Hello! Thank you for sharing this very well organized code. Additionally, I wonder if you could clarify a few points for me:

• F.relu (used in inference_CCAM.py) is commonly used during inference of multi-class CAMs to squash negatively contributing pixels and unrelated pixels together. However, in a binary case, the 0 means uncertain if background or foreground, and it does give you some additional information, no? Why did you choose to go like this, instead of simply checking ccam > 0.0 or taking the sigmoid function and checking for sigmoid(ccam) > 0.5?
• a) Which hyper-parameters did you use to train PoolNet? Was all of them set to default? b) Did you produce the .lst file yourself, pointing out to the locations of the saliency maps produced by inference_CRF.py? c) Did you use only images/pseudo saliency maps in train.txt, or all files listed in train_aug.txt?
• I noticed CCAM has a negative effect on the IoU of classes chair, diningtable and sofa (image). From these, only chair's IoU degeneration is reported in the paper, but not discussed. I guess you ran out of space? Do you have any insights on why that happened?

Thank you very much once again, cheers.

Thanks for your working! But when i runing the third step of the WSSS task , there is 'No module named 'core.puzzle_utils''. i can't find this module in the 'core' folder. Please help me.

Sorry to bother you again.When i use train_CCAM.py in CUSTOM to train a model using custom data, a got an EEOR:

`/data/yunzi/CCAM/CUSTOM$ OMP_NUM_THREADS=16 CUDA_VISIBLE_DEVICES=5 python train_CCAM.py --tag CCAM_Maket1501_MOCO --batch_size 1 --pretrained mocov2 --alpha 0.25 [i] CCAM_Maket1501_MOCO

[i] mean values is [0.485, 0.456, 0.406] [i] std values is [0.229, 0.224, 0.225] [i] train_transform is Compose( Resize(size=(256, 128), interpolation=bilinear, max_size=None, antialias=None) RandomHorizontalFlip(p=0.5) ToTensor() Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)) ) [i] #train data

[i] log_iteration : 2,587 [i] val_iteration : 12,936 [i] max_iteration : 129,360 Loading unsupervised mocov2 pretrained parameters!

[i] Architecture is resnet50 [i] Total Params: 23.54M

Traceback (most recent call last): File "train_CCAM.py", line 229, in visualize_heatmap(args.tag, images.clone().detach(), ccam, 0, iteration) File "/data/yunzi/CCAM/CUSTOM/utils.py", line 49, in visualize_heatmap for j in range(axes.shape[0]): AttributeError: 'AxesSubplot' object has no attribute 'shape'`

Do you have the same problem?

Hi,thanks for the work. But when i run/data/yunzi/CCAM/WSOL$ OMP_NUM_THREADS=16 CUDA_VISIBLE_DEVICES=0 python train_CCAM_CUB.py --experiment CCAM_CUB_MOCO --lr 0.0001 --batch_size 16 --pretrained mocov2 --alpha 0.75

I got a error :ModuleNotFoundError: No module named 'cmapy' where can i import this file?

class SimMaxLoss(nn.Module): def init(self, metric='cos', alpha=0.25, reduction='mean'): super(SimMaxLoss, self).init() self.metric = metric self.alpha = alpha self.reduction = reduction

def forward(self, embedded_bg):
    """
    :param embedded_fg: [N, C]
    :param embedded_bg: [N, C]
    :return:
    """
    if self.metric == 'l2':
        raise NotImplementedError

    elif self.metric == 'cos':
        sim = cos_simi(embedded_bg, embedded_bg)  #(N,N)
        loss = -torch.log(sim)
        loss[loss < 0] = 0
        _, indices = sim.sort(descending=True, dim=1)
        _, rank = indices.sort(dim=1)
        rank = rank - 1
        rank_weights = torch.exp(-rank.float() * self.alpha)
        loss = loss * rank_weights
    else:
        raise NotImplementedError

    if self.reduction == 'mean':
        return torch.mean(loss)
    elif self.reduction == 'sum':
        return torch.sum(loss)

Hello! First of all, you did a great job! Congratulations! Second, I have three questions I'd like to ask you about running the code.

1. First, in the case of WSOL, I didn't find the test file test.py in the code. Is the result on the CUB-200-2011 test set and imagenet-1k validation set mentioned in this paper calculated by def test in train_ccam_cub. py?

2. Secondly, the three evaluation Top-1 Loc, Top-5 Loc mentioned in the paper were not found in the code to calculate Top-1 Loc and Top-5 Loc, so there was no Top-1 Loc and Top-5 Loc in the running result of my CUB-200-2011 dataset. How to solve this problem?

3. Thirdly, the C2AM(Ours) in Table1 uses backbone DenseNet161 and Efficientnet-B7. I did not find the part of the code downloading and replacing these two backbone networks in the code, so that the results in C2AM(Ours) in Table1 could not be reproduced. How to solve this problem?

the backbone of CAM and CCAM is the same or different one? when i use the ccam to refine the CAM, should i start a new network to generate the ccam separately?