Skip to content

Releases: lightly-ai/lightly

API update for working with delegated access

Delegated Access and Documentation Updates

20 Jul 15:31
a3ba3f6
Compare
Choose a tag to compare

Documentation updates and API connections

Documentation Updates and Improved Video Loading

05 Jul 07:25
7de764e
Compare
Choose a tag to compare

PIRL and more API helpers

28 Jun 09:33
00ff462
Compare
Choose a tag to compare

Refresh docs and more API helpers

21 Jun 14:44
f2241f0
Compare
Choose a tag to compare

Examples for MAE and better docs

Masked Autoencoder, SSL debug utilities

09 Jun 14:55
b119a57
Compare
Choose a tag to compare

Masked Autoencoders

We implemented the paper Masked Autoencoders Are Scalable Vision Learners. https://arxiv.org/abs/2111.06377 is suggesting that a masked auto-encoder (similar to pre-training on NLP) works very well as a pretext task for self-supervised learning. See #721 for more details. Thanks to @Atharva-Phatak for helping us figure out a good implementation method.

Collapse detection helper for SimSiam

We added a helper for detecting a collapsing SimSiam network. See https://ar5iv.labs.arxiv.org/html/2011.10566#S4.SS1 for more details.

Plot positive and negative example images

We added a helper to plot positive and negative example images, which also allows seeing what the augmentations do. See #818 for more details.

Models

Video Dataset initialisation improvements

24 May 11:41
5137f84
Compare
Choose a tag to compare

Video Dataset initialisation speedup

When initialising a LightlyDataset on a directory with videos, all frames in all videos have to be counted to know the number of frames in the dataset and their filenames. This process now uses multihreading over videos and can thus be much faster.

Video Dataset initialisation bugfix

We fixed a bug that the number of frames was estimated wrongly based on the length of the video when using the pyav backend.

Video Dataset initialisation progress bar

When initialising the video dataset, a progress bar over the videos is shown. This is helpful information for datasets with many videos.

Models

Decoupled Contrastive Learning, Model heads with default parameters

17 May 14:50
9b016b2
Compare
Choose a tag to compare

Decoupled Contrastive Learning

We have implemented the Decoupled Contrastive Learning (DCL) loss. It allows faster training with smaller batch sizes than other self-supervised learning models.

Documentation: https://docs.lightly.ai/examples/dcl.html
Decoupled Contrastive Learning paper: https://arxiv.org/abs/2110.06848

Model heads with default parameters

All model heads have now default parameters following the values of the original papers.

Create custom metadata config

Custom metadata in the Lightly Platform can now easily be configured via ApiWorkflowClient.create_custom_metadata_config()

Progress bar for video dataset initialization

Constructing a LightlyDataset with large video datasets can take long, as all frames in all videos have to be counted. We added a tqdm progress bar for it.

Models