A pytorch implementation for the CREPE text classification model (https://arxiv.org/abs/1509.01626).

It is trained to predict the toxicity of a given piece of text.

The train data has to be stored as a .csv file in the ./data folder and contain 2 columns - the labels and the raw texts themselves.

Importnat note: the current implementation is a binary classifier, which means it has only one single sigmoid output and is being trained with the Binary Cross Entropy loss, hence the labels should be integer values (either 0 or 1) and don't need any additional encoding.

• MAX_LENGTH - defines the maximal length of each input text (longer texts are truncated, shorter texts are padded with zeros)

• CHANNELS - the number of filters (same on each hidden convolutional layer)

• KERNEL_SIZES - the list of kernel sizes for each convolutional layer

• POOLING_SIZE - the size of each pooling operation in the model

• LINEAR_SIZE - the size of the hidden fully connected layers (same for all of them)

• DROPOUT - the dropout value

• OUTPUT_SIZE - the number of the outputs (should be equeal to the number of classes in case it's a multilabel classification problem)

• EPOCHS - the number of epochs

• BATCH_SIZE - the batch size

• LEARNING_RATE- the learning rate

• MODEL_PATH - the path for all model-related files

• DATA_PATH - the training dataset path (folder + filename)

• EXPERIMENT_PREFIX - a prefix for the experiment - all the corresponding files will have this prefix before their names

• RUS - a hardcoded cyrillic alphabet string

• BEST_MODEL_PATH - the filename of the best model (folder + filename)

To train the model you need to define all the desired parameters in the constants.py file (or in a dev.env file in the future) After that you can run:

python train.py

The training process will procude multiple files in the model directory:

• A train log file called train.log

• Two models with the .pth.tar extension: the best model and last model

• Two json files with the evaluation metrics for the best and the last models