2021-03-19
Python Client
Embeddings are fully optional in the python client now
Previously, if you had no embeddings available and were operating in anonymous mode, you needed to provide arbitrary values to prevent processing failures. You no longer need to provide any embedding values, even in anonymous mode.
Python client supports multiple simultaneous instances / process
Multiple instances of the aquariumlearning
client package can now run simultaneously on the same host.
Python client allows editing issue state / name
Basic issue state management is now exposed in the python client's IssueManager
.
App
Project & Dataset Management
Project / Dataset pages (+ favorites)
All projects can now be viewed on a dedicated projects page. Project settings and datasets can be viewed on each individual project page. Each dataset and inference set has a dedicated information page that has settings, metadata, and links to explore the data.
Project + Dataset Metadata
You can now attach JSON formatted metadata to projects, datasets, and inference sets by going to their respective Settings pages.
Create Projects UI
You can now create a project in the web app (although you will still need to use the Python client to upload your datasets and inference sets).
Update label class map colors via UI
You can now update your label class map colors in the web app, either via the Display Settings in Explore view or through your Project Settings page.
Project preview images
You can now upload custom images as a preview image for your project so that they are visually distinct on the projects page.
Data Exploration
Left-side navbar
New vertical bar on the left side of screen to navigate between Cards Grid, Histogram, Embeddings, and Metrics. Switching between frames and crops on the Embedding view has also been moved to this sidebar as well.
Permalinks: Direct linking to dataset + view + filter
The currently active view and filter on the data exploration page will now be stored in the url so that links can be shared to specific views/tabs.
Autocomplete for user metadata + issue filters
In the Data Exploration view, we now have autocomplete for categorical user metadata fields and issue names, to make it easier to find what you’re looking for.
Breadcrumbs
New UI to show project -> dataset -> inference sets hierarchy on the explore/dataset pages. You can click into detailed Project, Dataset, and Inference Set pages via these breadcrumbs.
Embeddings
Better embedding point size defaults
Point sizes in the embeddings view now have improved size scaling based on their density.
Grid View for Embeddings
When you lasso multiple points in the Embeddings view, you can now choose to view them in a grid instead of clicking through the selected-points carousel.
Color embeddings by custom metrics
In addition to coloring embeddings by Aquarium's computed metrics, you can now view your data colored by your own custom metrics.
Color embeddings by numeric / continuous metadata
In addition to coloring embeddings by categorical metadata, you can now view your data colored according to numerical / continuous metadata values.
Max confidence display threshold
You can adjust this threshold to visually hide hide-confidence inferences. (Note that this is used purely for display. There is a min confidence threshold that is used for metrics and confusion matrix calculations).
Issues
Issue Membership on Cards
You can now tell at-a-glance the number of issues a card is a member of, and see the full list in the expanded detailed view.
Issue Cards respect “Cards per row”
Cards-per-row as set in Display Settings now also apply to the cards in issue views.
Metrics
Project-level default IOU + Conf thresholds
You can now set project-level IOU + confidence thresholds, which will be used for matching your ground truth (GT) and inferences, and for calculating metrics/confusion matrices. You can still adjust these dynamically in the Data Exploration view without affecting project defaults.
Stratified Metrics
You can now filter the metrics view based on per-label attributes. For example, if you attach an attribute called “label_size,” you can evaluate performance on just large or small labels independently. For more details, check out the docs here.
Numbers rendering on confusion matrix
You can now choose to show the absolute numbers in confusion matrix cells.
Search + Many Classes on metrics view
When your dataset has many classes (>30) it will show up as a table instead of a full confusion matrix in the metrics tab. Search bars have been added for finding the specific class you are looking for in the tables.
Project-level settings for metrics behavior on frames with no ground truth.Include_no_gt as a project-level toggle
The project settings UI now exposes an include_no_gt
flag which can be updated on existing projects. This changes how Aquarium's default metrics handle frames with zero labels (the default behavior is to match pycocotools / tensorflow object detection api and ignore frames with no support).
Other
Clientside Bucket Secrets
If your data is in a private storage bucket and you provide s3://
or gs://
data paths, you can set your corresponding AWS or GCS creds so that your assets are viewable in the app. Note that these creds remain clientside and aren’t ever exposed to Aquarium’s servers.
Inference contrast coloring mode
You can toggle “Use Inference Contrast Coloring” in display settings to view images with a high contrast coloring scheme between ground truth labels (green) and inferences (pink).
User generate-able API keys
Users can now create API keys to use with the aquariumlearning
client package. Navigate to your username in the top right corner of the page, User Settings, and then Generate New API Key.
“Go to Embedding”
Allows taking currently selected cards in a non-embedding view over to the embeddings view, with the same selections preserved.
Highly visible confidence values
If you have an inference set selected in the Data Exploration view , you will now be able to see confidence numbers within the colored label “legends” in each frame.
Better sorting/searching (projects page, issue summaries)
You can more easily sort and filter though lists of projects, datasets, and issues.
Display Settings persist per project
The display settings you choose for a particular project (e.g. number of cards per row) will persist the next time you view it. Display settings are currently persisted to your local device settings.
Last updated