Logging to handlers

You can organize your metadata into folder-like structures with the help of namespaces and handler objects.

If you assign a run namespace path to a variable, as on the highlighted line below, you get a namespace handler object.

import neptune

# You can upload the value through an absolute path
run["params/max_epochs"] = 20

# Or with a namespace handler and a relative path
params_namespace = run["params"]
params_namespace["learning_rate"] = 0.005

In this case, the params namespace is created when we assign the nested max_epochs field to the run. You can then access the params_namespace object and use it to nest any logged metadata under the params namespace.

Result

run
|--> params
     |--> max_epochs: 20
     |--> learning_rate: 0.005

In the pipeline logging example below, we show how to use a handler to organize metadata from a particular step of a run into a given namespace.

Using a generic Handler object

If you assign a run path that doesn't exist yet, you get a Handler object that can become anything (including a namespace, if you then nest other things under it).

Example: Handler becomes field

import neptune

run = neptune.init_run()

Because no namespaces or fields exist in the run yet, accessing the path run["train/acc"] returns a Handler object:

handler_object = run["train/acc"]

It's a reference that points to some location in the run – we just don't know what's there yet.

Let's assign something to the handler:

handler_object.append(0.71)

The handler now represents a FloatSeries.

Resulting run structure

run
|--> train
     |--> acc: FloatSeries

Once a field exists under the path which the handler points to, it only accepts the corresponding field methods:

handler_object.append(0.72)  # Works
handler_object.upload("image.png")  # Error: Can't upload an image to a FloatSeries

Example: Handler becomes namespace

In this example, we'll have an overall run structure that is identical to the example above, but our Handler object will reference a different location in the run.

import neptune

run = neptune.init_run()

If we try accessing the path run["train"], a Handler is returned, as that field does not exist yet and no value has been assigned to it:

handler_object = run["train"]

We can now assign something to handler_object. Let's assign the value 0.82 to the field acc and assign that to the handler:

handler_object["acc"] = 0.82

Since the handler object for run["train"] has something nested under it – in this case, a FloatSeries field – the handler now represents a namespace.

Resulting run structure

run
|--> train
     |--> acc: FloatSeries

Pipeline logging example

Let's say we have a data preprocessing step in our pipeline, and we would like to ensure that all of the metadata generated during this part is organized under a folder named "preprocessing".

We would start a Neptune run early in the script:

import neptune
from sklearn.datasets import fetch_lfw_people
from utils import *

dataset = fetch_lfw_people(min_faces_per_person=70, resize=0.4)
run = neptune.init_run()
dataset_config = {...}

To collect the metadata from this step into one place, we set up a preprocessing namespace inside the run. This will be the base namespace where all the preprocessing metadata is logged.

We define the path run["preprocessing"] and assign it to a handler object, which we'll then use to log metadata as if it were a run.

preprocessing_handler = run["preprocessing"]

# Log dataset details
preprocessing_handler["dataset/config"] = dataset_config

# Preprocess dataset
dataset_transform = Preprocessing(...)
...

# Log scaler and features files
preprocessing_handler["dataset/scaler"].upload(path_to_scaler)

This way, whenever we log something to preprocessing_handler, it'll end up under the path run["preprocessing/..."].

See full example script on GitHub 

Tip

The handler exposes similar methods as other Neptune objects.

For details, see API reference Field types: Namespace handler.