v1.10.1

DiffEqFlux v1.10.1

Diff since v1.10.0

Merged pull requests:

• fix default adjoints for GPU (#240) (@ChrisRackauckas)
• Add GPU-based MNIST example to the docs (#242) (@ChrisRackauckas)

v1.10.0

DiffEqFlux v1.10.0

Diff since v1.9.0

Closed issues:

• Gradients w.r.t. scalar indexing of a DiffEqArray (#160)
• Store and return the current best loss from sciml_train on Flux.jl optimizers (#170)
• Broadcasting problem in Zygote.gradient with FastChain() (#171)
• Issue with batch training of NeuralODE with signals as inputs (#225)
• Docstrings and documentation for fast layers (#231)

Merged pull requests:

• Utilize view strided fixes (#152) (@ChrisRackauckas)
• CompatHelper: bump compat for "Optim" to "0.21" (#222) (@github-actions[bot])
• Documenter (#228) (@Emmanuel-R8)
• gigantic doc update (#232) (@ChrisRackauckas)
• Deploy Docs (#233) (@kanav99)
• Add MLSL global optimization (#234) (@AStupidBear)
• Gigantic documentation update 2 (#235) (@ChrisRackauckas)

v1.9.0

DiffEqFlux v1.9.0

Diff since v1.8.1

Closed issues:

• BBO breaks (#199)
• Force stopping BBO from callback fails (#205)

Merged pull requests:

• Add callback kwarg to BlackBoxOptim dispatch (#201) (@abhigupta768)
• See what the tests say (#213) (@ChrisRackauckas)
• Add TikTak global optimization to sciml_train (#217) (@ranjanan)
• Move TikTak to requires.jl (#219) (@ranjanan)
• Add QuadDIRECT global optimizer (#220) (@ranjanan)
• analytical tgrad for neural ODEs (#223) (@ChrisRackauckas)

v1.8.1

DiffEqFlux v1.8.1

Diff since v1.8.0

Merged pull requests:

• Failed when loss function returned more than one value to pass to cb (#203) (@ChrisDeGrendele)
• Adding kwarg for Optim options (#208) (@YingboMa)

v1.8.0

DiffEqFlux v1.8.0

Diff since v1.7.1

Closed issues:

• Theoretical underpinning (#193)
• Solving Forward Backward SDE (#197)
• FastChain in loss (#198)

Merged pull requests:

• improve logging of sciml_train (#194) (@AStupidBear)
• Save best solution rather than last one in sciml_train (#195) (@ChrisDeGrendele)
• Add BlackBoxOptim dispatch to sciml_train (#196) (@abhigupta768)

v1.7.1

DiffEqFlux v1.7.1

Diff since v1.7.0

Closed issues:

• Setup progress bar of sciml_train to be platform agnostic (#158)
• sciml_train fails on GPU (#182)
• sciml_train error when using NewtonTrustRegion optimizer (#185)
• Neural SDE does not train diffusion parameters (#189)

Merged pull requests:

• Provide default progress loggers and use new ProgressLogging API (#184) (@devmotion)
• Test DiffEqBase promotion (#188) (@ChrisRackauckas)
• fix and better test neural SDE on FastChain (#191) (@ChrisRackauckas)

v1.7.0

DiffEqFlux v1.7.0

Diff since v1.6.0

Closed issues:

• Singular mass matrices not compatible with hardcoded adjoints (#34)
• StackOverflowError on a simple example (#149)
• Setup progress bar of sciml_train to be platform agnostic (#158)

Merged pull requests:

• Add progress bar to sciml_train, fixes #158 (#180) (@abhigupta768)
• support second order (#181) (@ChrisRackauckas)

v1.6.0

DiffEqFlux v1.6.0

Diff since v1.5.0

Closed issues:

• Setup NLopt optimizers with sciml_train (#161)

Merged pull requests:

• Setup NLopt optimizers with sciml_train (#177) (@abhigupta768)

v1.5.0

DiffEqFlux v1.5.0

Diff since v1.4.0

Merged pull requests:

• Alternative training AD modes (#176) (@ChrisRackauckas)

v1.4.0

DiffEqFlux v1.4.0

Diff since v1.3.2

Closed issues:

• Implementing Neural DAE's (#124)
• sciml_train: unexpected: call back args out of sync with loss() return values (#166)
• concrete_solve not defined (#168)
• sciml_train NeuralODE has zygote error, when processing 2D arrays of batch data (#169)

Merged pull requests:

• [WIP] Add neural DAE support (#159) (@abhigupta768)
• Fix loss in callback to be computed before updating parameters (#172) (@ChrisRackauckas)
• Allow ForwardDiff inside of sciml_train calls (#175) (@ChrisRackauckas)