jeff-sjtu / sampling-argmax Goto Github PK

hi @Jeff-sjtu
in the end of scripts/train_pose.py , i see DPG training? can u explain what it is?

Hi, in the function of norm_heatmap, it writes gumbel_heatmap = heatmap - log_eps / tau, which is different from gumbel_heatmap = (heatmap - log_eps)/ tau in the gumbel_softmax. Is this matter when training with annealing strategy？

Hi, do you replace the sampling softmax of softmax in any dense prediction task and is it still improved？

Thanks for your wonderful work. But I am confused about the custom backward in ClipIntegral, why the derivative of this operation is binarized as {-1, 1} by the comparison between weight and output_coord ?

sampling-argmax/sampling_argmax/models/simplepose.py

Line 87 in 12c8074

weight_mask = torch.ones(weight.shape, dtype=grad_output.dtype,

Thanks for your great work!
I've just read your paper and the discussions on OpenReview, and have the same question with the reviewer——

Although it's indicated in "We conduct an experiment of training the model with Eq. 5. The model only obtains 30.9 mAP on COCO Keypoint" that training with discrete density map can not work, I still wonder why sampling (works) and why "training with only sample outperforms soft-argmax"

@Jeff-sjtu 您好，很棒的一篇论文，看到论文中的图之后有一点不太懂的地方，希望您能帮忙解答一下
1）图中经过第三步之后的最终输出是一个xy坐标还是一个heatmap的最大值索引？
2）图中第二步的子分布的个数如何确定？

Hello, the paper mentions you use an annealing schedule going from high to low temperature. In the review its mentioned how the method is parameter free, I take it you found a single annealing schedule that worked effectively for the temperature during training? I was wondering since the code is not published yet if you could provide the schedule you used.

On a side note it's great to see more research on soft-argmax! Thank you to all of the authors 👏

Hi, thanks for your great work! I'm wondering for sampling-argmax and RLE, which is better in localization?