v0.0.1.dev13

Notable Changes:

• new Auto-Encoders:
  • Ae
  • TripletAe (Ae version of TripletVae)
  • AdaAe (Ae version of AdaVae)
  • AdaNegTripletAe (Ae version of AdaNegTripletVae)
• custom dataset MNIST example in the docs

Breaking Changes

• flattened disent.frameworks.vae and disent.frameworks.ae modules, unsupervised, weaklysupervised, and supervised submodules no longer exist.
• remove latent parameter classes from VAEs, VAEs now directly encode distributions with the encode_dists() function, this simplified a lot of other code.
• Datasets now only return 'x' in the observation dictionary if an augment is specified, ~5% performance boost
• some dependencies are optional, more work is still required to minimise dependencies
• Removed sample_random_traversal_factors, sample_random_cycle_factors from StateSpace and replaced with generic function sample_random_factor_traversal
• renamed all autoencoders AE to Ae

Other Changes:

• hdf5 dataset performance fix, now up to 5x faster when not loaded into memory
• all Auto-Encoders have new config options to disable the augment loss, recon loss, or detach the decoder so that no loss flows back through the encoder. VAEs can additionally have the regularisation loss disabled.
• new laplace latent distribution, can be specified in VAE configs.
• triplet loss helper functions
• flatness components metric helper functions for use elsewhere: compute_linear_score, compute_axis_score
• FftKernel augment module inheriting from torch.nn.Module, applies a channel-wise convolution to the input.
• to_standardised_tensor fix for non-PIL.Image.Image inputs
• more math helper functions:
  • torch_normalize normalise values along an axis between 0 and 1
  • torch_mean_generalized now supports the keepdim argument
• disent.visualise.visualise_module removed old redundant code adapted from disentanglement_lib
• disent.visualise.visualise_util additions
  • make_image_grid and make_animated_image_grid auto-detect border colour from input dtype
  • replaced cycle_factor with get_factor_traversal that accepts different modes: interval and cycle
• cleaned up experiments

++ many more additions and minor fixes ++

v0.0.1.dev12

Large Release

• utility additions
  • dct
  • kernels: gaussian + box
  • conv2d channel wise
  • differentiable sorting, spearman rank loss
• ground truth dataset with factors
• more reconstruction losses
  • kernel reconstruction losses
  • recon loss fixes
  • parameterised recon losses
• scaled hard averaging for adatvae and adanegtvae
• DataOverlapRankVAE - uses differentiable sorting to optimise spearman rank correlation coefficient instead of triplet loss
• DataOverlapTripletVAE - fixes, simplifications, moved out triplet mining
• removed unnecessary metric values
• Conv64Alt encoder and decoder that support normalisation layers for faster convergence
• FFT gaussian and box blur augments
• more experiment schedules
• more experiments

And much more...

v0.0.1.dev11

• fixed init files

v0.0.1.dev10

frameworks

• simplified ada frameworks
• moved schedules out of ada frameworks into configs
• extra kl divergence modes

metrics

• combined flatness components
  • axis alignment ratio
  • linearity ratio
  • incorrect swap ratio

experiments

• existing configs should be frozen -- changes should be added to experiment scripts below
• helper script
• experiment scripts

more

• and much more

v0.0.1.dev9

metrics

• flatness components
  • reworked linearity component, now uses PCA to measure linearity along single arbitrary basis, and variance of embeddings to measure linearity along axis.

dataset wrappers

• new random dist dataset
  • only triplets have some sort of order, otherwise everything is sampled randomly
• RandomEpisodeDataset now has RandomDataset as parent

torch util - math

• PCA functions

v0.0.1.dev8

• renamed flatness components metric (originally dual flatness)

v0.0.1.dev7

metrics

• dual flatness metric
  • measures linearity and ordering of latent traversals
  • ordering over randomly sampled embeddings compared to ground truth factors

utils

• torch math helper functions
  • Covariance matrix
  • Pearson's correlation matrix
  • Spearman's rank correlation matrix
  • Generalised mean (p from -inf to inf, special cases for harmonic, geometric, arithmetic, min, max, quadratic means)

state spaces

• fixes for pos_to_idx and idx_to_pos with array sizes with more than 1 dimension (excluding last factor or idx dim)

v0.0.1.dev6

new frameworks

• Beta-TCVAE (probably wrong, needs verification & correct loss scaling)

frameworks

• DFC-VAE input fixes & support for 1 channel

v0.0.1.dev5

new frameworks

• DIP-VAE
• Info-VAE
• EXPERIMENTAL: Data Overlap TVAE

frameworks

• rewrite of frameworks extending from single VAE class, hooks are now made available for easy overrides. Removed lots of duplicate code.

datasets

• Fully random paired datasets

v0.0.1.dev4

cfg

• renamed loss_rediction mode batch_mean to mean_sum
• changed default loss_reduction mode to mean
• changed default beta to match new loss_reduction mode

flatness metric

• added average angle along traversals
• simplified greatly

reconstruction loss

• more reconstruction losses based off of distributions
  • bernoulli
  • continuous_bernoulli
  • normal

shedules

• new schedules:
  • NoopSchedule: does absolutely nothing!
  • CosineWaveSchedule: smooth cosine wave
• Adjusted arguments of most schedules

bugs

• fixed various runtime bugs
  • logging crash for W&B
  • disabled checkpointing in trainer
  • flatness metric did not support axis size < 2