Movatterモバイル変換

luz 0.5.1

fixed a bug preventing additional arguments passed topredict to be forwarded tomodel$predict(#157)

Added mixed precision callback. (#127)
Added support for torch iterable datasets. (#135)
Fixed a bug when trying to resume models trained with learning rateschedulers. (#137)
Added support for learning rate schedulers that take the currentloss as arguments. (#140)
Added French translation of luz messages. (@cregouby #148)

drop_last=TRUE is now the default for trainingdataloaders created by luz (when eg. you pass a list or a torch datasetas data input) (#117)
The default profile callback no longer tracks intra step timings asit adds a non ignorable overhead. (#125)

Added support for arm Mac’s and the MPS device. (#104)
Refactor checkpointing in luz - we now also serialize optimizerstate and callbacks state. (#107)
Added aluz_callback_autoresume() allowing to easilyresume training runs that might have crashed. (#107)
Added theluz_callback_resume_from_checkpoint()allowing one to resume a training run from a checkpoint file.(#107)
Users can now chose if metrics should be called on both training andvalidation, only training or only validation. Seeluz_metric_set() for more information. (#112)
Improved how errors raised on user code, eg while calling metrics orcallbacks are raised. This helps a lot when debuging errors in callbacksand metrics. (#112)
loss_fn is now a field of the context, thus callbackscan override it when needed. (#112)
luz_callback_mixup now supports therun_valid andauto_loss arguments. (#112)
ctx now aliases to the defaultopt andopt_name when a single optimizer is specified (ie. mostcases) (#114)
Addedtfevents callback for logging the loss andgetting weights histograms. (#118)
You can now specify metrics to be evaluated duringevaluate. (#123)

Bug fix:acceleratorscpu argument isalways respected. (#119)
Handledrlang andggplot2 deprecations.(#120)
Better handling of metrics environments.
Faster garbage collection of dataloaders iterators, so we use lessmemory. (#122)
Much faster loss averaging at every step. Can have hight influencein training times for large number of iterations per epoch. (#124)

lr_finder() now by default divides the range betweenstart_lr andend_lr into log-spaced intervals,following the fast.ai implementation. Cf. Sylvain Gugger’s post:https://sgugger.github.io/how-do-you-find-a-good-learning-rate.html. Theprevious behavior can be achieved passinglog_spaced_intervals=FALSE to the function. (#82,@skeydan)
plot.lr_records() now in addition plots anexponentially weighted moving average of the loss (again, see SylvainGugger’s post), with a weighting coefficient of0.9 (whichseems a reasonable value for the default setting of 100learning-rate-incrementing intervals). (#82,@skeydan)

Added MixUp callback and helper loss function and functional logic.(#82,@skeydan).
Added aluz_callback_gradient_clip inspired by FastAI’simplementation. (#90)
Added abackward argument tosetupallowing one to customize howbackward is called for theloss scalar value. (#93)
Added theluz_callback_keep_best_model() to reload theweights from the best model after training is finished. (#95)

Allow users to provide the minimum and maximum number of epochs whencallingfit.luz_module_generator(). Removedctx$epochs from context object and replaced it withctx$min_epochs andctx$max_epochs (#53,@mattwarkentin).
Early stopping will now only occur if the minimum number of trainingepochs has been met (#53,@mattwarkentin).
Addedcuda_index argument toacceleratorto allow selecting an specific GPU when multiple are present (#58,@cmcmaster1).
Implementedlr_finder (#59,@cmcmaster1).
We now handle different kinds of data arguments passed tofit using theas_dataloader() method(#66).
valid_data can now be scalar value indicating theproportion ofdata that will be used for fitting. This onlyworks ifdata is a torch dataset or a list. (#69)
You can now supplydataloader_options tofit to pass additional information toas_dataloader(). (#71)
Implemented theevaluate function allowing users to getmetrics from a model in a new dataset. (#73)

Fixed bug in CSV logger callback that was saving the logs as a spacedelimited file (#52,@mattwarkentin).
Fixed bug in the length of the progress bar for the validationdataset (#52,@mattwarkentin).
Fixed bugs in early stopping callback related to them not workingproperly whenpatience = 1 and when they are specifiedbefore other logging callbacks. (#76)

ctx$data now refers to the current in usedata instead of always refering toctx$train_data. (#54)
Refactored thectx object to make it safer and avoidreturing it in the output. (#73)