fastai

fastai

fastai is a layered API on top of PyTorch. It popularized Transfer Learning and good defaults (One Cycle Policy).

When to Use

• Rapid Deep Learning Prototyping: Building vision, NLP, and tabular models with layered abstractions on top of PyTorch.
• Transfer Learning with State-of-the-Art Defaults: Automated discriminative learning rates, 1cycle scheduling, and mixup data augmentation.
• Computer Vision Tasks: Image classification, semantic segmentation, and object detection with minimal boilerplate.
• Learning Rate Optimization: Finding optimal learning rates automatically using lr_find().

Quick Start

from fastai.vision.all import *

# Train state-of-the-art vision classifier in 4 lines
path = untar_data(URLs.PETS)/'images'
def is_cat(x): return x[0].isupper()

dls = ImageDataLoaders.from_name_func(path, get_image_files(path), valid_pct=0.2,
                                     seed=42, label_func=is_cat, item_tfms=Resize(224))

learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)

Core Concepts

Computer Vision Classifier with DataBlock API

Building an image classification pipeline with automated transforms:

from fastai.vision.all import *

# Define flexible data pipeline
data_block = DataBlock(
    blocks=(ImageBlock, CategoryBlock),
    get_items=get_image_files,
    splitter=RandomSplitter(valid_pct=0.2, seed=42),
    get_y=parent_label,
    item_tfms=Resize(460),
    batch_tfms=aug_transforms(size=224, min_scale=0.75)
)

# Load data from directory
dls = data_block.dataloaders('/path/to/dataset', bs=64)

# Create transfer learning learner with pre-trained ResNet/ConvNeXt
learn = vision_learner(dls, resnet50, metrics=accuracy)

# Find optimal learning rate
suggested_lr = learn.lr_find()
print(f"Optimal Learning Rate: {suggested_lr.valley}")

# Fine-tune using 1cycle policy
learn.fine_tune(epochs=4, base_lr=suggested_lr.valley)

Tabular Deep Learning with Categorical Embeddings