Author: Liangwei Li
E-mail: [email protected]
This project provides a pytorch version solution to a classic Kaggle competition: Dogs and Cats. Every participant is supposed to train a model for this binary classification: to predict whether a picture represents a cat or a dog. In this project, one version is completed with pytorch-cpu and some other relative packages of python. It can record log and model state automatically at every epoch, which means it can start even if an interuption occurs.
Training data as well as test data are excluded in this project. You need to download it yourself.
Mainly, ResNet34 is used as the model to be trained. However, in .models, any possible model can be difined and thus trained.
To begin the training process on your own computer, some specific dependent packages are required, which are listed in requirements.txt. When the environment is set, your need to make some changes of the parameters "train_all_path, test_all_path " in config.py, which will be used to define the location of the data set. After this, run main.py and every thing should work normally.
To test the proformance of the model, make one change at the end of main.py as follows:
#train(model, training_all)
evaluate(model, dev_all)