interdigitalinc / hrfae

Official implementation for paper High Resolution Face Age Editing

HRFAE: High Resolution Face Age Editing

Official implementation for paper High Resolution Face Age Editing.

Dependencies

• Python 3.7
• Pytorch 1.1
• Numpy
• Opencv
• TensorboardX
• Tensorboard_logger

You can also create a new environment for this repo by running

conda env create -f env.yml

Load and test pretrained network

1. You can download the pretrained model by running:

   cd ./logs/001
   ./download.sh
   

2. Upload test images in the folder /test/input and run the test file. The output images will be saved in the folder /test/output. You can change the desired target age with --target_age.

   python test.py --config 001 --target_age 65
   

Train a new model

1. Pretrained age classifier

   To get age information, we use an age classifier pretrained on IMDB-WIKI dataset. We use the model released from paper Deep expectation of real and apparent age from a single image without facial landmarks by Rothe et al.

   To prepare the model, you need to download the original caffe model and convert it to PyTorch format. We use the converter caffemodel2pytorch released by Vadim Kantorov. Then name the PyTorch model as dex_imdb_wiki.caffemodel.pt and put it in the folder /models.

2. Preparing your dataset

   Download FFHQ dataset and unzip it to the /data/ffhq directory.

   Download age label to the /data directory.

   You can also train the model with your own dataset. Put your images in the /data directory. With the pretrained classifier, you can create a new label file with the age of each image.

3. Training

   You can modify the training options of the config file in configs directory.

   python train.py --config 001 
   

Google Colab

We also provide a colab version for quick test. To run it using Google Colab, please click here.

Citation

@article{yao2020high,
    title   = {High Resolution Face Age Editing},
    author  = {Xu Yao and Gilles Puy and Alasdair Newson and Yann Gousseau and Pierre Hellier},
    journal = {CoRR},
    volume  = {abs/2005.04410},
    year    = {2020},
}

License

Copyright © 2020, InterDigital R&D France. All rights reserved.

This source code is made available under the license found in the LICENSE.txt in the root directory of this source tree.

hrfae's Issues

CELoss may be wrong?

Hello:
I find some code that I think may be wrong
return nn.CrossEntropyLoss()(x, target_age)
https://github.com/InterDigitalInc/HRFAE/blob/master/trainer.py#L62

which should be
return nn.CrossEntropyLoss(x, target_age)
?

I don't know if it has some impact.

the test model looks like based on cuda device

Hi bor:
when I try to run test.py, and when the code go into line 58:
trainer.load_checkpoint(log_dir + 'checkpoint')
then an error happend at here.
'''
raise RuntimeError('Attempting to deserialize object on a CUDA '
RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.
'''
this looks like means I must run these code on a cuda device.
or because of this test model is trained on a cuda device, I must build a new model which supported on CPU only?

Not proper result when reducing size to 256x256

Hello there,

i have issue in smaller size ... changed width height in config to 256... and changed in image file also...

and getting output not proper ... can any one help me please ...

Originally posted by @kishanpanchotiya in #8 (comment)

dex_imdb_wiki.caffemodel.pt

Hi, I would like to train the model, but I had the problem when convert the caffe model to pt file. Is that possible to provide the dex_imdb_wiki.caffemodel.pt file for me?

how train in different resolution?

@Xu-Yao Thanks you very much for the code you shared.Could you please provide the code train for other resolution pictures and your dataset about 50,000 face images you made.

CelebA-HQ age label

Does anyone have the age tag for Celeba-HQ?

About age-equalized dataset

Hi @Xu-Yao
Thanks for sharing this code.

I checked the paper, and found that you are solving data imbalance using StyleGAN generated images (Section 4.1).
But unfortunately, not in the paper or here, there is no information on how to access this dataset (age-equalized).

If it is possible, could you share this dataset, please.

how to create a new age label?

Could you share the pretrained ageclassifer model?

It doesn't work very well

how much time

How much time do you spend on 70,000 pictures while training?

run download.sh failed

Hello,
looks like download weight URL died.
it returns HTTP ERROR 500 and get a 0 byte checkpoint file.
could you please fix it?
Thanks.

Pretrained model checkpoint backup

As I do not have access to Telecom resources anymore, the download link for the pre-trained model may soon be disabled. You can download the pre-trained checkpoint here: https://drive.google.com/file/d/11wK7cHOH--CWgNSbZucJ4nppon4BN9eJ/view?usp=share_link

Problem converting dex_imdb_wiki.caffemodel to .pt

Hi, I had the problem when convert the caffe model to pt file. Problem is" FileNotFoundError: [Errno 2] No such file or directory: 'protoc': 'protoc' "
Is that possible to provide the dex_imdb_wiki.caffemodel.pt file for me ?
Thanks.

dex_imdb_wiki.caffemodel.pt

could you please provide the dex_imdb_wiki.caffemodel.pt?

how to get faces of children

Thanks for your work very much.
there is a question that: when I test the pics prepared, i set '--target_age 5' , it can not generate faces of chilren well ?

实测效果不好

I raised a stupid question and solved it now，please delete it

Running time in the cpu environment

Hello, may I ask you why the encode structure takes a long time, about 33s

The results is not really good.Is it any pro?

I have try run the test.py and download the pretrained model. but the result is bad, so I'm confused any steps is wrong?

Docker image

Hi, will you provide a working docker image?

How to convert to ONNX model?

Do anyone try to convert trained model to ONNX model, please?
Don't know how to use torch.export because there're multiple networks.
Hope can get help here, thanks.

ResolvePackageNotFound: - openssl==1.1.1=h7b6447c_0

While trying to create a virtual environment using the steps mentioned in README.md , I faced this issue:

ResolvePackageNotFound: - openssl==1.1.1=h7b6447c_0

Did anyone face a similar issue?

Can't we change input data in configs file?

We have:

# Input data
input_w: 1024
input_h: 1024

in configs/001.yml file. Can't we change it to other pixel formats, for example, 3780x6720 or retaining the original size of images? Please feel free to throw some light on why it can't accept the image as is.

I tried changing it to given example but the program was immediately killed.

It gives weird outputs when the size is changed to 1024x1024. Check the images. The first one is original, second one is the output by the model.

no module named 'torch'

I have verified that I have installed pytorch, when I type conda list I see the following:

pytorch 1.5.0 py3.7_cuda102_cudnn7_0 pytorch

however, when I try to run this code after following your instructions and setting up an environment I receive the following error:

Traceback (most recent call last):
File "test.py", line 11, in
import torch
ModuleNotFoundError: No module named 'torch'

Any help with this issue would be greatly appreciated.