jiaminren / randwirenn Goto Github PK

Hi, I am excited to run your code. I run it on 4 TITAN Xp GPU. And yet, each epoch roughly takes 2 hours. Is that normal? What about your time to train an epoch?
Thanks!

what's the meaning of rand and world_size? They cannot be obtained from "os.environ". The program throws a error when it comes to here.

Hi, I have a question about weight initialization of node's additive aggregation here.

At model.py#L41, weights are initialized with torch.ones, not torch.zeros. This will make actual weights to be initialized with 0.7311, not 0.5000. Was it intended? (Though it won't affect much on training, I'm just curious about it.)

I try to run your code on 4 1080Ti GPUs, and I found that there is not enough memory under your default setting (256 batchsize, regular model). So I want to know how could you run the model on 8 Xp GPUs.
Thanks a lot.

请教一下，out = self.conv(out)的输入应该是三个变量，(inplanes, outplanes, stride)，为什么写成self.conv(out)，谢谢

In your implementation line 71~73 in model.py , nodes are divided into input and first-order ones. This may lead that second(or higher)-order(from input node) ones are ignored. Your implementation only consider the situation that the max distance between input and output node is 2. Am I right?

Hi, dear author,
thanks for your great work and open source code. I wonder is this repo. the official implementation of original paper? Thank you.
PS. It would be better if you can post your implementation performance benchmark in readme file.

请问这个代码里面是复现了论文的神经架构搜索吗？还是只训练了最好的架构。训练过程中网络的结构会改变吗？
Is this code retrieving the neural architecture search for the paper? Still only trained the best architecture. Will the structure of the network change during the training process?

Hi,
thanks for your contribution of NAS, it is very nice work.
but, the pytorch's data loader is very slow when you use a big dataset likes ImageNet2012.
Fortunately， I found mxnet's dataloader is a nice work, you can package ImageNet2012
into a rec file, then call mxnet's dataloader interface in your code, and convert mxnet data
to pytorch data.

Hello, I have an error message, I don't know if this file is missing. I hope to get your reply. Thank you。
FileNotFoundError: [Errno 2] No such file or directory: 'checkpoint/regular_c109_n32/conv2.yaml'

Hi, I’m author of https://github.com/seungwonpark/RandWireNN.

One SO question brought me here. At first, I was excited to see another implementation of RandWireNN. However, I’ve observed that much of your code resembles my implementation(plus, you even emailed me last Saturday), and was disappointed to see that there are no statement referencing my repo.

I’m not saying that this repo is just copy-and-paste of mine. There are several useful features that are not included in mine: multi-GPU support and tool for ImageNet data loading, and so on.

I would like to kindly ask you to add reference to my repository, if you have referred it.

Hi @JiaminRen Thanks for you work.
Use FP16 training on the basis of your code to achieve 75% accuracy

Have you realized the faster rcnn + FPN detection training of RandWireNN as backbone?
If you implemented it, is it implemented in mmdetection? Can you publish the training code? Thanks.

Hi, Jiamin. Many thanks for sharing your code.
I run the train.sh and found that
broadcast_params(model)
makes code stuck.

When I remove this code, the training is still stuck by
losses.update(reduced_loss.item(), input.size(0)) top1.update(reduced_prec1.item(), input.size(0)) top5.update(reduced_prec5.item(), input.size(0)).

I used single node 8 GPUs for training. The training procedure is the same as you described.
Do you know why this happens?