v0.1.7

Tsunami v0.1.7

Diff since v0.1.6

Merged pull requests:

• use Flux.get_device (#61) (@CarloLucibello)

v0.1.6

Tsunami v0.1.6

Diff since v0.1.5

Closed issues:

• add a get_device method (#49)

Merged pull requests:

• improve tests for Metal and add Foil example (#60) (@CarloLucibello)

v0.1.5

Tsunami v0.1.5

Diff since v0.1.4

Merged pull requests:

• Flux0.14 support + AMDGPU and Metal GPU support (#59) (@CarloLucibello)

v0.1.4

Tsunami v0.1.4

Diff since v0.1.3

Closed issues:

• add accumulate gradient (#38)
• add docs for gradient clipping (#39)
• save the model_state in the checkpoint instead of the full model (#55)

Merged pull requests:

• save Flux.state(model) in checkpoints (#56) (@CarloLucibello)
• accumgrad and clipgrad (#57) (@CarloLucibello)
• deprecation for fit(ckpt_path, trainer, ...) (#58) (@CarloLucibello)

v0.1.3

Tsunami v0.1.3

Diff since v0.1.2

Closed issues:

• add a NLP example (#45)
• use MLDatasets in the README example (#46)
• deprecate fit! in favor of fit (#53)

Merged pull requests:

• bert finetuning example (#48) (@CarloLucibello)
• simplify show and fix functor FluxModule (#51) (@CarloLucibello)
• Add Tsunami.fit + foreach_trainable (#52) (@CarloLucibello)
• make Tsunami.fit the main API (#54) (@CarloLucibello)

v0.1.2

Tsunami v0.1.2

Diff since v0.1.1

Closed issues:

• cleaner validation printing (#5)
• check licensing issues with pytorch lightning (#28)
• create a lightweight training framework (#43)

Merged pull requests:

• introduce the accelerator Foil (#44) (@CarloLucibello)
• more hooks + precision setting + other stuff (#47) (@CarloLucibello)

v0.1.1

Tsunami v0.1.1

Diff since v0.1.0

Closed issues:

• tag a first release (#16)
• reuse the progress bar, don't create a new one for each epoch (#32)
• add validate and test methods (#37)

Merged pull requests:

• overwrite progbar at each epoch (#36) (@CarloLucibello)
• implement Tsunami.test (#40) (@CarloLucibello)
• add validate and on_start_*_epoch hooks (#41) (@CarloLucibello)
• on_before_pullback_call (#42) (@CarloLucibello)

v0.1.0

Tsunami v0.1.0

Closed issues:

• change the name of the library? (#3)
• make fit_state its own type (#7)
• define show method for generic FluxModel (#8)
• no training progress bar is shown for small number of batches (#9)
• describe in pseudocode the fit! method (#11)
• add log function (#13)
• add seed_everything (#15)
• check and improve documentation (#18)
• deprecate training_epoch_end (#19)
• audit API (#20)
• account for batch_size when logging on_epoch (#22)
• make batch_idx optional in training_step etc... (#27)
• schedulers should become lr_schedulers? (#30)

Merged pull requests:

• Callbacks support (#10) (@CarloLucibello)
• create logging infrastructure (#21) (@CarloLucibello)
• [logging] weight with batchsize on_epoch reduction (#24) (@CarloLucibello)
• Tsunami.seed! (#26) (@CarloLucibello)
• change training_* to train_* etc.. (#29) (@CarloLucibello)