Questions about difference between code and paper

[chrissyguo]

Question 1:
In def forward_single(self, x) of ~mmdet/dense_heads/MIAOD_retina_head.py​:
y_head_cls = y_head_f_mil.softmax(2) * y_head_cls_term2.sigmoid().max(2, keepdim=True)[0].softmax(1)
It seems that this code (with max(2, keepdim=True)[0]) is not in accordance with the equation (5) in the paper:

What is the function of max(2, keepdim=True)[0]? And why softmax(1) after all computation? I thought the code should be:
y_head_cls = (y_head_f_mil.softmax(2)) * (y_head_cls_term2.sigmoid().softmax(1))according to my understanding.

Question 2:
For uncertainty calculation definition in the paper:

The instance uncertainty is defined as the prediction discrepancy of f_1 and f_2.

and the equation for discrepancy is:

After reweighting, the equation has changed to:

I thought the code of uncertainty for loss computation and the code of uncertainty for data selection should be the same.
However, the code in def l_wave_dis of ~mmdet/dense_heads/MIAOD_head.py​:

l_det_cls_all = (abs(y_head_f_1_single - y_head_f_2_single) * w_i.reshape(-1, self.C)).mean(dim=1).sum() * self.param_lambda

directly compute the discrepancy by subtraction (Manhattan Distance?), while the code in def calculate_uncertainty of ~mmdet/apis/test.py​ is:

loss_l2_p = (y_head_f_1 - y_head_f_2).pow(2)

which uses pow(2) for square computation (Euclidean Distance?). I would like to know why you have such designs in code.

Thank you so much!

[yuantn]

Thanks for your questions.

Answer 1:
The function of max(2, keepdim=True)[0] is to highlight the class with the highest score (i.e., to make a [2, 68400 (anchor number), 20 (class number)] tensor to a [2, 68400, 1] tensor). By using this maximum function, we only focus on the classes and bounding boxes that are most likely to be predicted as the foreground, and that is the goal of multiple instance learning. In fact, this is an auxiliary effect for the first term in this equation.

The effect of softmax(1) function is the same as that in the first term. Its effect is:

Answer 2:

There are indeed differences between the l_wave_dis in the paper and that in the code.

Computing the discrepancy by subtraction directly (called Manhattan Distance in Digital Image Processing but L1 loss in Deep Learning here) and by square (called Euclidean Distance in Digital Image Processing but L2 loss in Deep Learning here) are two ways to calculate loss. Their effects are basically the same but minorly different.

In addition, as shown in our previous experiments (which are not provided now), there is not much differences in performance between using these two types of loss. So we did not unify them in the paper and code before.

I will re-run the code using L1 loss, and enclose the performance in this issue for comparison to prove my inference above. Please wait a few days.

[yuantn]

Here are the results using L1 loss in Answer 2 and those in the paper:

Proportion (%) of Labeled Images	5.0	7.5	10.0	12.5	15.0	17.5	20.0
mAP (%) using L1 loss	46.84	59.36	63.22	67.36	69.33	70.90	71.48
mAP (%) in the paper	47.18	58.41	64.02	67.72	69.79	71.07	72.27
mAP (%) of the differences	-0.34	0.95	-0.8	-0.36	-0.46	-0.17	-0.79

Here is the output log: Google Drive | Baidu Drive (Extraction Code: 27ql)

[liaorongfan]

In tools/trains.py the code minimizes the uncertainty first and then to maximizing the uncertainty, which is not in line with the paper. I was wondering if whether there were some special concerns.

[yuantn]

In tools/trains.py the code minimizes the uncertainty first and then to maximizing the uncertainty, which is not in line with the paper. I was wondering if whether there were some special concerns.

Please kindly refer to here for explanation.

[liaorongfan]

@yuantn No mean for offences, just curious, sorry

[yuantn]

@yuantn No mean for offences, just curious, sorry

No need to be so cautious... I'm OK! 😁

I answered like that just because someone has asked your question before.

Feel free to ask any questions you want, but it would be better to open another issue if your question is different from the current issue.

[liaorongfan]

thanks a lot

[chrissyguo]

Thank you so much for answering!