leokarlin / brad Goto Github PK

Thank you for your impressive work,

I am a new researcher in the field of unsupervised learning. I have a question about the setting of the evaluation part. I notice that the classifier is fit with source labels. Does this break the initial purpose of unsupervised learning?

Look forward to your reply.
Regards,

Hi, can you release the Resnet18 pretrained weights on Pacs dataset for UDG task?
Thanks!

Hi! I am very appreciated of your work and it really inspires me. I am trying to reproduce the results in Table 3 of your paper. However, I couldn't find the training script and parameters in your official README. Would you please share the FUDA training parameters and scripts so that I could reproduce your results?

Thank you!

i know that FSDA setting need an imagenet pretrained resnet to init
but if the UDG setting need an imagenet pretrained resnet to init or random init?

Thank you for your great work!
It seems that both main_brad.py and main_brad_test.py are code for the UDA setting training and testing.
Could you please release the training and testing code for the UDG setting?
Thanks a lot for your effort.
@EliSchwartz @leokarlin @aarbelle @sivangl

Hi! Sorry for bothering you again. I am trying to resume training from the checkpoints you released at https://drive.google.com/file/d/1T7v2xwAWQGsAv11-CEwKLmUH-TmCkue9/view?usp=sharing by using --resume parameters when running main_brad.py. However, I got the following errors:

File "main_brad.py", line 179, in main_worker
missing_keys, unexpected_keys = model.load_state_dict(sd, strict=False)
File "/home/xxxxx/miniconda3/envs/brad/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1223, in load_state_dict
Traceback (most recent call last):
File "main_brad.py", line 649, in
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for DistributedDataParallel:
size mismatch for module.queue: copying a param with shape torch.Size([128, 65536]) from checkpoint, the shape in current model is torch.Size([128, 65088]).
size mismatch for module.queue_idx: copying a param with shape torch.Size([65536]) from checkpoint, the shape in current model is torch.Size([65088]).

I was using the default parameters in config.py and set --nproc_per_node=4. It seems that the model parameters ("65088" in "the shape in current model is torch.Size([65088])") vary as the batch size changes. I wonder what is the reason of this and I would sincerely appreciate it if you would provide the proper way to resume training the released weights.

Thank you!

Could anybody reproduce the result in Table 1 with the pretrained weights provided here? I modified the main_brad_test.py but the acc is only ~17% rather than 60.78%.

Hi,

I would appreciate it if you could kindly let me know from where the original dataset (.jpg) was downloaded?

Thanks.
Sara A. Al-Emadi