#4 has described potential label leakage in the code, but the reply has not covered it.

According to the author's kind explanation, the preprocessing code has once served necessary purposes in ASL's implementation. However, ASL was evaluated on fully labeled COCO. In partial label settings, it is worried that the same split-merge code can lead to potential label leakage, which is important but not solved in the previous closed issue.

Could the authors check if this code causes the main problem mentioned in #4 ? The code seems to result in a larger proportion of known labels than is actually set.

Thanks.

Thank you for your great work.

While reading the code, I find something confused me in data preprocessing:

# In coco_detection.py, line 105......
output = torch.zeros((3, len(self.classnames)), dtype=torch.long)
for obj in target:
    if obj['area'] < 32 * 32:
        output[0][self.cat2cat[obj['category_id']]] = 1
    elif obj['area'] < 96 * 96:
        output[1][self.cat2cat[obj['category_id']]] = 1
    else:
        output[2][self.cat2cat[obj['category_id']]] = 1
target = output
if self.mask is not None:
    masked = - torch.ones((3, len(self.classnames)), dtype=torch.long)
    target = self.mask[index] * target + (1 - self.mask[index]) * masked
# ......
# In trainers.py, line 109
target = target.max(dim=1)[0]

Why split the annotations into 3 parts, mask out, and then merge them? It is odd, and also seems to make the proportion of known labels larger than the settings. Only if the masks for 3 parts are all 0 will the label be dropped. For example, 87.5%(0.5^3) labels will be retained instead of 50% if set the proportion to 0.5 following the README.

It would be better if the authors replace the code with a simple one and reproduce the results.

Looking forward to the release code.
When will the code be released?

Dear author, I find the cls_ids.pickle file missing while reproducing the zero-shot experiment on the coco dataset.
Specifically, I do not find the cls_ids.pickle of coco in the download link (https://drive.google.com/file/d/154dkD7Ok1xxwTZb7hQTAf3FlNCN8Q6KM/view?usp=sharing) that you give.
Please tell me how to get the cls_ids.pickle or directly give the file.

Dear author, have you done any experiments on the openimages dataset and what are the results

Thank you for your great work.

While reading the code, I find something confused me in data preprocessing:

# In coco_detection.py, line 105......
output = torch.zeros((3, len(self.classnames)), dtype=torch.long)
for obj in target:
    if obj['area'] < 32 * 32:
        output[0][self.cat2cat[obj['category_id']]] = 1
    elif obj['area'] < 96 * 96:
        output[1][self.cat2cat[obj['category_id']]] = 1
    else:
        output[2][self.cat2cat[obj['category_id']]] = 1
target = output
if self.mask is not None:
    masked = - torch.ones((3, len(self.classnames)), dtype=torch.long)
    target = self.mask[index] * target + (1 - self.mask[index]) * masked
# ......
# In trainers.py, line 109
target = target.max(dim=1)[0]

Why split the annotations into 3 parts, mask out, and then merge them? It is odd, and also seems to make the proportion of known labels larger than the settings. Only if the masks for 3 parts are all 0 will the label be dropped. For example, 87.5%(0.5^3) labels will be retained instead of 50% if set the proportion to 0.5 following the README.

It would be better if the authors replace the code with a simple one and reproduce the results.

Hi!
Your work is so interesting. Please, could you provide guidelines and/or requirements for data preparation before training your code with new datasets?

Thank you in advance.

@sunxm2357 Hi thanks sharing the code base wonderful work i have few queries

1. can we training the existing pipeline for automotive data which multilable classification like showing in figure below
   

2. Since it is used for multilabel classification can we add features like prompts to improve the accuracy of the model
   Thanks in advance

I try to reproduce the codes according to the descriptions in the original paper combined with the open-sourced CoOp, however, I couldn't obtain the results in the paper. It will be better if the authors could release the open-sourced codes.

Hi there,

I have cloned the DualCoOp repo and installed all the dependencies except dassl. Where do I install dassl? It is its own repo so I'm not sure how to set it up accordingly "within" the DualCoOp repo. Thanks.

TypeError: multi_head_attention_forward got an unexpected keyword argument 'average_attn_weights'

Dear author.
I found your work so interesting.
Then, I have a question about training settings to reproduce the results in your paper.
Can you share each value of loss_w related to pp, respectively?
I really appreciate any help you can provide.

Hello,

I would like to thank you for releasing the code of your very inspiring work!

I am trying to reproduce your results and I was wondering why there is a multiplication by 20 line 232 of the "models.dualcoop.py" file:
output = 20 * F.conv1d(image_features_norm, text_features[:, :, None])

I have the same question regarding the multiplication by 5 line 237 of the same file:
output = 5 * (output * w).sum(-1)

Regards

Thank you for your great work!

Would you mind sharing your code? Thanks!

About whether the network can be replaced by VIT for this