Release code.

.gitignore

@@ -0,0 +1,4 @@
+.DS_Store
+__pycache__
+output
+*.pth

README.md

@@ -0,0 +1,71 @@
+# Deep Fusion Network for Image completion
+
+## Introduction
+
+Deep image completion usually fails to harmonically blend the restored image into existing content,
+especially in the boundary area.
+Our method handles with this problem from a new perspective of
+creating a smooth transition and proposes a concise Deep Fusion Network (DFNet).
+Firstly, a fusion block is introduced to generate a flexible alpha composition map
+for combining known and unknown regions.
+The fusion block not only provides a smooth fusion between restored and existing content,
+but also provides an attention map to make network focus more on the unknown pixels.
+In this way, it builds a bridge for structural and texture information,
+so that information can be naturally propagated from known region into completion.
+Furthermore, fusion blocks are embedded into several decoder layers of the network.
+Accompanied by the adjustable loss constraints on each layer, more accurate structure information are achieved.
+The results show the superior performance of DFNet,
+especially in the aspects of harmonious texture transition, texture detail and semantic structural consistency.
+More detail can be found in our [paper](https://arxiv.org/abs/1904.08060)
+
+![](imgs/github_teaser.jpg)
+
+If you find this code useful for your research, please cite:
+
+```
+@inproceedings{xin2019dfnet,
+  title={Deep Fusion Network for Image Completion},
+  author={Xin Hong and Pengfei Xiong and Renhe Ji and Haoqiang Fan},
+  journal={arXiv preprint},
+  year={2019},
+}
+```
+
+## Prerequisites
+
+- Python 3
+- PyTorch 1.0
+- OpenCV
+
+## Testing
+
+Clone this repo:
+
+``` py
+git clone https://github.com/hughplay/DFNet.git
+cd DFNet
+```
+
+Download pre-trained model from [Google Drive](https://drive.google.com/drive/folders/1lKJg__prvJTOdgmg9ZDF9II8B1C3YSkN?usp=sharing)
+and put them into `model`.
+
+### Testing with Places2 model
+
+There are already some sample images in the `samples/places2` folder.
+
+``` sh
+python test.py --model model/model_places2.pth --img samples/places2/img --mask samples/places2/mask --output output/places2 --merge
+```
+
+### Testing with CelebA model
+
+There are already some sample images in the `samples/celeba` folder.
+
+``` sh
+python test.py --model model/model_celeba.pth --img samples/celeba/img --mask samples/celeba/mask --output output/celeba --merge
+```
+
+## License
+
+<a rel="license" href="http://creativecommons.org/licenses/by-nc/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-nc/4.0/88x31.png" /></a><br />This work is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-nc/4.0/">Creative Commons Attribution-NonCommercial 4.0 International License</a>.
+

config.yaml

@@ -0,0 +1,133 @@
+---
+# 0. Datasets
+# You can define multiple datasets and use them later.
+places2_inpaint:
+  train_path: "/data/datasets/places2/places_train.info"
+  val_path: "/data/datasets/places2/places_valid.info"
+  mask_path: "/data/datasets/places2/mask_strokes_and_shapes.info"
+---
+# 1. Experiment common used information
+tag: 'inpaint.dfn.63'
+task: 'inpaint'
+description: 'Use conv2d, compare to pconv.'
+update:
+date: '2019-04-04'
+
+# Support attr: ['batch_size', 'dataset', 'date', 'device', 'dpflow_device', 'input_size', 'split', 'tag', 'task', 'username']
+# You can define extra attributes in 'extra'
+exp_id_format: [task, model, input_size, date, tag]
+dpflow_format: [username, dataset, split, input_size, batch_size, dpflow_device]
+
+extra:  # Will be used to generate `exp_id`
+  model: 'Conv.StyleLoss'
+
+computed: true           # whether auto computed variables are computed.
+
+exp_id: 'inpaint.Conv.StyleLoss.512x512.2019-04-04.inpaint.dfn.63'                                   # auto computed, depending on train settings.
+
+experiment_root: '/data/train_log/hongxin/inpaint'
+rrun_root: '/data/train_log/hongxin/rrun'
+
+model_dir: '/data/train_log/hongxin/inpaint/conv/dfn/places2/best/num/inpaint.Conv.StyleLoss.512x512.2019-04-04.inpaint.dfn.63/models'                                                                 # auto computed, experiment_root/exp_id/models
+tensorboard_dir: '/data/train_log/hongxin/inpaint/conv/dfn/places2/best/num/inpaint.Conv.StyleLoss.512x512.2019-04-04.inpaint.dfn.63/tb'                                                                 # auto computed, experiment_root/exp_id/tensorboard
+log_dir: '/data/train_log/hongxin/inpaint/conv/dfn/places2/best/num/inpaint.Conv.StyleLoss.512x512.2019-04-04.inpaint.dfn.63/log'                                                                 # auto computed, experiment_root/exp_id/log
+result_dir: '/data/train_log/hongxin/inpaint/conv/dfn/places2/best/num/inpaint.Conv.StyleLoss.512x512.2019-04-04.inpaint.dfn.63/results'                                                                 # auto computed, experiment_root/exp_id/results
+
+model_latest: 'latest.pth'
+model_best: 'best.pth'
+
+data_script: 'start_dpflow.py'
+---
+# 2. Training settings
+continue: true
+remove_old: false
+
+# Data settings
+dataset: 'places2_inpaint'
+split: 'train'
+input_size: [512, 512] # scalar or list [height, width]
+batch_size: 6 # per GPU
+seed: 2019
+dpflow_base_name: 'hongxin.places2_inpaint.train.512x512.6.8x4'      # auto computed, based on `dpflow_format`
+dpflow_replicas: 8 # Generally, same as total number of gpus
+worker_per_dpflow: 4
+
+# Trainer setting
+device:
+- {num: 1, gpu: 8, cpu: 16, memory: 51200}  # Memory in MiB
+
+model:
+  c_img: 3
+  c_mask: 1
+  c_alpha: 3
+  mode: 'nearest'
+  norm: 'batch'
+  act_en: 'relu'
+  act_de: 'leaky_relu'
+  en_ksize: [7, 5, 5, 3, 3, 3, 3, 3]
+  de_ksize: [3, 3, 3, 3, 3, 3, 3, 3]
+  blend_layers: [0, 1, 2, 3, 4, 5]
+
+optimizer:
+  name: 'Adam'
+  args:
+    lr: 0.0002
+
+epoch: 20
+iter_per_epoch: 37500
+lr_decay_epoch: 5
+lr_decay_ratio: 0.1
+
+loss:
+  c_img: 3
+  w_l1: 6.
+  w_percep: 0.1
+  w_style: 240.
+  w_tv: 0.1
+  structure_layers: [0, 1, 2, 3, 4, 5]
+  texture_layers: [0, 1, 2]
+
+log_level: 'INFO'
+action:
+  save_model: true
+  validate: false
+  tensorboard: true
+  model_graph: false # depend on tensorboard
+
+log_interval: 10
+model_save_interval: 1000 # (iters)
+validate_interval: 1000
+validate:
+  dataset: 'places2_inpaint'
+  split: 'val'
+  num: 20
+  input_size: [512, 512]
+  batch_size: 1 # To keep the original size of images
+---
+# 3. Testing settings
+# Auto compute: the following parameter will be compute automatically
+device:
+- {num: 1, gpu: 1, cpu: 4, memory: 10240}  # Memory in MiB
+
+model: 'best'
+data_tag: 'places2-1000'
+
+img: /data/train_log/hongxin/inpaint/data/sample512-1000/img
+mask: /data/train_log/hongxin/inpaint/data/sample512-1000/mask
+input_size: [512, 512]
+batch_size: 16
+
+action:
+  save: ['final']
+  metrics:
+
+dataset: 'places2'
+split: 'val'
+seed: 2019
+dpflow_base_name: 'hongxin.places2.val.512x512.16.1x2'             # auto computed
+dpflow_replicas: 1 # Generally, same as number of gpus
+worker_per_dpflow: 2
+---
+# 4. Training & Testing record
+train: []
+test: []