2.2.2
- Updated installation function
install_fastai
2.2.1
- Fixed breaking changes during model training
- Updated installation function
install_fastai
2.0.9
- PyTorch version is 1.9
- lr_find bug is fixed
2.0.8
2.0.7
- bs_finder is fixed
- better visualization on Colab. Issue with fig size is fixed
2.0.6
- new function to load_learner
- unet_config is Deprecated
- while installing fast.ai Mac OS, first, it downloads PyTorch 1.8, then 1.7.1. It is fixed, now.
nn_module()
function allows to rename the model, e.g. summary(model)
nn_module()
will not move the model to GPU, if gpu
argument is FALSE (by default it is TRUE)
- custom loss functions with
nn_loss()
. Based on Kaggle notebook
2.0.5
install_fastai
no more supports extensions. They need to be installed separately by users.
- PyTorch was upgraded from 1.7.0 to 1.7.1.
2.0.4
- stick to fastaudio 0.1.3 (resolve dependencies)
- add
geom_point
for interactive visualization within RStudio
- add TPU module into fastai
2.0.3
2.0.2
- Hugging Face integration, prediction
one_batch()
ability to add more arguments
- no need to call
options(reticulate.useImportHook = FALSE)
DataBlock
automatically places data into cuda
if available
2.0.1
nn_module
for model construction
fix_fit
for disabling the training plot
- all the
fit
functions now return the training history
fastai 2.0.0