Dear Hsinchu,
Thanks a lot for your opening code about your work. And when I run trainval_net_MEAA.py with adding "--mGPUs", there is an error "TypeError: forward() missing 4 required positional arguments: 'im_data', 'im_info', 'gt_boxes', and 'num_boxes'
". However, the error doesn't occur without "--mGPUs". Can you help me to solve it?

Best wishes!

Hello,
Currently, I am trying the code on colab. However, when it disconnects, I change resnet path field in the config file to start from learnt weights. But accuracy drops too much and loss also starts from very high point. So, how can we resume training from a checkpoint ?

It seems that nms_cuda.c are missing. I encountered this problem when I compile code.
I follow this jwyang/fpn.pytorch#1.

Thanks

I ready to predict the picture with the model,but i find model.faster_rcnn.vgg16_global and model.faster_rcnn.resnet_global are lack.

from model.faster_rcnn.vgg16_global import vgg16

from model.faster_rcnn.resnet_global import resnet

they are used in demo_global.py.SO i hope author can upload them.Thanks

 # initilize the network here.
  from model.faster_rcnn.vgg16_SCL import vgg16

  if args.net == 'vgg16':
    fasterRCNN = vgg16(foggy_classes, pretrained=True)

    # fasterRCNN = vgg16(pascal_classes, pretrained=False, class_agnostic=args.class_agnostic,lc=args.lc,gc=args.gc)
  elif args.net == 'res101':
    fasterRCNN = resnet(pascal_classes, 101, pretrained=False, class_agnostic=args.class_agnostic,context=args.context)
  elif args.net == 'res50':
    fasterRCNN = resnet(pascal_classes, 50, pretrained=False, class_agnostic=args.class_agnostic)
  elif args.net == 'res152':
    fasterRCNN = resnet(pascal_classes, 152, pretrained=False, class_agnostic=args.class_agnostic)
  else:
    print("network is not defined")
    pdb.set_trace()

  fasterRCNN.create_architecture()

It seems that some files are missing.

Traceback (most recent call last):
  File "demo_global.py", line 192, in <module>
    from model.faster_rcnn.vgg16_SCL import vgg16
ModuleNotFoundError: No module named 'model.faster_rcnn.vgg16_SCL'

Thanks

Hi,
I would like to ask the number of GPUs used for training and the training time.
Thanks.

How to generate an attention map，thanks

Hi,
I use cityscapes to be source domain which has 2964 images after filtering, night scene of bdd100k to be target domain which has 19864 images after filtering. There are 8 categories:
bicycle, bus, car, motorcycle, person, rider, train, truck and use python trainval_net_MEAA.py --cuda --net vgg16 --dataset cityscape --dataset_t bdd100k --use_tfb

I would like to ask how to avoid getting NaN loss, which mostly happens in rcnn loss. I get NaN loss at the 4th epoch and 4300 iterations.

By the way, does it matter if the source domain and target domain training samples are unbalanced?

And I modify these because pixel indexes is 0-based.

 x1 = float(bbox.find('xmin').text)
 y1 = float(bbox.find('ymin').text) 
 x2 = float(bbox.find('xmax').text) 
 y2 = float(bbox.find('ymax').text)

Thanks.

----UPDATE----
I followed this jwyang/faster-rcnn.pytorch#136 (comment) but only checkedxmin==xmax and ymin==ymax and boundary coordinate, more importantly, I made coordinates convert to integers instead of floating-point, because the BDD100K coordinates were originally floating point.
I trained the model for 20 epochs and didn't get NaN Loss error.
But, I get a error after save model at the final epoch but it doesn't matter.

Exception ignored in: <bound method _DataLoaderIter.__del__ of <torch.utils.data.dataloader._DataLoaderIter object at 0x7efdfc15f240>>
Traceback (most recent call last):
  File "/media/wyt/Data3/conda_env/DA-OD-MEAA/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 399, in __del__
  File "/media/wyt/Data3/conda_env/DA-OD-MEAA/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 378, in _shutdown_workers
  File "/media/wyt/Data3/conda_env/DA-OD-MEAA/lib/python3.6/multiprocessing/queues.py", line 337, in get
AttributeError: Can't get attribute 'Tensor' on None
Exception ignored in: <bound method _DataLoaderIter.__del__ of <torch.utils.data.dataloader._DataLoaderIter object at 0x7efe17fc7f60>>
Traceback (most recent call last):
  File "/media/wyt/Data3/conda_env/DA-OD-MEAA/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 399, in __del__
  File "/media/wyt/Data3/conda_env/DA-OD-MEAA/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 378, in _shutdown_workers
  File "/media/wyt/Data3/conda_env/DA-OD-MEAA/lib/python3.6/multiprocessing/queues.py", line 337, in get
AttributeError: Can't get attribute 'Tensor' on None

I tested the uploaded foggy cityscape trained model, but the result is different from the picture you provided. I do not modify any code. Why is this reason?