zengxianyu / jsws

Joint learning of saliency detection and weakly supervised semantic segmentation

Hi there 👋

zengxianyu/zengxianyu is a ✨ special ✨ repository because its README.md (this file) appears on your GitHub profile.

Here are some ideas to get you started:

• 🔭 I’m currently working on phd
• 🌱 I’m currently learning machine learning
• 👯 I’m looking to collaborate on computer vision

jsws's Issues

file naming bug

in your readme, it says
conda env create --file=environments.yaml
But the file in your repo is environment.yml, should fix that

cannot import name 'Logger'

Another similar issue:

File "weak_seg_full_sal_train.py", line 16, in <module>
    from logger import Logger
ImportError: cannot import name 'Logger'

There is a folder called logger, but it is empty

what do the loss values represent

in the following output:

iter 1500 loss_sal -0.9201; loss_cls 0.0697

what do the loss_sal and loss_cls mean?

Also should loss_cls be negative?

Calculate miou on training data set

I noticed that miou is only calculated on validation data set
What would i have to do if I want to calculate the miou on the training data set as well?

Paper code

Hello, I am a graduate student, now I want to reproduce your paper, what I downloaded is the version you uploaded before, I see you deleted all the previous ones, is there any mistake?Looking forward to your reply. I think it is very important to me，thanks

About training details

Thank you for repo. For the pseudo segmentation labels used in the second training stage, did you use the offline saved outputs from SSNet1 trained in the first stage?

runtime error

I was running your code, and after 30,000 iterations, I got this error:

Traceback (most recent call last): File "/home/s2bashar/anaconda3/envs/thesis/lib/python3.6/multiprocessing/queues.py", line 234, in _feed File "/home/s2bashar/anaconda3/envs/thesis/lib/python3.6/multiprocessing/reduction.py", line 51, in dumps File "/home/s2bashar/anaconda3/envs/thesis/lib/python3.6/site-packages/torch/multiprocessing/reductions.py", line 333, in reduce_storage RuntimeError: unable to open shared memory object </torch_20991_592962625> in read-write mode

Pred_seg

I noticed this line in the second loss calculation:

pred_cls = pred_seg.mean(3).mean(2)

pred_seg has the shape torch.Size([8, 21, 256, 256])
what is pred_seg, and what do the values in the shape represent?

SNet-S and SNet-m

Hi
Were do you specify SSNet-S. SSNet-M for training?

Could not find the 'trainself.txt'

when I run the 'self_seg_full_sal_train.py', I can't find the 'trainself.txt'，Can you tell me how to get it？ Thanks！

Why is background not considered in calculating loss?

In lines 178 and 179 of weak_seg_full_sal.py, while calculating loss 2, why is the background omitted?

pre trained models as backbone

Hi
Do you use any pre trained models as backbone for this project?

Some training problems

1）What is the F-measure and MAE result of stage1 in the multi-task setting？
2）Why are the Miou results of syn_stage1(val_voc) and test_stage1 different in the same model?
3）Why is the result of test_stage2 lower than that of test_stage1

Model

Can you provide trained models for us to test demo pictures?

IOU of individual classes

Is there a way to see the Iou of individual classes as opposed to just the overall mIou?

File missing

This file is missing in self_seg_full_sal_train.py line 52:
voc_train_split = './trainself.txt'

is this the same as trainaug.txt used in weak_seg_full_sal_train.py?

Semantic Loss function

In the paper, Lc is defined as follows:

In the code, It is:
loss_cls = F.binary_cross_entropy_with_logits(pred_cls[:, 1:], gt_cls[:, 1:]) + F.binary_cross_entropy_with_logits(pred_cls32x[:, 1:], gt_cls[:, 1:])
Which one of these are t_i and which one is t_i hat?
Also why are they two loss functions in one? as in what is the difference between pred_cls and pred_cls32x?

Missing files

youre missing the file val.txt for
voc_val_split = '/home/zeng/data/datasets/segmentation/VOCdevkit/VOC2012/ImageSets/Segmentation/val.txt'
as well as this one:
sal_val_img_dir = '/home/zeng/data/datasets/saliency/ECSSD/images'

Loss of Accuracy

So my computer has limited memory
it only runs 18,000 iterations, before dying
So when I restart from the step it died in, the mIOU decreases
so what i mean:
mIOU upto 18000 iter is 0.5547
then my memory limit is reached, and the program dies
i restart from that checkpoint, and mIOU is 0.5512 for the next 18000 iters
then my program gets killed because of memory issues
when i restart from the checkpoint of 36000, mIOU is 0.5178

Is there any way to fix this?

No file named evaluate_sal.py

Hi, I am trying to run the code, but I get this error:

File "weak_seg_full_sal_train.py", line 12, in <module>
    from evaluate_sal import fm_and_mae
ModuleNotFoundError: No module named 'evaluate_sal'

there indeed is no file named evaluate_sal.py

Stage 1 training killed

I am running stage one, But after iter 1300, it says killed
In the code it says it will run for 100000 iterations, but after 1300, it says killed

Is that supposed to happen?