• Contents
• DBPN Description
• Model Architecture
• Dataset
• Environment Requirements
• Script Description
• • Script and Sample Code
  • Script Parameters
  • • Training Script Parameters
    • Training Result
    • Evaluation Script Parameters
    • Evaluation result
• Model Description
• • Performance
  • • Training Performance
• ModelZoo Homepage

Contents

• DBPN Description
• Model Architecture
• Dataset
• Environment Requirements
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
• Model Description
  • Performance
    • Training Performance
    • Evaluation Performance
• Description of Random Situation
• ModelZoo Homepage

DBPN Description

The feed-forward architectures of recently proposed deep super-resolution networks learn representations of
low-resolution inputs, and the non-linear mapping from those to high-resolution output. However, this approach does not
fully address the mutual dependencies of low- and high-resolution images. We propose Deep Back-Projection Networks (
DBPN), that exploit iterative up- and down-sampling layers, providing an error feedback mechanism for projection errors
at each stage. We construct mutually-connected up- and down-sampling stages each of which represents different types of
image degradation and high-resolution components. We show that extending this idea to allow concatenation of features
across up- and down-sampling stages
(Dense DBPN) allows us to reconstruct further improve super-resolution, yielding superior results and in particular
establishing new state of the art results for large scaling factors such as 8× across multiple data sets.

Paper: Muhammad Haris, Greg Shakhnarovich, Norimichi Ukita

Model Architecture

The DBPN contains a generation network and a discriminator network.

Dataset

Train DBPN Dataset used: DIV2K

• Note: Data will be processed in src/dataset/traindataset.py

Validation and eval evaluationdataset
used: Set5 Set14

• Note:Data will be processed in src/dataset/dataset.py

Environment Requirements

• Hardware Ascend
  • Prepare hardware environment with Ascend processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Script Description

Script and Sample Code

The entire code structure is as following:

.DBPN
├─ README.md                    # descriptions about DBPN
├── scripts  
  └─run_distribute_train.sh     # launch ascend training(8 pcs)
  └─run_stranalone_train.sh     # launch ascend training(1 pcs)
  └─run_eval.sh                 # launch ascend eval
├─ src
  ├─ dataset
    └─ dataset.py                 # dataset for training and evaling
  ├─ loss
    ├─ withlosscell.py            # DBPN Gan loss Cell define
    └─ generatorloss.py           # compute_psnr losses function define
  ├─ model
    ├─ base_network.py             # the basic unit for build neural network
    ├─ dbpns.py                    # generator DBPNS define  T = 2
    ├─ ddbpnl.py                   # generator DDBPNL define  T = 6
    ├─ ddbpn.py                    # generator DDBPN define  T = 7
    ├─ dbpn.py                     # generator DBPN define T = 10
    ├─ dbpn_iterative.py           # generator DBPNiterative define
    ├─ dicriminator.py             # discriminator define  
    └─ generator.py                # get DBPN models
  ├─ trainonestep
    ├─ trainonestepgen.py          # training process for generator such as DBPNS、DDBPNL、DBPN
    └─ trainonestepgenv2.py        # training process for generator DDBPN which add clip grad
  ├─ util
    ├─ config.py                   # parameters using for train and eval
    └─ utils.py                    # initialization for srgan
  └─ vgg19
    └─ define.py                   # this is just 4 maxpool2d operator
├─ train_dbpn.py                   # train dbpn  script
├─ train_dbpngan.py                # train gan network script
├─ eval.py                         # eval
└─ export.py                       # export mindir script

Script Parameters

Training Script Parameters

# distributed training DBPN models
Usage: bash run_distribute_train.sh [DEVICE_NUM] [DISTRIBUTE] [RANK_TABLE_FILE] [TRAIN_GT_PATH] [VAL_GT_PATH] [VAL_LR_PATH] [MODE]

eg: bash run_distribute_train.sh 8 1 ./hccl_8p.json /data/DBPN_data/DIV2K_train_HR /data/DBPN_data/Set5/HR /data/DBPN_data/Set5/LR False

# standalone training DBPN models
Usage: bash run_standalone_train.sh [DEVICE_ID] [MODEL_TYPE] [TRAIN_GT_PATH] [VAL_GT_PATH] [VAL_LR_PATH] [MODE]

eg: bash run_standalone_train.sh 0 DDBPN /data/DBPN_data/DIV2K_train_HR /data/DBPN_data/Set5/HR /data/DBPN_data/Set5/LR False

Training Result

Training result of the model genearated image will be stored in Results.

Training result of the model loss will be stored in result.

You can find checkpoint file in ckpt.

Evaluation Script Parameters

• Run run_eval.sh for evaluation.

#evaling
bash run_eval.sh [DEVICE_ID] [CKPT] [MODEL_TYPE] [VAL_GT_PATH] [VAL_LR_PATH]
eg: bash scripts/run_eval.sh 0 /data/DBPN_data/dbpn_ckpt/gen_ckpt/D-DBPN-best.ckpt DDBPN /data/DBPN_data/Set5/HR /data/DBPN_data/Set5/LR

Evaluation result

Evaluation result will be stored in the result. Under this, you can find generator pictures.

Model Description

Performance

Training Performance

Parameters
Model Version	DBPN
Resource	CentOs 8.2; Ascend 910; CPU 2.60GHz, 192cores; Memory 755G
MindSpore Version	1.3.0
Dataset	DIV2K
Training Parameters	ddbpn:epoch=2000, batch_size = 16; dbpngan:epoch=1000,batch_size=4
Optimizer	Adam
Loss Function	BCELoss MSELoss VGGLoss
outputs	super-resolution pictures
Accuracy	Set5 compute_psnr 31.92(train ddbpn); Set5 compute_psnr 29.21(train dbpngan network)
Speed	1pc(Ascend): 3463 ms/step(ddbpn), 800.41ms/step(dbpngan); 8pcs: 1781ms/step(ddbpn) 818.22ms/step(dbpngan)
Total time	8pcs: 12h43m50s(ddbpn), 13h50m28s(dbpngan)
Checkpoint of DBPN models	58.8M (.ckpt file)
Scripts	DBPN script

ModelZoo Homepage

Please check the official homepage
.