• Contents
• SRGAN Description
• Model Architecture
• Dataset
• Pretrained model
• 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
    • Evaluation Performance
• ModelZoo Homepage

Contents

• SRGAN Description
• Model Architecture
• Dataset
• Pretrained model
• 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

SRGAN Description

Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of optimization-based super-resolution methods is principally driven by the choice of the objective function.Recent work has largely focused on minimizing the mean squared reconstruction error. The resulting estimates have high peak signal-to-noise ratios, but they are often lacking high-frequency details and are perceptually unsatisfying in the sense that they fail to match the fidelity expected at the higher resolution. In this paper, we present SRGAN,a generative adversarial network (GAN) for image superresolution (SR). To our knowledge, it is the first framework capable of inferring photo-realistic natural images for 4× upscaling factors. To achieve this, we propose a perceptualloss function which consists of an adversarial loss and a content loss. The adversarial loss pushes our solution to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved images and original photo-realistic images. In addition, we use a content loss motivated by perceptual similarity instead of similarity in pixel space. Our deep residual network is able to recover photo-realistic textures from heavily downsampled images on public benchmarks.

Paper: Christian Ledig, Lucas thesis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, Wenzhe Shi
Twitter.

Model Architecture

The SRGAN contains a generation network and a discriminator network.

Dataset

Train SRGAN 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/testdataset.py

Pretrained model

The process of training SRGAN needs a pretrained VGG19 based on Imagenet.

Training scripts|
VGG19 pretrained model

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

SRGAN

├─ README.md                   # descriptions about SRGAN
├── scripts  
 ├─ run_distribute_train.sh                # launch ascend training(8 pcs)
 ├─ run_eval.sh                   # launch ascend eval
 └─ run_stranalone_train.sh             # launch ascend training(1 pcs)
├─ src  
 ├─ ckpt                       # save ckpt  
 ├─ dataset
  ├─ testdataset.py                    # dataset for evaling  
  └─ traindataset.py                   # dataset for training
├─ loss
 ├─  gan_loss.py                      #srgan losses function define
 ├─  Meanshift.py                     #operation for ganloss
 └─  gan_loss.py                      #srresnet losses function define
├─ models
 ├─ dicriminator.py                  # discriminator define  
 ├─ generator.py                     # generator define  
 └─ ops.py                           # part of network  
├─ result                              #result
├─ trainonestep
  ├─ train_gan.py                     #training process for srgan
  ├─ train_psnr.py                    #training process for srresnet
└─ util
 └─ util.py                         # initialization for srgan
├─ test.py                           # generate images
└─train.py                            # train script

Script Parameters

Training Script Parameters

# distributed training
Usage: sh run_distribute_train.sh [DEVICE_NUM] [DISTRIBUTE] [RANK_TABLE_FILE] [LRPATH] [GTPATH] [VGGCKPT] [VLRPATH] [VGTPATH]

# standalone training
Usage: sh run_standalone_train.sh [DEVICE_ID] [LRPATH] [GTPATH] [VGGCKPT] [VLRPATH] [VGTPATH]

Training Result

Training result will be stored in scripts/srgan0/ckpt. You can find checkpoint file.

Evaluation Script Parameters

• Run run_eval.sh for evaluation.

# evaling
sh run_eval.sh [CKPT] [EVALLRPATH] [EVALGTPATH]

Evaluation result

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

Model Description

Performance

Training Performance

Parameters
Model Version	V1
Resource	CentOs 8.2; Ascend 910; CPU 2.60GHz, 192cores; Memory 755G
MindSpore Version	1.2.0
Dataset	DIV2K
Training Parameters	epoch=2000+1000, batch_size = 16
Optimizer	Adam
Loss Function	BCELoss MSELoss VGGLoss
outputs	super-resolution pictures
Accuracy	Set14 psnr 27.03
Speed	1pc(Ascend): 540 ms/step; 8pcs: 1500 ms/step
Total time	8pcs: 8h
Checkpoint for Fine tuning	184M (.ckpt file)
Scripts	srgan script

Evaluation Performance

Parameters	single Ascend
Model Version	V1
Resource	CentOs 8.2; Ascend 910; CPU 2.60GHz, 192cores; Memory 755G
MindSpore Version	1.2.0
Dataset	Set14
batch_size	1
outputs	super-resolution pictures

ModelZoo Homepage

Please check the official homepage.