v0.6.1rc6

ENH: implement validation sampling (#112)

* ENH: implement validation sampling
* play with CI issues (timeout with httpx)

v0.6.1rc3

Improvements around realistic crowdsourcing simulations  (#110)

See https://github.com/stsievert/salmon/pull/110#issue-610255620 for details.

v0.6.0

Breaking changes

• API: Endpoint /init_exp has been renamed to /init (#106).
• API: defaults for salmon.triplets.offline.OfflineEmbedding changed. (#100)
• API: class RR for "round robin" has been renamed to ARR for "active round robin." (#109)
• API: salmon.triplets.algs has been renamed to salmon.triplets.samplers. (#109)

Improvements since v0.5.2
The main improvement is better showing how to enable good adaptive performance with Salmon (#100). This showed clear benefits from using adaptive algorithms – fewer responses will be required with the sampler ARR. Some changes around this include the following:

• ENH/MAINT: better setting of max_epochs for adaptive runners. This increases the response rate at which adaptive and random sampling perform the same (2 responses/sec → 10 responses/sec) (#100).
• MAINT: change defaults on adaptive runners (mostly, change default optimizer from SGD to Adadelta)
• MAINT: keep number of queries for RR to be 1 * n (don't post the database if not required).

These changes enabled showing the improvements:

• DOC: show how response rate affects Salmon embedding performance for adaptive algs, and provide a reference configuration (#100).

Additionally, some other smaller features have been implemented:

• DOC: firm up salmon.triplets.offline.OfflineEmbedding (make new properties, better document, etc).
• API: allow users to provide authorization credentials (#106)
• API/DOC: improve attributes/documentation for salmon.triplets.offline.OfflineEmbedding in salmon/triplets/offline.py (#100)

Maintenance

• DOC: provide methods for troubleshooting launching process (#104)
• DOC/MAINT: allow the database to be properly reset (#107)
• DOC: clean documentation. Add more classes, better descriptions, reference papers, etc (#100)
• DOC: docs only changed on releases (#108).

v0.5.2

This is a small bug-fix release.

• MAINT, TST: It fixes a typo on ZIP file upload w/o numeric filenames (#101)

This release is based off a single PR, #101. Other features in that PR include:

• MAINT: Cleans examples/datasets (remove unused code, better organize) (#101)
• MAINT/TST: Adds random state back in – but only for initialization (and test it too) (#101)
• TST: fix zip file tests so they run with pytest (→ remove xfail) (#101)
• TST: test uploads of each example in examples/ (#101)

v0.5.1

This is primarily a bug fix release:

• BUG: remove bug where same random state used to ask queries (#96)

Other improvements:

• MAINT: don't return repeated queries from round robin sampling
• MAINT: Files are ordered if numeric filenames are given in targets.zip (#95).
• MAINT: Offline embedding improved (#95; adaptive learning rate method used).
• DOC: Better show adaptive example

v0.5.0

What's new in this release?

• ENH: Adaptive sampling now works as expected for triplets (#92)
• API: an endpoint to get the embedding is implemented (#93).
• MAINT: initialization (#91, #90).
• DOC: better documentation around offline embeddings (#92).

What's new since the v0.4.0 release?
These are the most important changes in the minor releases.

• API: implements OGD and sample weights for offline embeddings
• MAINT: remove skorch for 40× faster & lower memory usage (mostly #87, #90 is slightly related)
• DOC improvements (#71, #57)
• MAINT: Backend performance (#66, #74) and stability (#71, #78) improvements.
• VIZ: show more information on dashboards, including embeddings (#67) and algorithm timings (#71).

v0.4.5

This release limits memory usage, to help prevent Dask workers and the database from becoming overloaded.

v0.4.4

This release makes the following improvements:

• API: implements OGD and sample weights for offline embeddings
• DOC: warns experimentalists to use HTTP and new AMI
• MAINT: Remove skorch. This means Salmon embeddings are much quicker, about 40× faster than before with Skorch. (no data preprocessing is required, only a torch.from_numpy call).
• MAINT: Train for longer after each response received (200 epochs, not 1).

All of these were implemented in #87.

v0.4.3

DOC: detail how to create offline embeddings (#86)

* ENH: allow sample_weights to be specified
* DOC: offline responses

v0.4.2

New redirect