A machine learning program that attempts to predict whether a given year will have a snowy Christmas based on the first 11 months of the year

This demo shows the model being trained on 131 years of data from the Vancouver region. The test dataset used is the same dataset to demonstrate the model's ability to converge, or at least memorize data.

The model module can be imported. It provides 2 classes: Trainer and Predictor. As the names suggest, Trainer is used to train a fresh model from scratch, whereas Predictor loads a model from a file and given the data from a year, predicts the probability of snow.

Trainer takes in two lists of filepaths during initialization, one for training and the second for testing. Currently, only the format used by the Canadian government's daily climate observation dataset is accepted. Then, by using the .train method, you can specify the number of epochs the model attempts to learn from the dataset. Using the .save method, the class will export the trained model as a pytorch weights file.

Predictor takes in a path to a pytorch weights file, then initializes and loads the model. Then, the .predict or .predict_tensor methods, it will take the data from either a pandas dataframe or a pytorch tensor to predict whether or not the year will end up snowing during Christmas.

The Forecast model behind everything takes heavy inspiration from image classification models like ResNet. It uses similar blocks to it, but visualizes a year as a 6-channel vector with temperature and precipitation information. Then, it uses 1-dimensional convolutions to attempt to classify the vector as a year that will lead to a snowy or unsnowy Christmas.

The data used is from (climatedata.ca)[climatedata.ca], and consequently the weather data class knows only how to accept data from .csv files that contain all the headers CLIMATE_IDENTIFIER, LOCAL_YEAR, LOCAL_MONTH, LOCAL_DAY, MEAN_TEMPERATURE, MIN_TEMPERATURE, MAX_TEMPERATURE, TOTAL_PRECIPITATION, TOTAL_RAIN, and TOTAL_SNOW.