels.

Hi, I am member of OpenMMLab who develops MMSegmentation. Our vision is provide up-to-date methods and dataset(i.e., benchmark) for researchers and community around the world.

First, congrats for acceptance of NeurIPS'21. I think this dataset and benchmark would definitely help Remote Sensing Image field where semantic segmentation plays an important role.

Frankly speaking, right now we do not have too much human resources. Would you like to help us incpoorate your dataset into MMSegmentation? We appreciate all contibutors and users, here is our contributing details.

I think if LoveDA is provided by MMSegmentation, it could let more people use & cite this excellent work, especially for those who want to establish standard segmentation benchmark.

Looking forward to your reply. Wish you all the best.

Best,

Following up on the discussion from #23, I was wondering whether in the context of the segmantic segmentation task there could be a shift in class labels between the data on which the pretrained model hrnetw32.pth was trained on and the data provided in this repo.

Here I have visualised the true and predicted segmentations on training image 1338 for 2 different COLOR_MAP-s from the repo (render.py and data.loveda.py)

Based on the input image we can see that the colours are correct for the top left and bottom right visualisations. Also, the black colours in top right image corresponds to label IGNORE with RGB values (0,0,0) while in the bottom left the black colour has RGB values (7,7,7), which seems to be because in data.loveda.py the COLOR_MAP only has 7 classes and with indexing 0-6 with agriculture having label 7 in the masked images, it is not colour mapped.

This seems to be related to the difference between labels in the current repo:

Category labels: background – 1, building – 2, road – 3, water – 4, barren – 5, forest – 6, agriculture – 7. And the no-data regions were assigned 0 which should be ignored. The provided data loader will help you construct your pipeline.

and the ones described on CodaLab:

Classes indexes: Background - 0, Building - 1, Road - 2, Water - 3, Barren - 4, Forest - 5, Agriculture - 6

Could this class label offset be the case or perhaps there is an alternative explanation which I have not thought about?

The links mentioned on the README.md of this repository as well as the competition page for google drive of the dataset are broken as of 30-01-2022 and return a 404 error. Please update the link with a working one.

I found that in the training process, the input resolution is 512x512, while in the test phase, the input resolution is 1024x1024. Would you please tell me why?

Hello,

Thank you very much for making your excellent work open to the public.

May I ask you the meaning of line 228 in tools.py for Unsupervised Domain Adaptation? I found that when running bash ./scripts/predict_cbst.sh, it will generate a bug saying AttributeError: 'NoneType' object has no attribute 'info'. This bug is due to line 228 and also the default setting _default_logger=None. Hence, I wonder what this line is for. Also, I would like to let you know that after commenting the line 228, the command can be run successfully.

Many thanks for your help.

Would it be possible to release the evaluation script from codalab? File format detail is a bit confusing. For example, if I set empty regions as transparent or embed color palette within the image the evaluation script shows warning:

/opt/conda/lib/python2.7/site-packages/PIL/Image.py:870: UserWarning: Palette images with Transparency   expressed in bytes should be converted to RGBA images
  'to RGBA images')

Even if i remove the color palette I get the following error:

Traceback (most recent call last):
  File "/tmp/codalab/tmpS_IrwU/run/program/evaluate.py", line 157, in <module>
    metric.forward(gt[valid_inds], mask[valid_inds])
  File "/tmp/codalab/tmpS_IrwU/run/program/evaluate.py", line 22, in forward
    cm = sparse.coo_matrix((v, (y_true, y_pred)), shape=(self.num_classes, self.num_classes), dtype=np.float32)
  File "/opt/conda/lib/python2.7/site-packages/scipy/sparse/coo.py", line 182, in __init__
    self._check()
  File "/opt/conda/lib/python2.7/site-packages/scipy/sparse/coo.py", line 219, in _check
    nnz = self.nnz
  File "/opt/conda/lib/python2.7/site-packages/scipy/sparse/coo.py", line 196, in getnnz
    raise ValueError('row, column, and data array must all be the '
ValueError: row, column, and data array must all be the same length

I made sure all my images are 1024 × 1024 with a single uint8 channel. The class ids have been assigned as per the specification, with empty regions assigned with value 15

Classes indexes

Background - 0
Building - 1
Road - 2
Water - 3
Barren - 4
Forest - 5
Agriculture - 6

So, it would be helpful to see the evaluation script and generate compatible prediction images.

Hi, I would like to use the visualized results of Adaptseg and CLAN for comparison, could you provide the pre-training weights (Rural to Urban weights) of these two networks?

Hi,

Is there a way to run ./scripts/predict_test.sh without CUDA?

I am using the LoveDA dataset and pretrained model weights hrnetw32.pth as described in the ReadME.

Initially I got the error urllib.error.HTTPError: HTTP Error 403: Forbidden, which I fixed by setting pretrained=False as recommended here: https://github.com/Junjue-Wang/LoveDA/issues/9.

Then when rerunning the predict_test.sh, I got the error:

Traceback (most recent call last):
  File "predict.py", line 52, in <module>
    predict_test(args.ckpt_path, args.config_path, args.out_dir)
  File "predict.py", line 38, in predict_test
    model.cuda()
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/nn/modules/module.py", line 680, in cuda
    return self._apply(lambda t: t.cuda(device))
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/nn/modules/module.py", line 570, in _apply
    module._apply(fn)
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/nn/modules/module.py", line 570, in _apply
    module._apply(fn)
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/nn/modules/module.py", line 570, in _apply
    module._apply(fn)
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/nn/modules/module.py", line 593, in _apply
    param_applied = fn(param)
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/nn/modules/module.py", line 680, in <lambda>
    return self._apply(lambda t: t.cuda(device))
  File "/Users/kristjan/miniconda3/envs/mip/lib/python3.7/site-packages/torch/cuda/__init__.py", line 208, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

I then commented out the line 38: https://github.com/Junjue-Wang/LoveDA/blob/4d574ce08f84cbc8d27becf2bd9dce8fbb7f50f8/Semantic_Segmentation/predict.py#L38 and after rerunning predict_test.sh, I got the output:

Load model!
INFO:data.loveda:./LoveDA/Val/Urban/images_png -- Dataset images: 0
INFO:data.loveda:./LoveDA/Val/Rural/images_png -- Dataset images: 0
INFO:ever.core.logger:HRNetEncoder: pretrained = False
0it [00:00, ?it/s]

Traceback (most recent call last). File ""home/libowen/LoveDA-master/Semantic_Segmentation/predict.py", line 52, in smodule> predict test(argsckpt path, args.config path, args.out dir) File "/home/libowen/LoveDA-master/Semantic_Segmentation/predictpy", line 37, in predict test model.load_state_dictmodel_state_dict) File "home/libowen/.conda/envs/bw/ib/python3.8/site-packages/torch/nn/modules/module,py", line 1667, in load_state_dictraise RuntimeError('Error(s) in loading state_dic for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state dict for HRNetFusion. Missing keys) in state dict: "ackbone.het.conv1.weight"ackbone hmet bn1.weight, "ackbone hretbn1.bias""backbone hmet bn1.running mean"

I download pretrained weight and use predict.py to test some images, but meet this bug, what's the problem of the fuse_layers?

File "test4/Road/LoveDA-master/Semantic_Segmentation/module/baseline/base_hrnet/_hrnet.py", line 394, in forward y = y + self.fuse_layers[i][j](x[j]) RuntimeError: The size of tensor a (500) must match the size of tensor b (504) at non-singleton dimension 3

**Shall we run this work with One GPU? If possible how to set parameters？ **

I'v got the issue below:

PS F:\Models\LoveDA-master\Semantic_Segmentation> bash ./scripts/train_hrnetw32.sh NOTE: Redirects are currently not supported in Windows or MacOs. Init Trainer Set Seed Torch Traceback (most recent call last): File "train.py", line 79, in trainer = er.trainer.get_trainer('th_amp_ddp')() File "D:\ProgramData\Anaconda3\lib\site-packages\ever\api\trainer\th_amp_ddp_trainer.py", line 77, in init torch.cuda.set_device(self.args.local_rank) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\cuda_init_.py", line 311, in set_device device = _get_device_index(device) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\cuda_utils.py", line 34, in _get_device_index return _torch_get_device_index(device, optional, allow_cpu) File "D:\ProgramData\Anaconda3\lib\site-packages\torch_utils.py", line 537, in _get_device_index 'or an integer, but got:{}'.format(device)) ValueError: Expected a torch.device with a specified index or an integer, but got:None ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 108252) of binary: D:\ProgramData\Anaconda3\python.exe Traceback (most recent call last): File "D:\ProgramData\Anaconda3\lib\runpy.py", line 193, in run_module_as_main "main", mod_spec) File "D:\ProgramData\Anaconda3\lib\runpy.py", line 85, in run_code exec(code, run_globals) File "D:\ProgramData\Anaconda3\Scripts\torchrun.exe_main.py", line 7, in File "D:\ProgramData\Anaconda3\lib\site-packages\torch\distributed\elastic\multiprocessing\errors_init.py", line 345, in wrapper return f(*args, **kwargs) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\distributed\run.py", line 724, in main run(args) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\distributed\run.py", line 718, in run )(*cmd_args) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\distributed\launcher\api.py", line 131, in call return launch_agent(self._config, self._entrypoint, list(args)) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\distributed\launcher\api.py", line 247, in launch_agent failures=result.failures, torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

train.py FAILED

Failures: <NO_OTHER_FAILURES>

Root Cause (first observed failure): [0]: time : 2022-11-13_13:10:33 host : KWPAACQRFTY8V05 rank : 0 (local_rank: 0) exitcode : 1 (pid: 108252) error_file: <N/A> traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html

When training self-training 2urban scripts, such as CBST_train.py and IAST_train.py, there is a problem which is 'FileNotFoundError: No such file: '/home/xxx/ssuda/UDA/log/cbst/2urban/pseudo_label/3814.png''. I guess this is because that the batch size is set to 2, and as expected the problem is solved when batch size is modified to 1.

So, I wonder that if this is a bug or something what?

Thanks for your excellent works!