trendingtechnology / dmfn Goto Github PK

DMFN (Dense Multi-scale Fusion Network)

This is an unoffical repository for reproducing model DMFN from the paper [Image Fine-grained Inpainting]. The original repository is here, but author have not commit the rest of implement code yet.

Prerequisites

• Python3.5 (or higher)
• pytorch 1.0(or higher) with GPU
• numpy
• OpenCV
• scipy
• tensorboardX

RESULT

Note that the following result maybe not as good as the paper because they are trained only in 1 epoch. You can get the final result in original author's github.

train

test

loss

Prepair the dataset

Download the dataset of celebA, unzip and split it to test/train dataset (or you can use my train/test file in CelebA/ ).

How to test

You can specify the folder address by the option --dataset_path, and set the pretrained model path by --load_model_dir when calling test.py as the following

python test.py ---dataset_path celeba_data --data_file img_align_celeba_png\test.txt --load_model_dir pretrained/1epoch

I train it only 1 epoch with single GPU, you can train it yourself for better performance or in custom dataset.

How to train

Use train.py as the following

python train.py ---dataset_path celeba_data --data_file img_align_celeba_png\test.txt --batch_size 8 --lr 2e-4

You can load the pretrained model by the option --load_model_dir too.

TODO

• 中文博客(https://blog.csdn.net/h8832077/article/details/105166776)
• upload pretrained model you can download pretrained model here (baidu netdisk: ozpo)