ewrfcas / face-super-resolution Goto Github PK

Hi, what a amazing work! In my application in some low resolution of Asian face, I found that sometimes the eye appears green, I think this is the cause of the dataset, I want to in your training, on the basis of the application of hd Asia face dataset to training again, but I did not have a discriminator training parameters refer to "90000_G.pth", can give me this file? Thank you very much! (PS: I only have one 1080 ti, retraining costs too much)

Hi, I am realizing this model and I am wondering what data set did you use in this model? Do you have training details? Thank you very much!

Hello, when i check your code, i found you clip face area and do SR for face area then merge to larger images. do you try inference based on a image which include face directly? about training, you just training ESRGAN using face dataset?

Hi, your result is very impressive but I can not reproduce it. Is it possible that you can clarify the training details more specifically?

From the paper, there are two steps to train. One is to train without GAN and the other is to train with GAN. What are your two command line for these two steps and how long do you train for the first step?(The command line includes a lot of parameters like your lr and D_update_ratio and so on) Thank you very much.

I see that when we run test.py, we get two images as output - One where there is no shift in position and another focussed on the face. Why does the second one have more quality than the first?

Hello,

First of all, thank you for your work. There is an issue I should consult with you. I want to do a training to fit animated characters using trained weights, but I guess I need a .state file for that. Is it possible for me to access this file?

Thank you in advance, I wish you a good day.

.

could you upload pre-trained weights in google drive?

As I understood from train.py file, to continue training, I need to pass resume_state parameter. How can i get it or what value should it have to correctly continue training with existing weights?

I'm attempting to train on the full FFHQ 1024 set and am receiving an error despite setting hr_size=1024

RuntimeError: The size of tensor a (512) must match the size of tensor b (1024) at non-singleton dimension 3.

Does anyone know where in the model I need to adjust the numbers to allow for increased resolution?

On running the test.py, code is throwing an error saying No module named 'models.modules'. There is no modules file in the modules folder.

Hi, I'm wondering what dataset did you use to train & obtain weights for 90000_G.pth ?

Thanks

Your result is quite impressive. Do you also have 90000_D.pth uploaded? It would be really nice of you if you can upload it. Thank you.

What kind of GPU did you use for training and how long did you train for?
When I tried to train, I set LR 128, HR 512, but it showed insufficient memory. Any suggestions for improvement?

Hello author, when I use your code to train and test HR512512, LR128128, the code is all normal, but when the data set is changed to HR256256, LR6464, the above error occurs. May I ask How to solve this problem? I have changed the size of LR in the code to 64.

Hi, I saw in the paper they interpolate the PSNR model and GAN model with alpha. Did you also interpolate them or you just train the model using your train.py?

hi @ewrfcas there is NO error if I decrease the image size in train.py
If I change lr_height, lr_width and hr_height, hr_width in data_loader.py to my input size it's giving me the error.
But it's working fine with default sizes. what I am missing here

the error is

ret = torch.addmm(bias, input, weight.t())
RuntimeError: mat1 dim 1 must match mat2 dim 0

Hi!

Where is the requirements.txt file?

what dataset did you use in training this model!!

Hi author,

I am new to this and ur work are very interesting. May i check if there are any other models that i can use? or how can i train a new model based on new dataset that i have? Pleaseee help ty...

Thanks for your sharing, your work is quite interesting. and i wonder whether there are some relative papers to read?

Python version 3.6
Pytorch version 1.4

I downloaded the pre-trained models and running test.py

python test.py

But getting the below error

Traceback (most recent call last):
  File "test.py", line 84, in <module>
    output_img, rec_img = sr_forward(img)
  File "test.py", line 73, in sr_forward
    img_aligned, M = dlib_detect_face(img, padding=padding, image_size=(128, 128), moving=moving)
  File "/home/ec2-user/mydir/Face-Super-Resolution/dlib_alignment.py", line 68, in dlib_detect_face
    dets = dlib_detector(img, 0)
TypeError: __call__(): incompatible function arguments. The following argument types are supported:
    1. (self: _dlib_pybind11.fhog_object_detector, image: array, upsample_num_times: int=0) -> _dlib_pybind11.rectangles

Invoked with: <_dlib_pybind11.fhog_object_detector object at 0x7fd51a70d5e0>, None, 0

Please suggest.

When I try to run test.py it says:
ImportError: No module named SRGAN_model
what is the package that needs to be installed?

I get this error:

Traceback (most recent call last):
File "test.py", line 69, in
sr_model = SRGANModel(get_FaceSR_opt(), is_train=False)
File "Face-Super-Resolution/models/SRGAN_model.py", line 17, in init
super(SRGANModel, self).init(opt, is_train)
TypeError: super() argument 1 must be type, not classobj