Python is installed.

Following packages are installed:pandas, numpy, matplotlib, scikit-learn, pytorch, torchvision.

It is highly recommended that you use use high power gpu, such as nvidia Titan XP to train the model, otherwise, it will be slow.

There are many argument flags you can add. All the default settings are in utils/config.py file. The model will be saved in model_dir (default is ./experiments/). In model_dir, there will be best model weights, train.log, and loss/acc plots. You can change model_dir using '--model_dir' flags. Given the image folder is '../find_phone/', To train from scratch, run:

python train_phone_finder.py --data_dir ../find_phone/

To continue previous training, run:

python train_phone_finder.py --data_dir ../find_phone/ --resume

Here is the list of command how the provided model (in experiments folder) is generated:

python train_phone_finder.py --data_dir ../find_phone/ --gpu 1 --rotation --pretrained --freezed_layers 60 --epochs 100 --model_dir ./experiments
python train_phone_finder.py --data_dir ../find_phone/ --gpu 1 --rotation --pretrained --freezed_layers 3 --epochs 200 -- model_dir ./experiments --resume

Notice: --rotation flag make the images randomly rotate while training. Also the code will automatically split the dataset in 90%/10% train/validation split. Notice: After running the above commands, the validation accuracy should be 100%, and train accuracy should be around 75-80%, however, without random rotation and translation, the accuracy on train dataset should be around 99%. Notice: The default model is resnet18. Notice: During training, the validation accuracy will fluctuate. However, it is expected since the validation dataset has only around 10 images.

Pass the image path or path of the directory contains images as argument. Here are two exmaples: To predict single image:

python find_phone.py ../find_phone/0.jpg

To predict multiple images (given all images are store in one directory):

python find_phone.py ../find_phone/

Notice: It is recommended that you use the second command if you have multiple images, since with second command, you only need to load the model once which can save you some time.

My code uses resnet (arvix) to predict. I am using similar code structure of my current lab project

├── README.md          <- The top-level README for developers using this project.
├── data               <- Scripts to that load and transform images and labels
│   ├── dataloader.py
│   └── d_utils.py
│
├── models             <- Used to Create Models
│   └── model.py
├── utils         
│   ├── config.py      <- Script that contains default settings
│   ├── constants.py   <- Script that contains default constants
│   ├── eval_utils.py  <- Script that contains evaluation meters
│   └── logger.py      <- Script that creates logger
├── requirements.txt   <- The file that contains the names of required python packages
│
├── train_phone_finder.py   <- Script to train models 
│
├── train_phone_finder.py   <- Script to make predictions
│
└── trainer.py         <- Module to handle model training