numalogic

Background

Numalogic is a collection of ML models and algorithms for operation data analytics and AIOps.
At Intuit, we use Numalogic at scale for continuous real-time data enrichment including
anomaly scoring. We assign an anomaly score (ML inference) to any time-series
datum/event/message we receive on our streaming platform (say, Kafka). 95% of our
data sets are time-series, and we have a complex flowchart to execute ML inference on
our high throughput sources. We run multiple models on the same datum, say a model that is
sensitive towards +ve sentiments, another more tuned towards -ve sentiments, and another
optimized for neutral sentiments. We also have a couple of ML models trained for the same
data source to provide more accurate scores based on the data density in our model store.
An ensemble of models is required because some composite keys in the data tend to be less
dense than others, e.g., forgot-password interaction is less frequent than a status check
interaction. At runtime, for each datum that arrives, models are picked based on a conditional
forwarding filter set on the data density. ML engineers need to worry about only their
inference container; they do not have to worry about data movement and quality assurance.

Numalogic realtime training

For an always-on ML platform, the key requirement is the ability to train or retrain models
automatically based on the incoming messages. The composite key built at per message runtime
looks for a matching model, and if the model turns out to be stale or missing, an automatic
retriggering is applied. The conditional forwarding feature of the platform improves the
development velocity of the ML developer when they have to make a decision whether to forward
the result further or drop it after a trigger request.

Installation

numalogic can be installed using pip.

pip install numalogic

If using mlflow for model registry, install using:

pip install numalogic[mlflow]

Build locally

Install Poetry:

curl -sSL https://install.python-poetry.org | python3 -

To activate virtual env:
```
poetry shell
```
To install dependencies:
```
poetry install
```
If extra dependencies are needed:
```
poetry install --all-extras
```
To run unit tests:
```
make test
```
To format code style using black:
```
make lint
```

Usage

The following example shows a series of operations that can be performed using the package.

Model training

Let's use one of the available models to train on generated data.
Here we are going to use a sparse convolutional Autoencoder.
We can start by scaling the data.

from sklearn.preprocessing import MinMaxScaler
from numalogic.models.autoencoder.variants import Conv1dAE
from numalogic.models.autoencoder.pipeline import SparseAEPipeline

# Scale the train and test sets.
scaler = MinMaxScaler()
X_train = scaler.fit_transform(train_df.to_numpy())

# Define the model and train
seq_len = 36
# SparseAEPipeline is a sparse autoencoder trainer that follows sklearn's API pattern
model = SparseAEPipeline(
    model=Conv1dAE(in_channels=3, enc_channels=8), seq_len=seq_len, num_epochs=30
)
model.fit(X_train)

Model saving

Now that the model is trained, let's save it. Numalogic has built in support
for Mlflow's tracking and logging system.

Let's first start the mlflow server on localhost,
which has already been installed optionally via poetry dependency:

mlflow server \
        --default-artifact-root {directory}/mlruns --serve-artifacts \
        --backend-store-uri sqlite:///mlflow.db --host 0.0.0.0 --port 5000

from numalogic.registry import MLflowRegistrar

# static and dynamic keys are used to look up a model
static_keys = ["synthetic", "3ts"]
dynamic_keys = ["minmaxscaler", "sparseconv1d"]

registry = MLflowRegistrar(tracking_uri="http://0.0.0.0:5000", artifact_type="pytorch")
registry.save(
   skeys=static_keys, 
   dkeys=dynamic_keys, 
   primary_artifact=model, 
   secondary_artifacts={"preproc": scaler}
)

Model loading

Let's now load the model for inference on test data

registry = MLflowRegistrar(tracking_uri="http://0.0.0.0:8080")
artifact_dict = registry.load(
    skeys=static_keys, dkeys=dynamic_keys
)
scaler = artifact_dict["secondary_artifacts"]["preproc"]
model = artifact_dict["primary_artifact"]

Model inference

Finally, let's use the loaded artifacts to predict on the anomalous test data.
This can be a streaming or a batched data.

X_test = scaler.transform(outlier_test_df.to_numpy())

# predict method returns the reconstruction produced by the AE
test_recon = model.predict(X_test)

# score method returns the anomaly score, calculated using thresholds.
# A number less than 1 indicates an inlier, and greater than 1 indicates an outlier.
test_anomaly_score = model.score(X_test)

Contributing

We would love contributions in the numalogic project in one of the following (but not limited to) areas:

Adding new time series anomaly detection models
Making it easier to add user's custom models
Support for additional model registry frameworks

For contribution guildelines please refer here.

diego-ponce/numalogic