luodian/madan

Pytorch Code release for our NeurIPS paper "Multi-source Domain Adaptation for Semantic Segmentation"

madan's Issues

where can I find MADAN pretrained model cyclegan_gta_synthia2cityscapes?

Hi,
I'm trying to run MADAN script "cyclegan_gta_synthia2cityscapes.sh", and I can't find the pretrained model:
pretrained_models/cyclesynthia_drn26_iter2000.pth.
Do you know where I can find it?

TIA!

the proper weight of the semantic loss

Hi,

the training script scripts/CycleGAN/cyclegan_gta_synthia2cityscapes.sh contains: 
--DSC --DSC_weight 20, 
but the python main() doesn't accept "DSC" parameters and it fails. 

But, a semantic weight of 0.2 as default, appears in base_options.py: 
self.parser.add_argument('--general_semantic_weight', type=float, default=0.2, help='weight for semantic loss')

Which one is the correct one?

TIA

Error when loading state_dict

Hello @Luodian ,

Hope you enjoyed the winter holidays! Thank you so much for this code release, it was like second Christmas for me!

Anyway, I tried running the model and I got reasonably far; however, I get the following issue when I try to replicate the model.

-------------- End ----------------
CustomDatasetDataLoader
dataset [GTA5_Cityscapes] was created
initialize network with normal
initialize network with normal
initialize network with normal
Traceback (most recent call last):
  File "train.py", line 20, in <module>
    model = create_model(opt)
  File "/efs/spot/MADAN/cyclegan/models/__init__.py", line 20, in create_model
    model.initialize(opt)
  File "/efs/spot/MADAN/cyclegan/models/multi_cycle_gan_semantic_model.py", line 92, in initialize
    self.netPixelCLS_SYN = get_model(opt.weights_model_type, num_cls=opt.num_cls, pretrained=True, weights_init=opt.weights_init)
  File "/efs/spot/MADAN/cycada/models/models.py", line 12, in get_model
    net = models[name](num_cls=num_cls, **args)
  File "/efs/spot/MADAN/cycada/models/drn.py", line 256, in drn26
    out_map=out_map, finetune=finetune, **kwargs)
  File "/efs/spot/MADAN/cycada/models/drn.py", line 174, in __init__
    self.load_state_dict(state_dict)
  File "/home/ubuntu/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 845, in load_state_dict
    self.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for DRN:
	size mismatch for fc.weight: copying a param with shape torch.Size([1000, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([19, 512, 1, 1]).
	size mismatch for fc.bias: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([19]).

I think there's a semantic channel space (19, COCO-style) and then there's the 1000 dim vector, which I am not 100% sure where that comes from.

Let me know if you have any ideas, thanks!

where is the create_model?

in the eighth line of the train.py document, I can't find the create_model. Do you miss it?

How to train with synthia?

Hi,
@Luodian I have download the code and found that cyclegan_gta_synthia2cityscapes.sh and cyclegan_synthia2cityscapes are the same. Then i want to use the script of GTA5, however the initial model weights are from drn26_cycada_cyclegta2cityscapes.pth. So if i train synthia, how to init the model? The same with GTA5 use drn26_cycada_cyclegta2cityscapes.pth? What if I train from my another own dataset, how to init the model?
Another problem is that how could i train the model end-to-end which as you claimed in paper?
Expect your reply.
Thanks.

About DeepLabV2 model in the table 8,9

Thanks for sharing your great research and code.
I am a student studying deep learning.

MADAN: Multi-source Adversarial Domain Aggregation Network for
Domain Adaptation, 2021, IJCV (https://link.springer.com/content/pdf/10.1007/s11263-021-01479-3.pdf)
In the above paper, there was a part where you experimented with DeepLabV2. If possible, can you provide the code or detailed description of the DeepLabV2 you used? ([email protected])
Did you simply use resnet101 backbone + aspp module + upsampling?
I'm not sure about the difference between DeepLabV3 and DeepLabV2. Please give me some advice.
Thanks for reading.