NEWS.md

In fastai: Interface to 'fastai'

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

• minor fixes

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

• current stable version of fast.ai is 2.1.5
• lots of new callback ops
• freeze and unfreeze a model
• object detection module - icevision
• issue with exporting of a pickle file

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

• Initial release

fastai documentation built on June 22, 2024, 11:15 a.m.