mil-tokyo / dg_mmld

Pytorch Implementation of Domain Generalization Using a Mixture of Multiple Latent Domains

Domain Generalization Using a Mixture of Multiple Latent Domains

This is the pytorch implementation of the AAAI 2020 poster paper "Domain Generalization Using a Mixture of Multiple Latent Domains".

Requirements

• A Python install version 3.6
• A PyTorch and torchvision installation version 0.4.1 and 0.2.1, respectively. pytorch.org
• The caffe model we used for AlexNet
• PACS dataset (website, dateset)
• Install python requirements

pip install -r requirements.txt

Training and Testing

You can train the model using the following command.

cd script
bash general.sh

If you want to train the model without domain generalization (Deep All), you can also use the following command.

cd script
bash deepall.sh

You can set the correct parameter.

• --data-root: the dataset folder path
• --save-root: the folder path for saving the results
• --gpu: the gpu id to run experiments

Citation

If you use this code, please cite the following paper:

Toshihiko Matsuura and Tatsuya Harada. Domain Generalization Using a Mixture of Multiple Latent Domains. In AAAI, 2020.

@InProceedings{dg_mmld,
  title={Domain Generalization Using a Mixture of Multiple Latent Domains},
  author={Toshihiko Matsuura and Tatsuya Harada},
  booktitle={AAAI},
  year={2020},
  }

dg_mmld's Issues

Qustion about pseudo domain

I don't know what is the size of pseudo domain label.Where is the corrsponding code? Thank you

No such file or directory: '/data/unagi0/matsuura/model/alexnet_caffe.pth.tar'

_pickle.UnpicklingError: invalid load key, '<'

HI @milhidaka I am getting the following error
Traceback (most recent call last):
File "../main/main.py", line 90, in
num_classes=source_train.dataset.dataset.num_class, num_domains=disc_dim, pretrained=True)
File "../util/util.py", line 72, in get_network_fn
return nets_map[name]train
File "../model/caffenet.py", line 66, in caffenet
state_dict = torch.load("alexnet_caffe.pth.tar", map_location=lambda storage, loc: storage, pickle_module=pickle)
File "/home/vinodkk/miniconda2/envs/dg/lib/python3.6/site-packages/torch/serialization.py", line 358, in load
return _load(f, map_location, pickle_module)
File "/home/vinodkk/miniconda2/envs/dg/lib/python3.6/site-packages/torch/serialization.py", line 532, in _load
magic_number = pickle_module.load(f)
_pickle.UnpicklingError: invalid load key, '<'.
PACS/default/art1
path ../../../../../dataset/result/dg_mmld/PACS/default/art1

---

Source domain: photo, cartoon, sketch, Target domain: art_painting
Train: 7148, Val: 795, Test: 2048
args.model caffenet

Unable to extract tar file for caffe_alexnet model(using the link given).There seems to be some problem in file.

VLCS Dataset

Hi,

I wanted to try your method on the VLCS dataset, as described in your paper, but I am having a hard time finding the dataset. Can you point me to where you get the dataset?

Thanks!

PACS Dataset

Hi ! Thanks for sharing the work! But the link of PACS Dataset failed. Could you update the download link of PACS! Thank u

Question on the loss function formulation

Thanks for making your code public. I had a question about the loss function mentioned in the paper versus what has been implemented in the code:

In the paper you have eq.

However, in the implementation I find the following:

dg_mmld/train/general.py

Line 44 in bfd5942

total_loss = loss_class + loss_domain + loss_entropy * beta

I am trying to understand how to correlate the two. I would appreciate it if you could provide more details on how the implementation matches the equation in the paper.

Upload Images for README.md