Hello, thanks for sharing your code! I found a possible wrong in cfvqasimple.py Isn't out['logits_all'] = z_qkv # for optimization should be out['logits_all'] = z_qk # for optimization ? Or I took it wrong?

I can train all the other versions except smrl_cfvqa_rubi with batch_size = 256, I have to change it to 64 preventing CUDA out of memory. But the training is too slow, which takes a day to train a epoch with three 3090.

I wonder what is the difference between smrl_cfvqa_rubi and other versions making it so speical, and is it normal to train so slowly? or is it because I did something wrong?

Hello Yulei Niu. Thank you very much for your inspiring work, it has given me great inspiration. I have a question that I hope you can answer: In your paper, under equation 17, there is a sentence "Only c is updated when minimizing L_kl", but in the code, I don't see how it is implemented, it seems that L_kl is added to the whole loss.

Hi @yuleiniu Thank you for your great work!

I have some questions related to the core idea, hoping answering them will make the paper more clear for me.

1- In equations 11, 12, and 13 you are replacing the learned embeddings by a learnable constant, how this constant may be interpreted? what does it imply? 2- Why fixing this constant across the whole dimension, by multiplying it to ones? 3- Following this,

        z_qkv = self.fusion(logits, q_pred, v_pred, q_fact=True,  k_fact=True, v_fact=True) # te
        z_q = self.fusion(logits, q_pred, v_pred, q_fact=True,  k_fact=False, v_fact=False) # nie
        logits_cfvqa = z_qkv - z_q

if we neglect the non-linearity (z = torch.log(torch.sigmoid(z) + eps)), (z_qkv - z_q) will be interpreted as (z_k + z_q + z_v) - (2C + z_q) which means we can just rely from the beginning on z_k + z_v and remove the QA branch?! I think I missunderstand something here :D 4- Is it possible to replace the constant with other real example, such as augmented version of the input or something like that, what do u think?

Thanks in advance!

Hi,

First of all, great work and a good paper!

I just want to clarify a few things! I followed the readme file and re-trained the following variants: i) vqacp2 (smrl_baseline.yaml / smrl_cfvqa_sum.yaml /smrl_cfvqasimple_sum.yaml) ii) vqa2 (smrl_baseline.yaml / smrl_cfvqa_sum.yaml /smrl_cfvqasimple_sum.yaml) However, the evaluation results were different from the results reported in the paper.

1. Were the models originally trained for 22 epochs (as indicated in the YAML files)? if not, what's the recommended epoch to achieve results similar to the ones stated in the paper?
2. The paper reports the model's performance on the vqacp2 test set. However, the readme file states that there is no test set for vqacp2. Can I assume that the eval set and test set are the same for vqacp2?
3. Upon training, the model report results on logits_all, logits_vq, logits_cfvqa, logits_q and logits_v. How do I relate these results to the ones reported in the table?

Thank you for your time.

Hi, Recently I'm doing my own VQA work. In my case, the convergence speed in three categories (Y/N, Num., Other) are not same. Hence, the best result for each category may reveal in three different epoch. I'm little confused on how to choose the best result for each category.

So, in your work, how do you choose the best result for each category? Choose the highest result for each category from all epochs result? Or just choose one epoch which has highest All score and choose all results from this epoch only?

logs.txt I try to run the baseline as "python -m bootstrap.run -o cfvqa/options/vqacp2/smrl_baseline.yaml", and got the logs.txt. However, I cannot get the overall accuracy as similar in your paper (about 38.46), and how to get the accuracy of different question types (Y/N, other and number) from this log?

Should I run more epoch or change some super parameter?

Could you please tell me whether this error affects the experimental results?

[I 2021-10-12 03:55:06] ...trap/engines/engine.py.126: Saving best checkpoint for strategy eval_epoch.accuracy_top1
[I 2021-10-12 03:55:06] ...trap/engines/engine.py.420: Saving model...
Traceback (most recent call last):
  File "/data/gaokuofeng/anaconda3/envs/cfvqa/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/data/gaokuofeng/anaconda3/envs/cfvqa/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/data/gaokuofeng/anaconda3/envs/cfvqa/lib/python3.7/site-packages/block.bootstrap.pytorch-0.1.6-py3.7.egg/block/models/metrics/compute_oe_accuracy.py", line 8, in <module>
ModuleNotFoundError: No module named 'block.external'
[I 2021-10-12 03:55:07] ...trap/engines/engine.py.424: Saving optimizer...
[I 2021-10-12 03:55:10] ...trap/engines/engine.py.428: Saving engine...
[I 2021-10-12 03:55:10] ...trap/engines/engine.py.129: Saving last checkpoint
[I 2021-10-12 03:55:10] ...trap/engines/engine.py.420: Saving model...
[I 2021-10-12 03:55:11] ...trap/engines/engine.py.424: Saving optimizer...
[I 2021-10-12 03:55:14] ...trap/engines/engine.py.428: Saving engine...
[I 2021-10-12 03:55:14] ...trap/engines/engine.py.133: Ending training procedures

Hi @yuleiniu Thank you for your great work! I have two quick questions:

1. It seems to me that the core idea is very similar to Tang's unbiased SGG (CVPR'20) in that both works aim to remove the bad co-occurrence bias by subtracting the results with certain data blocked out (image modality/image patches). Is there any misunderstanding here?
2. The discussion of "good" and "bad" biases: it seems to me that the "bad" language bias can be removed by the proposed method; however, it seems to also remove the "good" ones. I didn't find a detailed discussion on your main motivation of removing the bad biases and retaining the good ones, neither further discussion beyond the Introduction section nor experimental proofs, in the paper. How do the good ones remain? Could you please elaborate on this?

I found the files in the networks folder that the block module is introduced in the header file, such as "from block.models.networks.mlp import MLP". How do I get the block module?

配置好环境后运行，在bootstrap/lib/logger.py里的json.dump()报TypeError: Object of type Tensor is not JSON serializable，请问作者有遇到这个问题吗？不解决是否可以，把logger.flush()注释掉后train和eval的结果都有了。