SacreD Experiment Manager (sdem)

Description

Sacred Experiment manager combines sacred and dvc to run experiments locally, on HPC clusters and across different users.

Currently sdem works with sacred local files, but there are plans to support mongo experiments.

Installation

Through pip

pip install sdem

From github

git clone [email protected]:defaultobject/sdem.git
cd sdem
pip install -e .

Example

The full example is shown in the example folder.

Run

There are two ways to run a file, the first is through the sdem cli. In The experiment folder run:

sdem --verbose run

which will sequentially run all models found in the models folder.

Alternatively we can directly run the models file:

python m_model.py -1

View table of results

sdem provides some convenient functions to see the results of the experiments ran. To automatically create a table of results go into the metrics folder and run

python table_of_results.py

which runs

results_df = get_ordered_table(
    '../',
    metrics=['train_rmse', 'test_rmse'],
    group_by=[
        'name', 
    ],
    results_by=[''],
    combine=True,
    flatten=True,
)

This will go through every sacred run, and group the metrics by name and compute the mean and std across all experiments in this group (ie across folds). This results in

    name          train_rmse_score    test_rmse_score
--  ------------  ------------------  -----------------
 0  linear_model  0.10 $\pm$ 0.01     0.10 $\pm$ 0.03

To view the results of each fold we can simply add this to the results_by argument

results_df = get_ordered_table(
    '../',
    metrics=['train_rmse', 'test_rmse'],
    group_by=[
        'name', 
    ],
    results_by=['fold'],
    combine=True,
    flatten=True,
)

which results in

    name            fold  train_rmse_score    test_rmse_score
--  ------------  ------  ------------------  -----------------
 0  linear_model       0  0.11 $\pm$ nan      0.09 $\pm$ nan
 1  linear_model       1  0.11 $\pm$ nan      0.06 $\pm$ nan
 2  linear_model       2  0.09 $\pm$ nan      0.14 $\pm$ nan
 3  linear_model       3  0.11 $\pm$ nan      0.09 $\pm$ nan
 4  linear_model       4  0.10 $\pm$ nan      0.11 $\pm$ nan

Get results for a given experiment

As shown in predictions.py we can load (unpacked) pickles and configs for the run experiments:

res_list, config_list = get_results_that_match_dict(
    {
        'fold': 0,
    },
    '../'
)

print(f'Number of experiments found {len(res_list)}')

This will return the (unpacked) pickles and configs of all experiments that match the passed dictionary, in this case it will return the one with fold equal to zero.

Installation

Setup Mongo

To install mongodb on a mac see here https://docs.mongodb.com/manual/tutorial/install-mongodb-on-os-x/.

mongo

use sacred
db.createUser(
  {
    user: "default",
    pwd: "default",
    roles: [ { role: "userAdminAnyDatabase", db: "admin" } ]
  }
)

Install

pip install requirements.py
pip install -e .

SEML

SEML requires a config file:

mkdir ~/.config/seml/

and for the responses use:

username: default
password: default
port: 27017
database: sacred
host: localhost

DVC Setup

Google API Setup

Follow https://dvc.org/doc/user-guide/setup-google-drive-remote#using-a-custom-google-cloud-project

• Create project here https://console.developers.google.com/

• Open OAuth consent screen

• Create OAuth client Credentials

• Enable Google Drive API

• Will need to authenticate on first use

In repo

dvc init
dvc remote add gremote gdrive://<folder_url_id>
dvc remote modify gremote gdrive_client_id <client ID>
dvc remote modify gremote gdrive_client_secret <client secret>
dvc remote default gremote