licongguan / ilm-assl Goto Github PK

Hello! Thank you for your excellent work on open source. But I wonder where does the active learning code run? Or do i need to run it myself manually?

While running the code, I encountered the error message "RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [256]] is at version 3; expected version 2 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later."
Upon further investigation, I found that the issue is related to "Error detected in CudnnBatchNormBackward0. No forward pass information available. Enable detect anomaly during forward pass for more information." As I am not familiar with distributed training, I would like to inquire if the community has any insights into this problem.

Here are the specific output logs:
log_20230810_161918.txt

Hello Author, can STAL be trained on a single GPU, such as the 3090?

hello,I want to know how to resolve ''Could not find a version that satisfies the requirement cityscpaesscripts (from versions: none)'' when I use pip install cityscpaesscripts in windows environment.

I don't find the file "python ILM-ASSL/datasets/rename.py"

Hi, Thank you for sharing a nice work!
I am wondering if there is a log for training since I want to know whether I'm training correctly to reproduce the paper's result.

I'm working with GTA2Cityscapes 1% protocol to reproduce the result. By now, I'm in 50 epoch and the mIoU is around 43 which is far behind the paper result (which is 70).

Thanks in advance!
Joo Young Jang

I noticed that when you use Dataloader in your code, you use cfg["data_list"] directly. In fact, there are two data_lists in cfg, how do you make sure that you are loading the correct list? Also, there are misspellings of variables in your published code, have you ever run your published code?

The download link from the synthia website is not working,is there any other way to download it?

Hi, I'm interested with Self training + Active Learning Concept and want to reproduce the results as paper suggested

However, as I saw the log of 2.2%, 5% that you sent, I am confused about the iterative loop.

As far as I understand, 2.2% and 5% is 2nd iteration and the pretrained weight should be from 1% model's final output.
However, the log is telling that pretrained model is from Imagenet.

If I'm wrong, please let me know.

Sincerely, Joo Young Jang

Following result is 5% starting log.

set random seed to 1
[Info] Load ImageNet pretrain from '/media/dell/Elements/DATA/core/models/resnet101.pth'
missing_keys: []
unexpected_keys: ['fc.weight', 'fc.bias']
[Info] Load ImageNet pretrain from '/media/dell/Elements/DATA/core/models/resnet101.pth'
missing_keys: []
unexpected_keys: ['fc.weight', 'fc.bias']
[Info] Load ImageNet pretrain from '/media/dell/Elements/DATA/core/models/resnet101.pth'
missing_keys: []
unexpected_keys: ['fc.weight', 'fc.bias']
[Info] Load ImageNet pretrain from '/media/dell/Elements/DATA/core/models/resnet101.pth'
missing_keys: []
unexpected_keys: ['fc.weight', 'fc.bias']
[2022-10-25 20:21:07,206][ base.py][line: 41][ INFO] # samples: 2150
[2022-10-25 20:21:07,208][ base.py][line: 41][ INFO] # samples: 2150
labeled: 4825
labeled: 4825
[2022-10-25 20:21:07,223][ base.py][line: 41][ INFO] # samples: 2825
[2022-10-25 20:21:07,226][ base.py][line: 41][ INFO] # samples: 2825
unlabeled: 2825
unlabeled: 2825
[2022-10-25 20:21:07,242][ base.py][line: 41][ INFO] # samples: 500
[2022-10-25 20:21:07,242][ base.py][line: 41][ INFO] # samples: 500
[2022-10-25 20:21:07,242][ builder.py][line: 28][ INFO] Get loader Done...
[2022-10-25 20:21:07,242][ builder.py][line: 28][ INFO] Get loader Done...
No checkpoint found in 'checkpoints/ckpt.pth'
[2022-10-25 20:21:07,255][ lr_helper.py][line: 65][ INFO] The kwargs for lr scheduler: 0.9
[2022-10-25 20:21:07,257][ lr_helper.py][line: 65][ INFO] The kwargs for lr scheduler: 0.9
epoch [ 0 : ] sample_rate_target_class_conf [0.10357965 0.0711445 0.04550922 0.09053792 0.05731867 0.11211205
0.07049027 0.05302454 0.08098789 0.0774475 0.04990353 0.0625474
0.04623231 0.07916455]
epoch [ 0 : ] criterion_per_class tensor([0.0901, 0.8649, 0.4227, 0.1821, 0.7204, 0.9933, 0.4074, 0.9022, 0.9825,
0.0326, 0.3988, 0.7301, 0.9449, 0.5666, 0.6234, 0.9697, 0.8400, 0.9904,
0.5960], device='cuda:0')
epoch [ 0 : ] sample_rate_per_class_conf tensor([0.9406, 0.1396, 0.5967, 0.8455, 0.2890, 0.0069, 0.6126, 0.1011, 0.0181,
1.0000, 0.6215, 0.2790, 0.0570, 0.4481, 0.3893, 0.0314, 0.1654, 0.0099,
0.4176], device='cuda:0')

Which one should I use when I want to inference? Teacher model? or Student model?