Ugness/PiCANet-Implementation

Pytorch Implementation of PiCANet: Learning Pixel-wise Contextual Attention for Saliency Detection

PiCANet-Implementation

Pytorch Implementation of PiCANet: Learning Pixel-wise Contextual Attention for Saliency Detection

New method on implementing PiCANet

• Issue#9
• Conv3d version is deleted.

• batchsize 1
• batchsize 4

(DUTS) Top 10 Performance Test with F-score (beta-square = 0.3)

batchsize:4

• Issue#9

Step	Value	Threshold	MAE
214000	0.8520	0.6980	0.0504
259000	0.8518	0.6510	0.0512
275000	0.8533	0.6627	0.0536
281000	0.8540	0.7451	0.0515
307000	0.8518	0.8078	0.0523
383000	0.8546	0.6627	0.0532
399000	0.8561	0.7882	0.0523
400000	0.8544	0.7804	0.0512
408000	0.8535	0.5922	0.0550
410000	0.8518	0.7882	0.0507

Execution Guideline

Requirements

Pillow==4.3.0
pytorch==0.4.1
tensorboardX==1.1
torchvision==0.2.1
numpy==1.14.2

My Environment

S/W
Windows 10
CUDA 9.0
cudnn 7.0
python 3.5
H/W
AMD Ryzen 1700
Nvidia gtx 1080ti
32GB RAM

Execution Guide

• For training,
• Please check the Detailed Guideline if you want to know the dataset structure.

    usage: train.py [-h] [--load LOAD] --dataset DATASET [--cuda CUDA]
                    [--batch_size BATCH_SIZE] [--epoch EPOCH] [-lr LEARNING_RATE]
                    [--lr_decay LR_DECAY] [--decay_step DECAY_STEP]
                    [--display_freq DISPLAY_FREQ]
    optional arguments:
      -h, --help            show this help message and exit
      --load LOAD           Directory of pre-trained model, you can download at
                            https://drive.google.com/file/d/109a0hLftRZ5at5hwpteRfO1A6xLzf8Na/view?usp=sharing
                            None --> Do not use pre-trained model. Training will start from random initialized model
      --dataset DATASET     Directory of your Dataset
      --cuda CUDA           'cuda' for cuda, 'cpu' for cpu, default = cuda
      --batch_size BATCH_SIZE
                            batchsize, default = 1
      --epoch EPOCH         # of epochs. default = 20
      -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                            learning_rate. default = 0.001
      --lr_decay LR_DECAY   Learning rate decrease by lr_decay time per decay_step, default = 0.1
      --decay_step DECAY_STEP
                            Learning rate decrease by lr_decay time per decay_step,  default = 7000
      --display_freq DISPLAY_FREQ
                            display_freq to display result image on Tensorboard

• For inference,
• dataset should contain image files only.
• You do not need masks or images folder. If you want to run with PairDataset structure, use argument like
  --dataset [DATAROOT]/images
• You should specify either logdir (for TensorBoard output) or save_dir (for Image file output).
• If you use logdir, you can see the whole images by run tensorboard with --samples_per_plugin images=0 option

    usage: image_test.py [-h] [--model_dir MODEL_DIR] --dataset DATASET
                     [--cuda CUDA] [--batch_size BATCH_SIZE] [--logdir LOGDIR]
                     [--save_dir SAVE_DIR]

    optional arguments:
      -h, --help            show this help message and exit
      --model_dir MODEL_DIR
                            Directory of pre-trained model, you can download at
                            https://drive.google.com/drive/folders/1s4M-_SnCPMj_2rsMkSy3pLnLQcgRakAe?usp=sharing
      --dataset DATASET     Directory of your test_image ""folder""
      --cuda CUDA           cuda for cuda, cpu for cpu, default = cuda
      --batch_size BATCH_SIZE
                            batchsize, default = 4
      --logdir LOGDIR       logdir, log on tensorboard
      --save_dir SAVE_DIR   save result images as .jpg file. If None -> Not save

• To report score,
• dataset should contain masks and images folder.
• You should specify logdir to get PR-Curve.
• The Scores will be printed out on your stdout.
• You should have model files below the model_dir.
• Only supports model files named like "[N]epo_[M]step.ckpt" format.

    usage: measure_test.py [-h] --model_dir MODEL_DIR --dataset DATASET
                       [--cuda CUDA] [--batch_size BATCH_SIZE]
                       [--logdir LOGDIR] [--which_iter WHICH_ITER]
                       [--cont CONT] [--step STEP]

    optional arguments:
      -h, --help            show this help message and exit
      --model_dir MODEL_DIR
                            Directory of folder which contains pre-trained models, you can download at
                            https://drive.google.com/drive/folders/1s4M-_SnCPMj_2rsMkSy3pLnLQcgRakAe?usp=sharing
      --dataset DATASET     Directory of your test_image ""folder""
      --cuda CUDA           cuda for cuda, cpu for cpu, default = cuda
      --batch_size BATCH_SIZE
                            batchsize, default = 4
      --logdir LOGDIR       logdir, log on tensorboard
      --which_iter WHICH_ITER
                            Specific Iter to measure
      --cont CONT           Measure scores from this iter
      --step STEP           Measure scores per this iter step

Detailed Guideline

Pretrained Model

You can download pre-trained models from https://drive.google.com/drive/folders/1s4M-_SnCPMj_2rsMkSy3pLnLQcgRakAe?usp=sharing

Dataset

PairDataset Class

• You can use CustomDataset.
• Your custom dataset should contain images, masks folder.
  • In each folder, the filenames should be matched.
  • eg. images/a.jpg masks/a.jpg

DUTS

You can download dataset from http://saliencydetection.net/duts/#outline-container-orgab269ec.

• Caution: You should check the dataset's Image and GT are matched or not. (ex. # of images, name, ...)
• You can match the file names and automatically remove un-matched datas by using DUTSDataset.arrange(self) method
• Please rename the folders to images and masks.

Directory & Name Format of .ckpt files

"models/state_dict//<#epo_#step>.ckpt"

• The step is accumulated step from epoch 0.