If you log an URL as a string, e.g., a link to a heavy output such as video visualization, it is rendered as a clickable URL.

This way, there is no need to copy the content and paste it into a new browser tab. Instead, you can navigate to it right away from the Runs table or Run details.

If an invitation to a workspace gets stuck somewhere or is accidentally deleted, the admin can resend it to the user.

Project owners can now set storage limits for projects individually.

Upgrade of Neptune dependencies to newer versions to keep up with all the improvements and continued support: Elasticsearch (=7.11, including 8), MySQL (8), Kafka (you can use any 2.x and 3.x version).

Azure Blob Storage is now supported in addition to S3 and PVC.

Further updates to the on-prem documentation to provide you with the best practices and instructions for common (and less common) scenarios.

Natively track uncommitted changes in the experiment environment.

Easily compare and analyze the differences in your code between runs to debug experiments quicker and reproduce them without losing any critical modifications in the source code.

Automatically track packages used in the experiment environment to streamline the process of reproducing environments without the risk of misconfiguration.

When fetching the runs table you can specify which fields (columns) should be fetched through the columns parameter. If unspecified all possible fields will be fetched.

Neptune stores most of its data in S3-compatible object storage. In situations where access to such storage is not feasible, we support a POSIX-compliant file system as a storage medium.

Historically, we delivered MooseFS (an open-source distributed POSIX-compliant file system in many aspects similar to NFS) as part of our installation package. Some time ago, we deprecated MooseFS usage and started actively moving all customers away from it. In Neptune 2.3, MooseFS is no longer supported.

We’re happy to help you with moving to supported storage. In case of questions, contact us at support@neptune.ai.

We are planning small changes to improve the UX of the Project management API:

• create_project method will no longer require the key parameter. If not provided we will create a project key for you.

• It will be possible to find programmatically what was the reason for the error. E.g. which parameter is at fault when a validation error occurs.

A connection mode can be set as an environment variable. If you want to turn temporarily off logging to Neptune you can do so by changing to the debug connection mode without modifying the code.

To facilitate better programmatic access to the model registry, for e.g. to create an automatic workflow validating models on staging with the newest validation dataset you can fetch the models table, in the same manner, you would the model versions table or the runs table.

We've patched a severe security vulnerability in Neptune on-premises version. This update doesn't contain new features, but it addresses the RemoteCodeExecution vulnerability issue announced on March 31st. 

Now instead of the full project name “YOUR_WORKSPACE/PROJECT_NAME,” you can provide just the project name in the code if the project name is unique.

Our neptune.new API is introducing improved, simplified and consistent way of tracking and fetching experiment metadata which will allow you not only to track more objects like histograms, audio, video, charts etc. but also resume existing experiments.

With the new client library you will also be able to: 

• organize experiment’s metadata into logical groups like ‘train’, ‘validation’ etc.

• track data stored on disk in case of connection failure

• track data with Neptune without Internet connection (offline mode)

Neptune.new has support for ML pipelines and distributed and parallel computing e.g. Apache Spark. 

There is also preparation for upcoming new functionality - Artifact (Model Registry) that will facilitate exporting models to production (CI/CD pipelines).

Quickstart example

import neptune 

exp = neptune.init(project='neptune/sandbox')  

# track hyperparameters of your experiment exp['parameters/batch_size'] = 5 exp['parameters/algorithm'] = 'ConvNet'  

# track the training process by logging your training metrics 
for epoch in range(100): 
exp['train/accuracy'].log(0.98)       exp['train/loss'].log(0.053) exp['my_debug/predictions'].log(neptune.Image('misclassifed.png'))  

# track the final results 
exp['f1_score'] = 0.66 exp['trained_model'].save('path/to/your/model.h5')

You can think of experiments as dict-like structures. Look at the following example:

exp['parameters/batch_size'] = 5
exp['parameters/algorithm'] = 'ConvNet'
exp['accuracy'].log(0.95)
exp['accuracy'].log(0.96)
exp['accuracy'].log(0.97)
exp['results/trained_model'].save('path/to/your/model')

that will result in the following experiment structure:

parameters:
 batch_size: Integer
 algorithm: String
accuracy: FloatSeries
results:
 trained_model: Files

Experiment structure will consists of variables that can be organized into namespaces. 

We've got feedback that you would like to stop and restart an experiment without creating a new one. We introduced this option in the neptune-client 1.0.

You can now update your runs even if they finished. Check the docs for more: https://docs.neptune.ai/how-to-guides/neptune-api/resume-run

In the new neptune-client 1.0 we are going to introduce namespaces.

With namespaces, you will be able to organized tracked metadata into meaningful steps of your pipeline e.g

# in preparation script 
exp['preparation/input_dataset'] = neptune.Artifact('test.csv', version='0.3')   

# in training script 
exp['train/accuracy'].log(0.96)

# in validation script 
exp['validation/accuracy'] = 0.89

Namespaces can be nested, for example:

exp['parameters/model/lr'] = 0.05

With the new neptune-client 1.0 you will be able track data to your experiment from many processes (running on different machines) with no problem.

Neptune will be fully compatible with distributed computing frameworks like e.g. Apache Spark, so you can use them to synchronize your processes. On top of that, Neptune is going to provide some synchronization methods that will help you handle more sophisticated workflows.

We've patched a severe security vulnerability in Neptune on-premises version. This addresses the RemoteCodeExecution vulnerability issue announced on December 19th, 2021.

Small improvement in displaying numbers in scientific notation, trailing zeros will be removed

• 1.200000e6 -> 1.2e6

• 1.000000e10 -> 1.0e10

We introduced 4-bytes UTF-8 support in tags:

• are case-sensitive (search included)

• 👀support emojis and Chinese characters

• works from the UI and client libraries

We’ve got feedback that sometimes you want to compare experiments across multiple views or pages. You can do it now with all of the bulk operations: compare, manage tags, abort, and delete.

.