course/application_example/ict

• ICT
• • Pretrained model
  • • Model trained by MindSpore
    • Model trained by PyTorch
  • Dataset
  • Train
  • • Transformer Training Parameter description
    • Train Transformer Model
    • • Example
      • Output
    • Upsample Training Parameter description
    • Train Upsample Model
    • • Example
      • Output
  • Infer
  • • Output

ICT

Image completion aims to fill in the missing parts of an image with visually realistic and semantically appropriate content, and is widely used in the field of image processing. Convolutional neural networks have made great progress in the field of computer vision due to their powerful texture modeling capabilities. However, convolutional neural networks do not perform well in understanding the global structure. The development of transformers in recent years has proved its ability to model long-term relationships, but the computational complexity of the transformer hinders its application in processing high-resolution images. ICT combines the advantages of these two methods into the image completion task. First, the transformer is used to reconstruct the appearance prior, and the multivariate coherent structure and some rough textures are restored, and then the convolutional neural network is used for texture complementation, which enhances the high-resolution image. The local texture details of the rough prior guided by the rate mask image.

Pretrained model

We tested the model on the imagenet dataset, noting that since the transformer is relatively slow, we only tested a subfolder n02410509 of the validation set, which took about 8 minutes.

Attention, this model relies on VGG19, the weights file of VGG19 has been extracted and available for download.

Model trained by MindSpore

Transofrmer	Upsample	PSNR$\uparrow$	MAE$\downarrow$
ImageNet_best.ckpt	InpaintingModel_gen_best.ckpt	27.330217	0.021267721

Model trained by PyTorch

Transofrmer	Upsample	PSNR$\uparrow$	MAE$\downarrow$
ImageNet.ckpt	InpaintingModel_gen.ckpt	26.976389	0.02261999

Dataset

At first, you should download dataset by yourself. ImageNet dataset is supported.

After you get the dataset, make sure your path is as following:

# ImageNet dataset
.imagenet/
    └── train
    |    ├── n04347754
    |    |      ├── 000001.jpg
    |    |      ├── 000002.jpg
    |    |      └── ....
    |    └── n04347756
    |           ├── 000001.jpg
    |           ├── 000002.jpg
    |           └── ....
    └── val
         |      ├── n04347754
         |      ├── 000001.jpg
         |      ├── 000002.jpg
         |      └── ....
         └── n04347756
                ├── 000001.jpg
                ├── 000002.jpg
                └── ....

In the image completion task, we also need a mask dataset to mask the image to obtain images with damaged pixels. The mask_dataset can be downloaded by itself. After decompression, the file has the following directory structure:

# Mask dataset
.mask/
├── testing_mask_dataset/
   ├── 000001.png
   ├── 000002.png
   ├── 000003.png
   └── ....

After downloading the weights file and dataset, your folder directory structure should look like this:

.ict/
├── ckpts_ICT/                      # The weight checkpoint folder
│  ├── ms_train/
│  ├── origin/
│  └── VGG19.ckpt
├── mask/                           # Mask dataset folder
│  ├── testing_mask_dataset/
│  │   ├── 00000.png  
│  │   ├── 00001.png
│  │   ├── 00002.png
│  │   └── ....
├── Guided_Upsample/                # Second stage Upsample
├── Transformer/                    # Fitst stage Transformer
├── images/                         # Folder for displaying pictures
├── run.py                          # One stage infer or eval
├── ict.ipynb                       # Executable case file
├── kmeans_centers.npy              # Cluster center dependency file
└── readme.md

Train

Transformer Training Parameter description

Parameter	Default	Description
data_path		Indicate where is the training set
mask_path		Indicate where is the mask
ckpt_path		The path of resume ckpt
device_id	0	Device id
device_target	GPU	Device type
save_path	./checkpoint	Save checkpoints path
batch_size	2	The number of train batch size
train_epoch	5	How many epochs
random_stroke	False	Use the generated mask
use_ImageFolder	False	Using the original folder for ImageNet dataset
prior_size	32	Input sequence length = prior_size * prior_size
learning_rate	3e-4	Value of learning rate
beta1	0.9	Value of beta1
beta2	0.95	Value of beta2

Train Transformer Model

Before starting to train the model, please get the dataset path data_path and mask dataset path mask_path.

When you change the dataset path, you have to change data_path, mask_path the two parameter.

Attention, if you want to modify the path, use absolute paths to reduce unnecessary errors.

Run the train.py to start to train the model. With the ckpt_path parameter, you can resume training from an existing model.

Example

python train.py --mask_path '../mask/testing_mask_dataset' --data_path '/data0/imagenet2012/train' --use_ImageFolder

Output

The following is a partial display of the training output

Epoch: [0 / 5], step: [32 / 1281167], loss: 2.3844809532165527
Epoch: [0 / 5], step: [34 / 1281167], loss: 2.360657215118408
Epoch: [0 / 5], step: [36 / 1281167], loss: 2.3280274868011475
......
Epoch: [0 / 5], step: [75372 / 1281167], loss: 1.682691216468811
Epoch: [0 / 5], step: [75374 / 1281167], loss: 1.682668685913086
......

Upsample Training Parameter description

Parameter	Default	Description
input		Indicate where is the training set
mask		Path to the kmeans
ckpt_path		The path of resume ckpt
device_id	0	Device id
device_target	GPU	Device type
save_path	./checkpoint	Save checkpoints path
kmeans	./kmeans_centers.npy	Path to the VGG
vgg_path	./VGG19.ckpt	Indicate where is the kmeans center
image_size	256	The size of origin image
prior_size	32	The size of prior image from transformer
prior_random_degree	1	During training, how far deviate from
use_degradation_2	False	Use the new degradation function
mode	1	1 is train or 2 is test
mask_type	2	The type of mask
max_iteration	25000	How many run iteration
batch_size	32	The number of train batch size
D2G_lr	0.1	Value of discriminator/generator learning rate ratio
lr	0.0001	Value of learning rate
beta1	0.9	Value of beta1
beta2	0.9	Value of beta2

Train Upsample Model

Before starting to train the model, please get the dataset path input and mask dataset path mask.

When you change the dataset path, you have to change input, mask the two parameter.

Attention, if you want to modify the path, use absolute paths to reduce unnecessary errors.

Run the train.py to start to train the model. With the ckpt_path parameter, you can resume training from an existing model.

Example

python train.py --input '/data0/imagenet2012/train' --mask '../mask/testing_mask_dataset'

Output

The following is a partial display of the training output

Epoch: [1], step: [0 / 40037], psnr: 15.069424, mae: 0.1978589
Epoch: [1], step: [100 / 40037], psnr: 18.389233, mae: 0.13529176
Epoch: [1], step: [200 / 40037], psnr: 19.784353, mae: 0.1134068
......
Epoch: [1], step: [24800 / 40037], psnr: 26.257063, mae: 0.038081832
Epoch: [1], step: [24900 / 40037], psnr: 26.259693, mae: 0.038054496

Infer

After training, you can use testset image to test your model.

Put your image in the folder, then run Transformer/infer.py to generate image priors.

Attention, ckpt_path is the path of the trained transformer model. As before, it is recommended to use absolute paths for all paths, including image paths and mask paths.

python infer.py --ckpt_path '../ckpts_ICT/origin/Transformer/ImageNet.ckpt' --image_url '../input' --mask_url '../mask/testing_mask_dataset' --GELU_2 --save_url '../save'

Then we run Guided_Upsample/infer.py to combine the image prior information to restore the image to its original resolution.

Attention, the parameter prior must be the same as the save_url in the above command.

python infer.py --ckpt_path '../ckpts_ICT/origin/Upsample/ImageNet/InpaintingModel_gen.ckpt' --input '../input' --mask '../mask/testing_mask_dataset' --prior '../save' --save_path '../save'

Output

PSNR: 25.320496, MAE: 0.022787869

In addition, we also provide run.py to complete inference in one stage.

Attention, since the directory is switched in the program, please use an absolute path, using a relative path may cause run error.

python run.py --transformer_ckpt '../ckpts_ICT/origin/Transformer/ImageNet.ckpt' --upsample_ckpt '../ckpts_ICT/origin/Upsample/ImageNet/InpaintingModel_gen.ckpt' --input_image '../input' --input_mask '../mask/testing_mask_dataset' --save_place '../save' --ImageNet

Visualize Results

The following picture is the processed picture of the inference result, the first picture is the real picture, the second picture is the damaged picture processed by the mask, and the third picture is the output picture of the model.