mil-tokyo/mcd_da

Maximum Classifier Discrepancy for Domain Adaptation

This is the implementation of Maximum Classifier Discrepancy for digits classification and semantic segmentation in Pytorch. The code is written by Kuniaki Saito. The work was accepted by CVPR 2018 Oral.

Maximum Classifier Discrepancy for Domain Adaptation: [Project][Paper (arxiv)].

Getting Started

Go to classification or segmentation folder and see the instruction for each task.

Citation

If you use this code for your research, please cite our papers (This will be updated when cvpr paper is publicized).

@article{saito2017maximum,
  title={Maximum Classifier Discrepancy for Unsupervised Domain Adaptation},
  author={Saito, Kuniaki and Watanabe, Kohei and Ushiku, Yoshitaka and Harada, Tatsuya},
  journal={arXiv preprint arXiv:1712.02560},
  year={2017}
}

mcd_da's Issues

sharing svhn2mnist result

hello, can you share your result of svhn2mnist accuracy?

I set svhn as the source and mnist as the target. Is it normal that the target accuracy is higher?

Reproduce the segmentation result

I have tried to rerun the segmentation part of the code in python 2.7
and I was unable to reproduce the result in the paper.
I have noticed several differences.

1. I have rewritten the label map part in the GTA dataset since GTA needs to map to Cityscape and the original code didn't do it.
2. I was using torch 1.3.1
3. I update the loss to NLLLoss rather than NLLLoss2d
4. I fixed a few grammar bugs which probably happened because of the change of time in the code
   here is my result:

I would love to see if anyone was able to reproduce the segmentation result, and I would be really appreciated if you could give me that piece of code such that I could see where went wrong.

THANKS, A LOT!

About toy dataset

Where can I find the code of the toy dataset, which is figure 4 in the paper ??

When do segmentation task, you also need labels for target domain.

It is an unsupervised domain adaptation method, however, you also need labels for target domain in the segmentation task. I am a little confused here.

Replicating the results

Does anybody succeded in getting the results as the paper claimed?
So far by running the VisDA experiment with ResNet101 using default settings I get this:

Mean class accuracy 62.54% (v.s 71.9 reported in paper)

About datasets on classification

drn_d_105-12b40979.pth

(pytorch1.1) zgm@zgm-icv:~/Lufei/MCD_DA/segmentation$ python adapt_trainer.py gta city --net drn_d_105 Downloading: "https://tigress-web.princeton.edu/~fy/drn/models/drn_d_105-12b40979.pth" to /home/zgm/.cache/torch/checkpoints/drn_d_105-12b40979.pth Traceback (most recent call last): File "adapt_trainer.py", line 69, in <module> is_data_parallel=args.is_data_parallel) File "/home/zgm/Lufei/MCD_DA/segmentation/models/model_util.py", line 51, in get_models model_list = get_MCD_model_list() File "/home/zgm/Lufei/MCD_DA/segmentation/models/model_util.py", line 41, in get_MCD_model_list model_g = DRNSegBase(model_name=net_name, n_class=n_class, input_ch=input_ch) File "/home/zgm/Lufei/MCD_DA/segmentation/models/dilated_fcn.py", line 109, in __init__ pretrained=pretrained, num_classes=1000, input_ch=input_ch) File "/home/zgm/Lufei/MCD_DA/segmentation/models/drn.py", line 343, in drn_d_105 model.load_state_dict(model_zoo.load_url(model_urls['drn-d-105'])) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/site-packages/torch/hub.py", line 439, in load_state_dict_from_url _download_url_to_file(url, cached_file, hash_prefix, progress=progress) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/site-packages/torch/hub.py", line 354, in _download_url_to_file u = urllib.request.urlopen(req).read() File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/urllib/request.py", line 223, in urlopen return opener.open(url, data, timeout) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/urllib/request.py", line 532, in open response = meth(req, response) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/urllib/request.py", line 642, in http_response 'http', request, response, code, msg, hdrs) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/urllib/request.py", line 570, in error return self._call_chain(*args) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/urllib/request.py", line 504, in _call_chain result = func(*args) File "/home/zgm/.conda/envs/pytorch1.1/lib/python3.6/urllib/request.py", line 650, in http_error_default raise HTTPError(req.full_url, code, msg, hdrs, fp) urllib.error.HTTPError: HTTP Error 403: Forbidden
So can you provide the drn_d_105-12b40979.pth, thanks so much！

question for location

I don't know why it will result in this error。

Segmentation labels

Thank you for sharing the code! I am trying to run the segmentation task, and I am really confused about the segmentation labels.

What code do you run to transfer the original colored ground truth images of GTAV into the grayscale images with the 20 classes that you report on?
The same question about Cityscrapes - the dataset has 33 classes. How do you turn them into 20?

cityscapes/info.json

Thank for your code ! I am trying to run the segmentation task. But there is error :
File "eval.py", line 223, in eval_city
with open(join(devkit_dir, 'data', dset, 'info.json'), 'r') as fp:
IOError: [Errno 2] No such file or directory: '/data/ugui0/dataset/adaptation/taskcv-2017-public/segmentation/data/cityscapes/info.json'
can you help me ?

Model Selection

The classification code seems to be running "test" on every epoch and is printing the test accuracy.
Which of these accuracies do you report?
How is the model selection done?

I understand each experiment is repeated 5 times. But each time, is the last epoch accuracy considered for the mean or the max(accuracy at each epoch)?

Loss is becoming negative

Very nice work! During my training, I found that loss can become negative:

Train Epoch: 198 [0/100 (0%)]	Loss1: 0.024868	 Loss2: 0.022132	  Discrepancy: 0.018226

Test set: Average loss: -0.0588, Accuracy C1: 9449/10000 (94%) Accuracy C2: 9509/10000 (95%) Accuracy Ensemble: 9554/10000 (96%) 

recording record/usps_mnist_k_4_alluse_no_onestep_False_1_test.txt
Train Epoch: 199 [0/100 (0%)]	Loss1: 0.012343	 Loss2: 0.020431	  Discrepancy: 0.030520

Test set: Average loss: -0.0581, Accuracy C1: 9419/10000 (94%) Accuracy C2: 9518/10000 (95%) Accuracy Ensemble: 9537/10000 (95%) 

recording record/usps_mnist_k_4_alluse_no_onestep_False_1_test.txt

Do you think this is normal?

learning rate in segmentation

it seems that when I use lr=1e-3, the validation mIoU is very low,such as 4%.

Any suggestions?

Asking for the visualization code

Congratulations, you have done a great job.

am confused that whether you can publish the codes about visualization in figure 4? As hown bellow:

Thank you very much. Looking forward to your replay.

Question about the handwritten digit experiment

In Table 1 the paper shows very high scores (more than 90%) in the digit classification tasks. So the question is that the testing set is merely the target-domain testing set or the combination of both testing sets in source and target domains. For example, in the unsupervised domain adaptation task of SVHN--> MNIST, the classification score can reach as high as 96.2%. Is the testing set for getting 96.2% ONLY MNIST TEST SET or the combination of SVHN TEST SET and MNIST TEST SET? Thanks in advance.

Could you release the code on the VisDA dataset?

Thank you for sharing your code. Could you release the code on the VisDA dataset? I find it somehow difficult to re-implement the accuracy? Thanks.

synth traffic dataset

Hi, could you share the files of synth traffic dataset or where I can download it? I can't find this dataset in the Internet. Thanks.

About the classifiers.

Thanks for the code !

I have a question about the classifier. What is the differences between F1 and F2 classifier?
I can't find any differences in the code of classification folder.
Can you tell me the differences on code implementation ?

The misnt_data.mat dataset

could you give me the dataset in your paper named "misnt_data.mat"? I can't find it in the internet.

gta train val files

Just wondering if you could provide the additional train/val . txt files for gta.
The code needs to read "train.txt" and "val.txt" in

MCD_DA/segmentation/datasets.py

Line 244 in 37cb1bc

"gta": "", ## Fill the directory over images folder. put train.txt, val.txt in this folder

Are they the ids stored in the split.mat@gta and saved externally as .txt?

Code for " Strong-Weak Distribution Alignment for Adaptive Object Detection "

So many great papers! Thank you!
I've found your CVPR 2019 paper Strong-Weak Distribution Alignment for Adaptive Object Detection which achieves great performance on adaptive object detection. You said the code is coming soon, but I can't wait any longer. :) Hope sooner, thanks. :)

The difference between trainer and trainer_one_step

Hello, would you please explain the pros and cons between two types of implementations?

Did they achieve similar performance?

Why did you comment loss_dis.backward() @ classification > train_onestep function?

actually, i am not familiar with torch, but i think loss_dis does not applied if you comment
loss_dis.backward().

could you please explain about it?

thank you in advance.

Missing Loss_weight file

For Segmentation task: It seems like that there is a class weight file but I don't find this file. I am wondering that if you can help me or does I have something wrong?

USPS->Mnist source-only result can't reach 0.634

The paper said that the source-only accuracy of USPS->Mnist is 0.634. But I can only get the accuracy of nearly 0.2~0.3 in USPS->Mnist by using the same network. Does anyone get the same source-only result with me? How to solve it?

[Request] fcnvgg

I founded that there is no vgg_fcn file which is defined in models/model_util.py.
I would really appreciate it if you could share the network for FCN8s, since I'm not really sure about how to split the base FCN8s network into SegBase and SegPixelClassifier.
Thank you in advance.

The link (http://crcv.ucf.edu/data/adaptationseg/ICCV_dataset.zip) is lost.

Hi, this link http://crcv.ucf.edu/data/adaptationseg/ICCV_dataset.zip is lost. Do you know where to download this dataset instead?

what does ReLabel mean?

ReLabel(255, args.n_class - 1),  # Last Class is "Void" or "Background" class

actually, the args.n_class - 1 means 'bicyle' in cityscapes.

so you will map all unsure pixels to bicycle?

Some problems with building my own datasets on classificaion

Thank you very much for the code you provided.Recently, I have been studied the dataset you provided for classification. Mnist serves as the source domain data and usps serves as the target domain data.However, when I build my own data to replace mnist and usps data, there is always an error. I hope to get your help.Error is shown as follows:

   IndexError: index 149 is out of bounds for axis  0 with  size 1