Questions about the competing method in Figure 5 (c)
Closed this issue · 6 comments
Hello,
Thank you for the fantastic work. May I ask a question about Figure 5 (c)? I notice the Core-set is listed as a competing method on COCO, which is an important result, and I believe this is the first paper that presents it.
In fact, I am trying to realize the same result on COCO. However, I find that to get the result I need to infer the model on the whole COCO training set (~ 118k images) and compute the distance for every pair of images, which is memory and time consuming: the distance matrix size is nearly 40GB. Moreover, running mixed integer programming (MIP) is even more expensive.
Therefore, may I ask how did you work this out? Did you run the full Core-set algorithm that contains the MIP? Or only the Greedy K-Centroid was used?
Additionally, could you please provide the initial COCO training set (the 2% one)? As active learning is prone to variation in initialization. Starting from the same initialization can make sure the following works to have a fair comparison with yours.
Looking forward to your answer :-)
Hello,
Thanks for your attention to our work.
For Core-set, we use the function in this repository for implementation. So we only use the Greedy K-Centroid method.
For COCO training set, we set the random seed here before initialization. You may use the same random seed 666.
Hope this answer would be useful to you :-)
Cool!. Thank you for the quick reply. It's very helpful. :-)
Hi. Sorry for bothering again. May I ask how many epochs did you train for the competing methods that did not require adversarial training? Is it still 26 epochs?
Yes.
Hi. But I found that a simple ResNet50 RetinaNet model with 26 epoch training can achieve ~14 AP using your initial set (2% images), which is far better than the numbers reported in the paper. Is there anything wrong?
Oh sorry, just found a bug in my code.... My bad :-(