fastai: Learner, DataLoaders & Transfer Learning

fastai is a deep learning library built on top of PyTorch, providing high-level APIs that accomplish common tasks (image classification, NLP, tabular data) with dramatically less boilerplate -- while still allowing access to raw PyTorch when its higher-level abstractions aren't enough.

Training an Image Classifier

from fastai.vision.all import *

# Sets up the full data pipeline: finds images, splits 20% for
# validation, resizes to a consistent size (required for batching)
dls = ImageDataLoaders.from_folder(
    path,
    valid_pct=0.2,
    item_tfms=Resize(224),
    batch_tfms=aug_transforms(),  # augmentation applied per-batch on GPU
)

# vision_learner defaults to a PRETRAINED model -- transfer learning
# accessible with minimal code, no manual weight downloading needed
learn = vision_learner(dls, resnet34, metrics=accuracy)

# fine_tune(): trains new final layers first (frozen backbone), then
# unfreezes and trains everything at a lower rate -- a proven recipe
# packaged into one method call
learn.fine_tune(3)

The Data Block API

# Declaratively configures each pipeline step -- flexible enough for
# many dataset organizations, far less code than a manual PyTorch
# Dataset class
dblock = DataBlock(
    blocks=(ImageBlock, CategoryBlock),
    get_items=get_image_files,
    splitter=RandomSplitter(valid_pct=0.2),
    get_y=parent_label,
    item_tfms=Resize(224),
)
dls = dblock.dataloaders(path)

Finding a Good Learning Rate

# Trains briefly while increasing the learning rate, plotting loss
# vs. rate -- avoids guessing or expensive full-training trial-and-error
learn.lr_find()

Beyond Vision: Tabular & Text

# Same high-level philosophy applied to tabular data -- categorical
# columns become learned embeddings, continuous columns are normalized
dls = TabularDataLoaders.from_df(
    df,
    cat_names=['occupation', 'workclass'],
    cont_names=['age', 'hours_per_week'],
    y_names='salary',
)
learn = tabular_learner(dls, metrics=accuracy)