We've been working on a redesign of our web application. You can now opt into the beta version and give it a try. Log in to the app and click the link in the banner, or access it later from the user menu in the top-right corner.
Improved performance of generating thumbnails for image series.
Fixed an issue with the scrollbar when the legend tooltip is visible in the chart widget.
Enabled compression in table view API requests, which improves loading of the runs table.
Improved rendering of several widgets (charts, image galleries) on a single page. Before this change, large widgets would cause the browser to hang, whereas they now load instantly. Scrolling performance is also smoother, and the local state of widgets is preserved while scrolling.
Fixed an issue where grouped model versions don't expand.
Enabled compression in table view API requests, which improves the loading of the runs table.
With our first major release of the Neptune client library, the "new" API is now officially the main one. See Upgrading for how to get the new package, and check out the full list of changes in the neptune 1.0 upgrade guide.
In image widgets, you can now scroll the image if it's larger than the widget preview size. In interactive CSV tables, the data is now rendered properly for column names that contain dots.
We fixed a couple of issues: a saved dashboard now retains the y-axis scaling if it has been changed to logarithmic, and deleted projects no longer cause the app to crash when viewing usage statistics. We also further tweaked the image gallery.
As a workspace member, you no longer lose access to a project you created if you set the visibility to "workspace" and then back to "private". We also ensured that data is deleted more thoroughly when deleting projects, dashboards, objects, and other things that contain user data.
The image gallery has received further improvements. In the Source code view, we tweaked the Git info a bit – for example, remote URLs are now rendered as clickable links. As for the table view, due to a technical limitation, we now limit the number of indexed characters in String fields to 1024 characters. This means that for runs table search, only the first 1024 characters will be considered.
When you've grouped runs in the table view, you can now bulk-select all runs of a group for comparison. We've also updated the image gallery with multiple enhancements and new features, such as the carousel view mode.
We want to make the log() method more intuitive, so we've released the methods that will eventually replace it: append() and extend(). They're used to create a series of values, appending either a single value or a collection of values at a time, respectively. (It was always possible to append one or multiple values with log(), but we're now separating those into their own methods.)
In the runs table, we fixed a Firefox scrollbar issue and disabled the "Apply" button if you go to manage tags but no changes have actually been made. In the run and comparison views, we fixed a bug where unauthorized users would be able to save changes to dashboards. We also tweaked the image gallery a bit.
We updated the create_booster_summary() function to log tree dataframes in CSV instead of HTML format, and changed the default value of the log_trees_as_dataframe parameter to False. We also made some package and import related fixes.
We made some import fixes and package-related updates to several of our integration libraries: neptune-prophet 0.4.1, neptune-tensorflow-keras 1.2.0, neptune-fastai 0.10.4, neptune-xgboost 0.10.1, neptune-sacred 0.10.1, kedro-neptune 0.1.4, and neptune-sklearn 0.9.7.
Methods that fetch objects in table format (such as fetch_runs_table()) no longer query trashed objects.
The way we detect the entrypoint file, source files, and Git repositories is now more strict and consistent than before.
Due to recent PyTorch Lightning changes, we've removed the jsonschema requirements.
The following alternate ways to install neptune-client now work properly: pip install git+https://github.com/neptune-ai/neptune-client.git and pip install . (after cloning the neptune-client repository).
Ahead of the 1.0 release, we've marked the following as deprecated:
Implicit casting of unsupported types to String when using log() or assign() (a.k.a. =). What does this mean? So far, if you attempt to log a value that is not among the supported neptune.types (such as a tuple) Neptune silently logs it as a string. As of 1.x, you're expected to use str(value) if you want it to be logged as a string; otherwise an exception will be thrown in the case of unsupported types.
We've revamped the image gallery! To analyze images logged as series, you now have more options for displaying and navigating the samples throughout the training process.
Integration: neptune-tensorflow-keras 1.1.0
We fixed a bug that crashed the integration when the input layer of a neural network didn't have the input_shape parameter defined or the model was not built.
We've improved the Neptune integration with 🤗 Transformers, which was released with version 4.22.0 . For the updated functionality, see the 🤗 Transformers guide.
As part of the management API, you can now move runs or other objects to trash with the trash_objects() function.
Python client library: neptune-client 0.16.7
When resuming a run or other object, we've introduced a new, unified parameter for specifying the ID of the object: with_id.
When fetching a table of objects (for example, with project.fetch_runs_table() you can use the new columns parameter to specify which fields to include as columns in the table, rather than having all fields returned.
Neptune no longer hangs indefinitely if the synchronization thread has stopped when wait() is called. The process also finishes faster when you call stop().
Integration: neptune-tensorflow-keras 1.0.0
We've added a bunch of new functionality to our Keras integration, such as saving the model summary and parameters used for fitting the model. See the integration changelog for details.