Deep Learning for Computer Vision final project

Deep Learning for Computer Vision final project

Team: DLCV1

Member & Contribution:

• 林彥廷 (R06943184): 主程式撰寫、模型訓練 (50%)
• 王擎天 (R06945055): 副程式撰寫、模型訓練、海報設計 (50%)

Overview:

This project contains code to predict image's type from different domain using moment matching.

Description:

Folders:

• script: folder contains scripts
• src: folder contains source code
• model: folder contains saved models which automatically download from network

Files:

• script/get_dataset.sh: script which downloads training and testing dataset
• script/download_from_gdrive.sh: script which downloads googledrive data
• script/parse_data.sh: script which loads training dataset and converts to torch dataset
• script/predict.sh: script which predicts images
• script/evaluate.sh: script which evaluates the model
• script/predict_for_verify.sh script which generates mini-batch average validation accuracy and loss plot
• src/models/classifier.py: classifier model
• src/models/loss.py: loss function
• src/models/pretrained.py: pretrained model
• src/models/model.py: Model and function for prediction and evaluation
• src/parse_data.py: load data in folder and convert them to torch dataset
• src/predict.py: prediction main function
• src/evaluate.py: evaluation main function
• src/train.py: training function
• src/utils.py: code for parsing and saving
• src/util/dataset.py: customized dataloader
• src/util/visual.py: code for visualization
• src/create_path_csv.py:main function to create image path csv file for image folder

Dataset:

Download training and testing dataset to folder named "dataset_public":

bash ./script/get_dataset.sh

WARNING:

Usage:

Create image-path csv file for image folder:

User can use this script to create image-path csv file

python3 src/create_path_csv.py $1

• $1 is the folder containing the images

Example: (path: /home/final-dlcv1)

python3 src/create_path_csv.py dataset_public/test

The result will look like following text: image_name,label test/018764.jpg,-1 test/034458.jpg,-1 test/050001.jpg,-1 test/027193.jpg,-1 test/002637.jpg,-1 test/017265.jpg,-1 test/048396.jpg,-1 test/013178.jpg,-1 test/036777.jpg,-1 ......

Predict labels of images:

User can use this script to predict labels of images

bash ./script/predict.sh $1 $2 $3 $4 $5

• $1 is the domain of images (Option: infograph, quickdraw, real, sketch)
• $2 is the folder containing the images
• $3 is the csv file contains image paths
• $4 is the folder to saved the result file
• $5 is the batch size

Example 1: Predict images from real domain (path: /home/final-dlcv1)

bash script/predict.sh real dataset_public dataset_public/test/image_path.csv predict 256

Example 2: Predict images from sketch domain (path: /home/final-dlcv1)

bash script/predict.sh sketch dataset_public dataset_public/sketch/sketch_test.csv predict 256

Example 3: Predict images from infograph domain (path: /home/final-dlcv1)

bash script/predict.sh infograph dataset_public dataset_public/infograph/infograph_test.csv predict 256

Example 4: Predict images from quickdraw domain (path: /home/final-dlcv1)

bash script/predict.sh quickdraw dataset_public dataset_public/quickdraw/quickdraw_test.csv predict 256

Evaluate the result file:

User can use this script to evaluate the reuslt file with answer file, it will print result on the screen

bash ./script/evaluate.sh $1 $2

• $1 is the predicted file csv
• $2 is the answer file csv

Example (path:/home/final-dlcv1)

bash ./script/evaluate.sh predict/real_predict.csv test/test_answer.csv

Reference

• https://gntuedutw-my.sharepoint.com/:p:/g/personal/b03901093_g_ntu_edu_tw/EUuiRqkxp69GjGDOzVcJeHoBIY61MIbEEbBVsMgNBfcEBg?rtime=uLaDbmDq1kg
• Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K. and Wang, B., 2018. Moment Matching for Multi-Source Domain Adaptation. arXiv preprint arXiv:1812.01754.