cvlab-yonsei / pisr Goto Github PK
View Code? Open in Web Editor NEWAn official implementation of "Learning with Privileged Information for Efficient Image Super-Resolution" (ECCV2020) in PyTorch.
An official implementation of "Learning with Privileged Information for Efficient Image Super-Resolution" (ECCV2020) in PyTorch.
Hi,
I am interested to know whether this compression method could be applied to very deep cnn like EDVR?
Thx,
Lei
In step2_train_student.py, you only set the teacher_model into eval type, I wonder whether the teacher model is trained while loss.backward(). Since the eval() only freeze the BN layer and Dropout layer, but the require_grad is still True.
Hi, Thank you for this great work.
I am a bit confused with Teacher mode. why the Teacher model needs LR images.
Hi,
I note that there are codes named visualizers, how can I use the visualizers to visualize my test results?
Hope for your reply!
can you tell me how to test the model
Hi
Thanks for your code.
I have another question about loss function.
loss in the paper is written as :
but I see the loss in the code as :
I wonder if these two equations are the same.
"torch.log(2std) + numerator / (std)" this part is almost the same with one in the paper
but why "mu.shape[1] * np.log(2math.pi)/2" term exists?
sorry for many questions but It'll save a lot of my time if you answer me ;)
thanks in advance.
Hi, thanks for providing a code and wonderful research.
This model upscale 2, 3, 4 times. Can we set this model to 8 times scale up?
Traceback (most recent call last):
File "step1_train_teacher.py", line 242, in
main()
File "step1_train_teacher.py", line 236, in main
run(config)
File "step1_train_teacher.py", line 204, in run
writer, visualizer, last_epoch+1)
File "step1_train_teacher.py", line 153, in train
eval_type='val')
File "step1_train_teacher.py", line 123, in evaluate_single_epoch
avg_loss = total_loss / (i+1)
UnboundLocalError: local variable 'i' referenced before assignment
Hello!
I have read your paper. You mentioned in the article that the low-resolution image learned from high-resolution image contains more high-frequency signals, which aroused my interest. If you can provide test code, I would be very grateful.
By the way, this is a awesome project!
when I train the student model and run 'python step2_train_student.py --config configs/fsrcnn/step2.yml', it returns the error that RuntimeError: CUDA out of memory. Tried to allocate 1.57 GiB (GPU 0; 10.76 GiB total capacity; 4.81 GiB already allocated; 1.21 GiB free; 8.77 GiB reserved in total by PyTorch). I reduce the batchsize to 4 in step2.yaml and use 4 gtx1080 Ti. The error still exists. Could you give me some suggestions?
hello, i have an issue is about the compared model of FSRCNN(origin), and your model of FSRCNN training by PISR.
the origin model of fsrcnn is trained without div2k, but your model has.
so, maybe your method is no good as your paper shows, because if the origin model of fsrcnn is trained with the same trainset like yours, it will better than before, right?
if i am wrong, could you tell me why, please?
thanks you very much!
Hi
Thanks for your code.
I'm reading the code with your paper and I saw that you stated initializing a student network with parameters from
teacher network helped training faster.
But I couldn't see where the code is.
May be it's because I'm not so familiar with pytorch.
It'll very helpful if you tell me where the code initializing student network parameter is.
hi~, the url of pretrained_model does not exist, could you release again,thanks
作者您好?为啥您的模型输入要求图片维度为1呢?为什么不直接输入3维的彩色图像呢?仅输入Y通道是不是为了能够取得更好的PSNR、SSIM值?那直接输入3维度图像效果如何呢?
Hello!
First of all, thank you for being able to open source your project, this project is great.
When I was reproducing your code, I did not find the test code that can get the subjective visual results. May I ask where to set the display or save the subjective visual results after inference in the project?
Looking forward to your reply!
Thank you for your brilliant work and sharing this code!
After I construct the 'data' folder as required, I encounter an unexpected error during training:
Traceback (most recent call last):
File "step2_train_student.py", line 274, in
main()
File "step2_train_student.py", line 268, in main
run(config)
File "step2_train_student.py", line 229, in run
dataloaders = {'train':get_train_dataloader(config, get_transform(config)),
File "/home/lyh/PISR/datasets/dataset_factory.py", line 33, in get_train_dataloader
dataset = get_train_dataset(config, transform)
File "/home/lyh/PISR/datasets/dataset_factory.py", line 23, in get_train_dataset
name=name, train=True, transform=transform,
File "/home/lyh/PISR/datasets/dataset.py", line 20, in init
begin, end)
File "/home/lyh/PISR/datasets/base_dataset.py", line 107, in init
self._init_repeat()
File "/home/lyh/PISR/datasets/base_dataset.py", line 143, in _init_repeat
assert n_images != 0
AssertionError
How can I fix this? Thank you so much for any possible solutions!
Hi.
I used my own training set with 1000 data.
However, when I print the total_size in train_single_epoch function it becomes 16000.
Where does the extra data come from?
Thanks.
Thank you for share your work,Thank you very much! But I only find fsrcnn's model, I want know how to train my own model?use the same encoder? Thank you very much.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.