Hi @soo89,

Wonderful work.

Could you please help understand how can I train this semi-supervised model on my own dataset. I only have images and xml files. I have not worked with VOC/COCO datasets before. Could you please give some suggestions about where should I start? Do I need to split my dataset into labelled and unlabelled datasets to perform semi-supervised learning?

I appreciate your response.

in paper, the consistency_conf_loss is JSD. but in the code, the consistency_conf_loss is sum of 2 KL How do I understand this consistency_conf_loss_a and consistency_conf_loss_a. JSD should be like this

thanks

Hi Jisoo, thanks for your nice work. could you tell me what --dataset is? Does it mean using the different configuration settings for different sizes of inputs? For example, VOC300 refers to the setting for an ssd300 model while VOC512 refers to ssd512?

I'm interested in this paper. So I'd like to train this model on custom datasets, but failed to train. Could you tell me which version of pytorch and cuda you use? thank you!

Hi author, Thanks for your wonderful work and your released code. I am confused about the code for the Localization consistency loss in https://github.com/soo89/CSD-SSD/blob/30b184c86a87b0fc2d301cd8157b11d8cfe7da1e/train_csd.py#L297.

consistency_loc_loss_x = torch.mean(torch.pow(loc_sampled[:, 0] + loc_sampled_flip[:, 0], exponent=2))
  
consistency_loc_loss_y = torch.mean(torch.pow(loc_sampled[:, 1] - loc_sampled_flip[:, 1], exponent=2))
                
consistency_loc_loss_w = torch.mean(torch.pow(loc_sampled[:, 2] - loc_sampled_flip[:, 2], exponent=2))
                
consistency_loc_loss_h = torch.mean(torch.pow(loc_sampled[:, 3] - loc_sampled_flip[:, 3], exponent=2))

For y, w, and h you use '-' when calculating the mean squared loss. I am confused about why you use '+' for x when calculating the mean squared loss.

Thanks for you great work! read the code, I find some code which I can't to understand, as below, why it need to flip append loc and append conf?

by the way, we compare loc and flip loc in

consistency_loc_loss_x = torch.mean(torch.pow(loc_sampled[:, 0] + loc_sampled_flip[:, 0], exponent=2))
...

but, how we guarantee each item between loc_sampled and loc_sampled_flip can match correctly?

Hope your kindly response, much thanks.

hi, in the experiment section, you mentioned that labeled and unlabeled data are randomly shuffled and selected. I wanted to know why it's necessary to have each mini-batch containing both labeled as well as unlabeled data?

What happens if one of mini-batch contains only unlabeled data (since the majority of data is unlabeled)? Can we handle that case? What would be the value of loss?

I'm new to semi-supervised. thanks.

Hello,

Thanks for your nice work. I just have few questions about training on GPUs.

1. Do you train your CSD on a single GPU or multiple GPUs? If trained on single GPU, do you mind sharing the series code(what type of GPU)? I am asking this because I am not able to train CSD with batch size of 32 on a single 2080 Ti and I'm not sure whether the problem is on my server or the nature of the code.

2. What's your estimated training time on VOC(07 labeled and 12 unlabeled) in your configuration? It looks like training a baseline SSD on a single GPU would take me several days in my case so it would be nice if you could share the stats....

Thanks for your time and enjoy your day!

Hello, Thanks for your nice work!

When I downloaded your code, I tried many times but failed. I thought maybe beacause of your "weight" folder that had the 70 M pretrained model.

Could you please delete the weight folder and offer another pretrained model link?

Thank you very much!

Hi, thanks for providing such nice work! I have a minor question about the loss weight schedule. For the consistency loss that you've designed, is there any experimental results that ramp-down weight scheduling affects the performance? Ramp-up seems plausible for stable training but I'm not sure why ramp-down needed.

Hi,

Thanks for sharing your code. I would like to inquire more about the JSD loss for classification consistency. specifically this line Link

Why is .detach() specifically used here for computing the losses?

And, I didn't understand how it impacts when you are detaching either output in each line? If I print requires_grad on those tensors it shows True.