This repo contains code for the blog post Candidate Generation Using a Two Tower Approach With EG Traveler Data
Expedia Group makes use of DataBricks and for this reason some (interesting) libraries are required.
This library makes use of spark-tfrecord please visit its website and follow the instructions to install on your machine.
The data has been preprocessed for you and can be found INSERT LINK HERE.
You'll first want to setup your python environment which can be done by using the environment-local.yml
file provided in the repo. Then install eg_two_tower with pip.
pip install -e <path to repo dir>
The code for running the data transformations are in eg_two_tower.etl. We ran the ETL on DataBricks with the
environment Databricks Runtime 11.3 LTS for Machine Learning.
If you want to run the ETL on your own machine you'll need either spark-tensorflow-connector or Spark-TFRecord in order for spark to write to tf-records. Since we were using DataBricks spark-tensorflow-connector is already installed.
This script is for training and replicating the experiments in the blog post. Most of the "important" hyperparameters have defaults set with them in the script and for any other hyper-parameters such as the size of the embeddings for the categorical features should be set by modifying the script directly.
To run the experiments in the blog post modify these options as follows
BN + log(q) + l2_norm: : BatchNorm, log(q) correction, and L2 norm
- prob_correction = True
- numerical_input_batch_norm = True
- output_l2 True
BN + log(q): BatchNorm and log(q) correction
- prob_correction = True
- numerical_input_batch_norm = True
- output_l2 = False
BN: BatchNorm Only
- prob_correction = False
- numerical_input_batch_norm = True
- output_l2 = False
We ran the model on DataBricks using a GPU instance, but if you'd like to run it locally depending on your setup you may want to experiment with some of the options below.
- Learning Rate: We found that the model is very sensitive to the learning rate (at least in its current setup) but a learning rate of 0.001 (the default for ADAM in tf) or 0.0001 should suffice (NOTE: we did not experiment with any burn-in for the optimizer)
- Batch size: we trained the model with a size of 2048 but depending on your hardware you might want to change this.
- If you encounter any sort of OOM errors then you may want to modify the training code (train_loop)
so that the model objects are saved (model weights), deleted, and then cleaned up but the python garbage collector.
Additionally, you'll want to make use of
tf.keras.backend.clear_session(). At the time of writing this we beleive the error comes from the .eval() method of the tf.keras Model object, refer to this github issue in the official tensorflow-recommenders repo for more info.
Copyright 2023 Expedia, Inc.
SPDX-License-Identifier: Apache-2.0