ODSC East 2018 materials for "Experimental Reproducibility in Data Science with Sacred" workshop
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
data
experiments
hpo
ingredients
notebooks
.gitignore
README.md
requirements.txt
slides.pdf
utils.py

README.md

Experimental Reproducibility in Data Science with Sacred

Demonstrate how to use Sacred to track machine learning experiments on popular kaggle titanic competition

Getting Started

  • Kaggle Titanic Competition
  • Sacred

Prerequisites

Setup

  1. Clone repo: git clone https://github.com/gilt/odsc-2018
  2. Setup Virtual Environment: python3 -m venv /PATH/TO/odsc-2018/sacred-demo
  3. Install python packages: pip install -r requirements.txt

Run Experiment

Running an experiment with all defaults

python experiments/model_accuracy.py - Notice that different runs yield different results since we have not
controlled the random seed.

But if we fix the seed by running: python experiments/model_accuracy.py with seed=0 we should end up with the same results on every run.

Running a variant

Variants in experiment

To run a different variant of our experiment: python experiments/model_accuracy2.py print_config with variant_rand_params

Similarly, we have a config option called save_submission which is False by default. We can turn it on from the CLI, which causes a submission file to be generated and tracked as an artifact. python experiments/model_accuracy2.py with seed=0 save_submission=True

Variants in ingredients

We also defined a variant_simple in our preprocessing ingredient. To run this variant: python experiments/model_accuracy2.py with preprocess.variant_simple seed=0

We can even use print_config to show a dry run of config and what's changed from the default python experiments/model_accuracy2.py print_config with seed=0 dataset.variant_split save_submission=True

Vary a bunch of stuff

python experiments/model_accuracy2.py with variant_rand_params save_submission=True dataset.variant_presplit

Running an experiment and logging metrics

Using Sacred's Metrics API , we can track the performance of a model with each training step.

In track_metrics.py, we run MLPClassifier and track training/validation performance over n_epochs.

python experiments/track_metrics.py -m sacred with n_epochs=10 seed=0

Running with a mongo observer

  1. Launch local mongo instance: mongod
  2. Run Experiment (result will be stored in sacred database in mongo): python experiments/model_accuracy.py -m sacred

See Results

Mongo

To look at all our runs on mongo:

mongo
use sacred
db.runs.find().pretty()

Sacredboard

Start local Sacredboard server and connect to local MongoDB instance listening on 27017, database name sacred: sacredboard -m sacred

Model Blending Workflow

Run the following to simulate various experiments with random parameter search:

python experiments/model_accuracy2.py -m sacredblender \
    with variant_rand_params dataset.variant_presplit save_submission=True

Note: we switched to a new database sacredblender in case there's any garbage in the sacred database. This is required because we've hardcoded the lookup to sacredblender database

Now run an experiment that blends the top 3 runs based on holdout performance:

python experiments/model_accuracy2.py -m sacred \
    with dataset.blend preprocess.variant_all save_submission=True

Note: we switched back to sacred database here. We also went with the default parameters for the meta blender model

You can always drop a database by doing the following:

mongo
use sacredblender
db.dropDatabase()

Hyperopt Hyper Parameter Optimization

We show a way to use hyperopt, a python library for hyper parameter optimization, with sacred to run many experiments.

Define Search Space

Search configs are defined in hyperopt_hpo_configs as dictionaries. There is one config for each of the experiments defined in the experiments folder. If a new config or experiment is added, they should be included and appropriately mapped in gather_experiments_and_configs.

Run

Usage

usage: hyperopt_hpo.py [-h] [--num-runs [NUM_RUNS]]
                       [--mongo-db-address [MONGO_DB_ADDRESS]]
                       [--mongo-db-name [MONGO_DB_NAME]]
                       [--experiment-file-name [EXPERIMENT_FILE_NAME]]

Logging Hyperopt Tests to Sacred

optional arguments:
  -h, --help            show this help message and exit
  --num-runs [NUM_RUNS]
                        Number of Hyperopt Runs. Default is 10
  --mongo-db-address [MONGO_DB_ADDRESS]
                        Address of the Mongo DB. Default is 127.0.0.1:27017
  --mongo-db-name [MONGO_DB_NAME]
                        Name of the Mongo DB. Default is sacredblender
  --experiment-file-name [EXPERIMENT_FILE_NAME]
                        Which hpo params to used. Add new in
                        hpo/hyperopt_hpo_configs.py. Default is model_accuracy

For --experiment-file-name - Key must exist in exp_and_configs dict

Example

To do 3 hyperopt runs of the sacred experiment in experiments/model_accuracy3 to the default local mongo instance, sacredblender collection.

python hpo/hyperopt_hpo.py --num-runs 3 --experiment-file-name model_accuracy3

Note that we are writing by default to sacredblender here. The blended model can be created after running HPO and selecting the 3 best models to be blended.

References