Repo for training a baseline model for the AI Rush Spam classification challenge. The model is just a simple ResNet50v2 model. We first fine-tune the last layer a few epochs and then train the full model until convergence. Note that the dataset has unlabeled data, but to train the baseline model we only use the labeled data.

1. The function that's bind to NSML infer needs to output a dataframe with the two columns filename and y_pred, where filename is the name of the file that's being predicted and y_pred is an integer value of the predicted class.

spam/spam_classifier/networks contains the neural network definitions.

spam/spam_classifier/models contains your full model code, this handles things like training, evaluation, instantiation of datasets and neural networks, chaining of different neural networks etc.

spam/spam_classifier/datasets contains logic related to loading data.

spam/training/experiments contains all of your experiment configs, setting things like which model and dataset to use, to parameters related to models and networks.

spam/training/train.py glues everything together with NSML. Loads the experiment config and starts the training.

nsml_train.py This file is just used do make sure that the project runs properly on NSML. It should always be the entrypoint when running nsml run but you can ignore it otherwise.

To run the baseline model training, stand in the airush2020/spam folder and run

nsml run -e nsml_train.py -d spam-1 -m "A good message" -g 1 -a "--experiment_name v1"

Using a geometric mean over all spam classes.

score = (f1_score_monotone * f1_score_screenshot * f1_score_unknown) ** (1 / 3)

There are four different classes; normal, monotone, screenshot, and unknown. The latter three are spam classes. The mapping between class names and the integer values used in the labels file is normal: 0, monotone: 1, screenshot: 2, unknown: 3, unlabeled: -1.

Normal review images, that should not be detected as spam.

Images that are not real review images, and mostly containing one color.

Images that are not real review images, and contain screenshots from the Naver shopping website.

Images that are not real review images, but don't fall under any of the classes above.

Images that have not been labeled yet. Can be used to train the model in a semi-supervised way if you want.

See AI Rush dataset documentation.