deepfillv2

The PyTorch implementations and guideline for Gated Convolution based on ICCV 2019 oral paper: free-form inpainting (deepfillv2).

We are focusing on Gated Conv so do not implement original paper completely, and implement it as a coarse-to-fine manner.

1 Implementations

Before running it, please ensure the environment is Python 3.6 and PyTorch 1.0.1.

1.1 Train

If you train it from scratch, please specify following hyper-parameters. For other parameters, we recommend the default settings.

python train.py     --epochs 40
                    --lr_g 0.0001
                    --batch_size 4
                    --perceptual_param 10
                    --gan_param 0.01
                    --baseroot [the path of training set, like Place365]
                    --mask_type 'free_form' [or 'single_bbox' or 'bbox']
                    --imgsize 256

if you have more than one GPU, please change following codes:
python train.py     --multi_gpu True
                    --gpu_ids [the ids of your multi-GPUs]

1.2 Test

At testing phase, please download the pre-trained model first. (including deepfillv2 network for ~~RGB images~~ and grayscale images)

For small image patches, make sure that all the dataset settings are the same as training part:

python test.py 	    --load_name '*.pth' (please ensure the pre-trained model is in same path)
                    --baseroot [the path of testing set]
                    --mask_type 'free_form' [or 'single_bbox' or 'bbox']
                    --imgsize 256

There are some examples:

The corresponding ground truth is:

1.3 PSNR experiment on 15 images

item	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
mask region	13.23	18.23	17.06	13.16	19.41	11.47	29.58	15.51	14.71	25.99	20.54	19.21	15.86	11.79	10.73
full image	26.50	45.01	32.35	29.59	31.65	24.57	48.44	30.27	32.24	51.18	35.15	36.75	30.56	27.21	29.13

2 Acknowledgement