Evidently and DVCLive

Log Evidently metrics with DVC via DVCLive.

This is a community-contributed integration. Author: Francesco Calcavecchia.

TL;DR: You can use Evidently to calculate metrics, and DVCLive to log and view the results.

Jupyter notebook with en example:

Overview

DVCLive can be used to track the results of Evidently. In the following we demonstrate it through an example.

How it works

Evidently calculates a rich set of metrics and statistical tests. You can choose any of the pre-built reports or combine individual metrics to define what you want to measure. For example, you can evaluate prediction drift and feature drift together.

You can then generate the calculation output in a Python dictionary format. You should explicitly define which parts of the output to send to DVCLive Tracking.

Step 1. Install DVCLive and Evidently

Install DVCLive, Evidently, and pandas (to handle the data) in your Python environment:

$ pip install dvc dvclive evidently pandas

Step 2. Load the data

Load the data from UCI repository and save it locally. For demonstration purposes, we treat this data as the input data for a live model. To use with production models, you should make your prediction logs available.

$ wget https://archive.ics.uci.edu/static/public/275/bike+sharing+dataset.zip
$ unzip bike+sharing+dataset.zip
import pandas as pd

df = pd.read_csv("raw_data/day.csv", header=0, sep=',', parse_dates=['dteday'])
df.head()

Step 3. Define column mapping

You should specify the categorical and numerical features so that Evidently performs the correct statistical test for each of them. While Evidently can parse the data structure automatically, manually specifying the column type can minimize errors.

from evidently.pipeline.column_mapping import ColumnMapping

data_columns = ColumnMapping()
data_columns.numerical_features = ['weathersit', 'temp', 'atemp', 'hum', 'windspeed']
data_columns.categorical_features = ['holiday', 'workingday']

Step 4. Define what to log

Specify which metrics you want to calculate. In this case, you can generate the Data Drift report and log the drift score for each feature.

from evidently.report import Report
from evidently.metric_preset import DataDriftPreset


def eval_drift(reference, production, column_mapping):
    data_drift_report = Report(metrics=[DataDriftPreset()])
    data_drift_report.run(
        reference_data=reference, current_data=production, column_mapping=column_mapping
    )
    report = data_drift_report.as_dict()

    drifts = []

    for feature in (
        column_mapping.numerical_features + column_mapping.categorical_features
    ):
        drifts.append(
            (
                feature,
                report["metrics"][1]["result"]["drift_by_columns"][feature][
                    "drift_score"
                ],
            )
        )

    return drifts

You can adapt what you want to calculate by selecting a different Preset or Metric from those available in Evidently.

Step 5. Define the comparison windows

Specify the period that is considered reference: Evidently will use it as the base for the comparison. Then, you should choose the periods to treat as experiments. This emulates the production model runs.

#set reference dates
reference_dates = ('2011-01-01 00:00:00','2011-01-28 23:00:00')

#set experiment batches dates
experiment_batches = [
    ('2011-01-01 00:00:00','2011-01-29 23:00:00'),
    ('2011-01-29 00:00:00','2011-02-07 23:00:00'),
    ('2011-02-07 00:00:00','2011-02-14 23:00:00'),
    ('2011-02-15 00:00:00','2011-02-21 23:00:00'),
]

Step 6. Run and log experiments in DVC

There are two ways to track the results of Evidently with DVCLive:

  1. you can save the results of each item in the batch in one single experiment (each experiment corresponds to a git commit), in separate steps

  2. or you can save the result of each item in the batch as a separate experiment

We will demonstrate both, and show you how to inspect the results regardless of your IDE. However, if you are using VSCode, we recommend using the DVC extension for VS Code to inspect the results.

Step 7

Option 1. One single experiment

from dvclive import Live

with Live() as live:
    for date in experiment_batches:
        live.log_param("begin", date[0])
        live.log_param("end", date[1])

        metrics = eval_drift(
            df.loc[df.dteday.between(reference_dates[0], reference_dates[1])],
            df.loc[df.dteday.between(date[0], date[1])],
            column_mapping=data_columns,
        )

        for feature in metrics:
            live.log_metric(feature[0], round(feature[1], 3))

        live.next_step()

You can then inspect the results using

$ dvc plots show

In a Jupyter notebook environment, you can show the plots as a cell output simply by using Live(report="notebook").

Option 2. Multiple experiments

from dvclive import Live

for step, date in enumerate(experiment_batches):
    with Live() as live:
        live.log_param("begin", date[0])
        live.log_param("end", date[1])
        live.log_param("step", step)

        metrics = eval_drift(
            df.loc[df.dteday.between(reference_dates[0], reference_dates[1])],
            df.loc[df.dteday.between(date[0], date[1])],
            column_mapping=data_columns,
        )

        for feature in metrics:
            live.log_metric(feature[0], round(feature[1], 3))

You can the inspect the results using

$ dvc exp show
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  Experiment                 Created    weathersit    temp   atemp     hum   windspeed   holiday   workingday   step   begin                 end
 ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  workspace                  -               0.231       0       0   0.062       0.012     0.275        0.593   3      2011-02-15 00:00:00   2011-02-21 23:00:00
  master                     10:02 AM            -       -       -       -           -         -            -   -      -                     -
  ├── a96b45c [muggy-rand]   10:02 AM        0.231       0       0   0.062       0.012     0.275        0.593   3      2011-02-15 00:00:00   2011-02-21 23:00:00
  ├── 78c6668 [pawky-arcs]   10:02 AM        0.155   0.399   0.537   0.684       0.611     0.588        0.699   2      2011-02-07 00:00:00   2011-02-14 23:00:00
  ├── c1dd720 [joint-wont]   10:02 AM        0.779   0.098   0.107    0.03       0.171     0.545        0.653   1      2011-01-29 00:00:00   2011-02-07 23:00:00
  └── d0ddb8d [osmic-impi]   10:02 AM        0.985       1       1       1           1      0.98        0.851   0      2011-01-01 00:00:00   2011-01-29 23:00:00
 ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

In a Jupyter notebook environment, you can access the experiments results using the Python DVC api:

import dvc.api

dvc.api.exp_show()

Last updated