Hip Landmark detection using Convolutional Neural Networks

Pytorch Implementation of Hip Landmark detection using Convolutional Neural Networks

Authors

• Imad Eddine Ibrahim Bekkouch i.bekkouch@innopolis.university
• Bulat Ibragimov b.ibragimov@innopolis.ru
• Ramil Kuleev r.kuleev@innopolis.ru
• Bulat Maksudov b.maksudov@innopolis.ru
• Tamerlan Aidinovich t.mustafaev1994@gmail.com

Innopolis University

University of Copenhagen

Innopolis

Getting Started

Please follow the instructions to get an up and running version of our code running on your local machine.

Prerequisites

Please make sure you have the following installed.

1. Python 3.6
2. Pytorch 1.0
3. termcolor 1.1.0
4. Numpy 1.15+

Running the code

First step is to unzip the office31 data

cd ./data
tar xvzf office31.tar.gz --keep-newer-files

Run the main file with using any of the following arguments:

usage: main.py [-h] [--image_size IMAGE_SIZE] [--num_classes NUM_CLASSES]
               [--train_iters TRAIN_ITERS] [--batch_size BATCH_SIZE]
               [--num_workers NUM_WORKERS] [--pre_lr PRE_LR] [--lr1 LR1]
               [--lr2 LR2] [--random_seed RANDOM_SEED]
               [--labeled_target_ratio LABELED_TARGET_RATIO]
               [--validation_source_ratio VALIDATION_SOURCE_RATIO]
               [--mode {train,test}]
               [--source {usps,svhn_extra,mnist,svhn,amazon,webcam,dslr}]
               [--target {usps,svhn_extra,mnist,svhn,amazon,webcam,dslr}]
               [--model_path MODEL_PATH] [--graph_path GRAPH_PATH]
               [--sample_path SAMPLE_PATH] [--mnist_path MNIST_PATH]
               [--usps_path USPS_PATH] [--svhn_path SVHN_PATH]
               [--off_path OFF_PATH] [--log_step LOG_STEP]
               [--iter_dom_adap ITER_DOM_ADAP] [--log_pre LOG_PRE]
               [--verbose VERBOSE]

optional arguments:
  -h, --help            show this help message and exit
  --image_size IMAGE_SIZE
  --num_classes NUM_CLASSES
  --train_iters TRAIN_ITERS
  --batch_size BATCH_SIZE
  --num_workers NUM_WORKERS
  --pre_lr PRE_LR       learning rate for the source domain pre-training
  --lr1 LR1             encoder"s and classifier"s learning rate for domain
                        adaptation
  --lr2 LR2             discriminator"s learning rate
  --random_seed RANDOM_SEED
  --labeled_target_ratio LABELED_TARGET_RATIO
                        ratio of target domain used for labeling
  --validation_source_ratio VALIDATION_SOURCE_RATIO
  --mode {train,test}
  --source {usps,svhn_extra,mnist,svhn,amazon,webcam,dslr}
                        source domain
  --target {usps,svhn_extra,mnist,svhn,amazon,webcam,dslr}
                        target domain
  --model_path MODEL_PATH
  --graph_path GRAPH_PATH
  --sample_path SAMPLE_PATH
  --mnist_path MNIST_PATH
  --usps_path USPS_PATH
  --svhn_path SVHN_PATH
  --off_path OFF_PATH
  --log_step LOG_STEP
  --iter_dom_adap ITER_DOM_ADAP
  --log_pre LOG_PRE     number of iterations to print the losses and accuracy
                        for pre training
  --verbose VERBOSE

Code structure

Quick explanation of the code structure:

.
├── data                      # Data Folder
│   ├── office31              # Office31 dataset
│   │   ├──amazon             # Amazon pictures
│   │   ├──webcam             # Webcam pictures
│   │   └──dslr               # Digital SLR pictures
│   ├── mnist                 # MNIST digits dataset
│   ├── usps                  # USPS digits dataset
│   └── svhn                  # SVHN digits dataset
├── model                     # Code, pictures and saved models
│   ├── code                  # Source Code
│   │   ├──data_loader.py     # Data loaders for Office31, mnist, usps and svhn datasets
│   │   ├──main.py            # Main
│   │   ├──model.py           # Model definition (Classier, Encoder and Discriminator)
│   │   └──solver.py          # Domain Adaptation Algorithm definition
│   ├── graphs                # Accuracy and loss graphs
│   └── models                # Saved Models
├── LICENSE                   # MIT LICENCE
└── README.md                 # README

License

This project is licensed under the MIT License - see the LICENSE file for details# Domain-Generalization