gsunshine / enjoy-hamburger Goto Github PK

According to the paper, isn't the gradient of the MDs should be one-step gradient? However, the code of NMF dose not apply torch.with_on_grad() on the local_inference of NMF and _MatrixDecomposition2DBase. Could you please provide some explanation on this difference?

Hello,

I have a question regarding the batch normalization. Is there a reason why you choose such a small momentum 3e-4 for the batch normalization?

Thank you in advance.

Hi, @Gsunshine

I notice that the implementation of one-step gradient in your code only consists the build of coefficient, as demonstrated by this line.

In my opinion, the function F is composed of two parts: the construction of coefficient and base, but the base part is omitted.

Is this a deliberate design or a mistake?

Thank you for your great work. Best wishes.

When are you going to release either the arxiv version or the blog?

请问一下，我用train.sh能够将模型跑通，因为是单卡，GPUS=1，我想深入了解下模型的参数，想尝试debug，但是已经设置了--gpus 1，norm_cfg = dict(type='BN', requires_grad=True)，仍然报错Default process group has not been initialized, please make sure to call init_process_group.请问有什么办法解决吗？

Thank you for your excellent work! I note the sample code in the blog said there should be a fixed point iteration before the one-step grident, so that it's guaranteed to be a contraction mapping right? I search the code with keyword "fix", "fixed", "iterations"and so on, but I cannot find fixed point iteration. So where is it?

I've been working on applying Hamburger to other detection models (to be specific, Mask2Former & SparseInst), mainly by inserting Hamburger after the neck to align the multi-scale features, but the training process always collapses after only several iterations because of the nan output.
Given that the training recipe is rather general, and further reducing the lr does no help, I guess this indicates the gradient propagation is unstable? (p.s. applying @torch.no_grad() to local_inference() is also unhelpful)
Thus I'm wondering what's the intrinsic cause for this? have you ever met similar cases? or any suggestions for a fix?

Any idea would be appreciated.

Hi there,
I set up environment on docker with torch=1.11.0, cuda=11.3, mmcv-full=1.5.0. When I ran the code, I got this trace back info shown in picture

the "hamenet_light_van_tiny_512x1024_160k_cityscapes.py" is the config file modified by myself based on ade20k one. The error msg seems to be unrelated with data, so it shouldn't be a problem.
Could you please give me any suggestion?
Thanks!

I give try to read the arxiv paper but failed to understand the mathematical intuition of the update rules.

In section 2.2.2,

I didn't understand the last line, "...and softmax is applied column-wise and $T$ is the temperature. Further, we can obtain a hard assignment by a one-hot vector when $T \rightarrow 0$." What is $T$ (Temperature) here?

I couldn't get the justification for the replacement of $\arg \min$ with softmax. And what is their update rule doing for $D\leftarrow XC^{T}\text{diag}(C1_n)^{-1}$

All of these questions might be irrelevant to the GitHub issue, but I was really struggling to understand these facts. It will be a great help if these things are described here. Thanks in advance.

hello
Thank you for the models you provided in the field of semantic segmentation. I encounter this error while training hamnet_van_base_512x512_160k_ade20k model using Custom Dataset. I have two more classes. Separated from the background, I made 20K as the ade20k dataset. In debugging, I saw that the addressing of the data set was correct.
Please help me.
correct_k.mul_(100.0 / target[target != ignore_index].numel())) ZeroDivisionError: float division by zero

源码中没有VAN 主干网络吗==！什么时候加上呢，期待