Serving for Inference and Training
Once features and training sets are registered, they can be served via Featureform's Python API.
When a feature is defined, it is materialized into the inference store associated with the definition. If a feature has a timestamp, only its most recent value is stored in the inference store. It can be thought of as a feature cache for inference time.
To serve a feature we initialize a serving client, specify the names and variants of the features we wish to serve, and provide the entity whos feature values to look up. Underneath the hood, Featureform will keep a mapping between a feature's name and variant and the tables that they are actually stored in. This allows user's to stick to the Featureform level of abstraction.
fpf = client.features([("fpf", "quickstart")], {"passenger": "1"})
The
wait() function can be used to wait until a feature table is fully materialized before attempting to lookup the feature. It has an optional timeout in second to wait until for the materialization completion.fpf = client.features([("fpf", "quickstart")], {"passenger": "1"}).wait(timeout=60)
When a training set is defined, it is materialized into the offline store associated with the definition. The label from the training set is zipped with the features to form a point-in-time correct dataset. Once that's complete we can initialize a ServingClient and loop through the our dataset
import featureform as ff
client = ff.ServingClient(host)
dataset = client.training_set(name, variant)
for row in dataset:
# print(row.features(), row.label())
# Train Model
The Dataset API takes inspiration from Tensorflow's Dataset API, and both can be used very similarly.
import featureform as ff
client = ff.ServingClient(host)
for row in client.training_set(name, variant).repeat(5):
# Run through 5 epochs of the dataset
import featureform as ff
client = ff.ServingClient(host)
buffer_size = 1000
for row in client.training_set(name, variant).shuffle(buffer_size):
# Run through a shuffled dataset
import featureform as ff
client = ff.ServingClient(host)
for batch in client.training_set(name, variant).batch(8):
# Run batches of features and labels
The
wait() function can be used to wait until the specified training set is ready to be served before attempting to call it. It has an optional timeout in seconds to wait for the training set to be complete.import featureform as ff
client = ff.ServingClient(host)
for row in client.training_set(name, variant).wait(timeout=60):
# Run batches of features and labels
The commands above can be chained together.
import featureform as ff
client = ff.ServingClient(host)
dataset = client.training_set(name, variant).repeat(5).shuffle(1000).batch(64)
for feature_batch, label_batch in dataset:
# Run through a shuffled dataset 5x in batches of 64