mindspore/model_zoo/research/cv/IPT

• Pre-Trained Image Processing Transformer (IPT)
• • Model architecture
  • • The overall network architecture of IPT is shown as below
  • Dataset
  • Requirements
  • • Hardware (Ascend)
    • Framework
    • For more information, please check the resources below
  • Script Description
  • • Scripts and Sample Code
    • Script Parameter
  • Training Process
  • • For pre-training
    • For fine-tuning
  • Evaluation
  • • Evaluation Process
    • Evaluation Result
  • Performance
  • • Inference Performance
  • ModeZoo Homepage

Pre-Trained Image Processing Transformer (IPT)

This repository is an official implementation of the paper "Pre-Trained Image Processing Transformer" from CVPR 2021.

We study the low-level computer vision task (e.g., denoising, super-resolution and deraining) and develop a new pre-trained model, namely, image processing transformer (IPT). To maximally excavate the capability of transformer, we present to utilize the well-known ImageNet benchmark for generating a large amount of corrupted image pairs. The IPT model is trained on these images with multi-heads and multi-tails. In addition, the contrastive learning is introduced for well adapting to different image processing tasks. The pre-trained model can therefore efficiently employed on desired task after fine-tuning. With only one pre-trained model, IPT outperforms the current state-of-the-art methods on various low-level benchmarks.

If you find our work useful in your research or publication, please cite our work:
[1] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. "Pre-trained image processing transformer". CVPR 2021. [arXiv]

@inproceedings{chen2020pre,
  title={Pre-trained image processing transformer},
  author={Chen, Hanting and Wang, Yunhe and Guo, Tianyu and Xu, Chang and Deng, Yiping and Liu, Zhenhua and Ma, Siwei and Xu, Chunjing and Xu, Chao and Gao, Wen},
  booktitle={CVPR},
  year={2021}
 }

Model architecture

The overall network architecture of IPT is shown as below

Dataset

The benchmark datasets can be downloaded as follows:

For super-resolution:

Set5,
Set14,
B100,
Urban100.

For denoising:

CBSD68.

For deraining:

Rain100L

The result images are converted into YCbCr color space. The PSNR is evaluated on the Y channel only.

Requirements

Hardware (Ascend)

Prepare hardware environment with Ascend.

Framework

MindSpore

For more information, please check the resources below

MindSpore Tutorials
MindSpore Python API

Script Description

This is the inference script of IPT, you can following steps to finish the test of image processing tasks, like SR, denoise and derain, via the corresponding pretrained models.

Scripts and Sample Code

IPT
├── eval.py # inference entry
├── train.py # pre-training entry
├── train_finetune.py # fine-tuning entry
├── image
│   └── ipt.png # the illustration of IPT network
├── readme.md # Readme
├── scripts
│   ├── run_eval.sh # inference script for all tasks
│   ├── run_distributed.sh # pre-training script for all tasks
│   └── run_finetune_distributed.sh # fine-tuning script for all tasks
└── src
    ├── args.py # options/hyper-parameters of IPT
    ├── data
    │   ├── common.py # common dataset
    │   ├── bicubic.py # scripts for data pre-processing
    │   ├── div2k.py # DIV2K dataset
    │   ├── imagenet.py # Imagenet data for pre-training
    │   └── srdata.py # All dataset
    ├── metrics.py # PSNR calculator
    ├── utils.py # training scripts
    ├── loss.py # contrastive_loss
    └── ipt_model.py # IPT network

Script Parameter

For details about hyperparameters, see src/args.py.

Training Process

For pre-training

python train.py --distribute --imagenet 1 --batch_size 64 --lr 5e-5 --scale 2+3+4+1+1+1 --alltask --react --model vtip --num_queries 6 --chop_new --num_layers 4 --data_train imagenet --dir_data $DATA_PATH --derain --save $SAVE_PATH

Or one can run following script for all tasks.

sh scripts/run_distributed.sh RANK_TABLE_FILE DATA_PATH

For fine-tuning

For SR tasks:

python train_finetune.py --distribute --imagenet 0 --batch_size 64 --lr 2e-5 --scale 2+3+4+1+1+1 --model vtip --num_queries 6 --chop_new --num_layers 4 --task_id $TASK_ID --dir_data $DATA_PATH --pth_path $MODEL --epochs 50

For Denoising tasks:

python train_finetune.py --distribute --imagenet 0 --batch_size 64 --lr 2e-5 --scale 2+3+4+1+1+1 --model vtip --num_queries 6 --chop_new --num_layers 4 --task_id $TASK_ID --dir_data $DATA_PATH --pth_path $MODEL --denoise --sigma $Noise --epochs 50

For deraining tasks:

python train_finetune.py --distribute --imagenet 0 --batch_size 64 --lr 2e-5 --scale 2+3+4+1+1+1 --model vtip --num_queries 6 --chop_new --num_layers 4 --task_id $TASK_ID --dir_data $DATA_PATH --pth_path $MODEL --derain --epochs 50

Or one can run following script for all tasks.

sh scripts/run_finetune_distributed.sh RANK_TABLE_FILE DATA_PATH MODEL TASK_ID

Evaluation

Evaluation Process

Inference example:
For SR x4:

python eval.py --dir_data $DATA_PATH --data_test $DATA_TEST --test_only --ext img --pth_path $MODEL --task_id $TASK_ID --scale $SCALE

Or one can run following script for all tasks.

sh scripts/run_eval.sh DATA_PATH DATA_TEST MODEL TASK_ID

Evaluation Result

The result are evaluated by the value of PSNR (Peak Signal-to-Noise Ratio), and the format is as following.

result: {"Mean psnr of Set5 x4 is 32.68"}

Performance

Inference Performance

The Results on all tasks are listed as below.

Super-resolution results:

Scale	Set5	Set14	B100	Urban100
×2	38.36	34.54	32.50	33.88
×3	34.83	30.96	29.39	29.59
×4	32.68	29.01	27.81	27.24

Denoising results:

noisy level	CBSD68	Urban100
30	32.37	33.82
50	29.94	31.56

Derain results:

Task	Rain100L
Derain	41.98

ModeZoo Homepage

Please check the official homepage.