Process Transformer
Transformer Neural Model for Business Process Monitoring Tasks
Tasks
- Next Activity Prediction
- Time Prediction of Next Activity
- Remaining Time Prediction
Install
pip install processtransformer
Usage
import argparse
import tensorflow as tf
from processtransformer import constants
from processtransformer.data import loader
from processtransformer.models import transformer
parser = argparse.ArgumentParser(description="Process Transformer - Next Activity Prediction.")
parser.add_argument("--dataset", required=True, type=str, help="dataset name")
parser.add_argument("--task", type=constants.Task,
default=constants.Task.NEXT_ACTIVITY, help="task name")
parser.add_argument("--epochs", default=1, type=int, help="number of total epochs")
parser.add_argument("--batch_size", default=12, type=int, help="batch size")
parser.add_argument("--learning_rate", default=0.001, type=float,
help="learning rate")
# Load data
data_loader = loader.LogsDataLoader(name = args.dataset)
(train_df, test_df, x_word_dict, y_word_dict, max_case_length,
vocab_size, num_output) = data_loader.load_data(args.task)
# Prepare training examples for next activity prediction task
train_token_x, train_token_y = data_loader.prepare_data_next_activity(train_df,
x_word_dict, y_word_dict, max_case_length)
# Create and train a transformer model
transformer_model = transformer.get_next_activity_model(
max_case_length=max_case_length,
vocab_size=vocab_size,
output_dim=num_output)
transformer_model.compile(optimizer=tf.keras.optimizers.Adam(args.learning_rate),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
transformer_model.fit(train_token_x, train_token_y,
epochs=args.epochs, batch_size=args.batch_size)
See complete code examples within the github repository for other tasks, including preparing raw process data for transformer model.
