Intermediate

Computer Vision with FastAI

Build state-of-the-art image classifiers in just a few lines of code using vision_learner, fine_tune, built-in data augmentation, and powerful interpretation tools.

Your First Image Classifier

This is the power of FastAI — a complete, state-of-the-art image classifier in 4 lines:

Python

from fastai.vision.all import *

# 1. Get data
path = untar_data(URLs.PETS)

# 2. Create DataLoaders
dls = ImageDataLoaders.from_name_re(
    path, get_image_files(path/'images'),
    pat=r'/([^/]+)_\d+.jpg$',
    item_tfms=Resize(224),
    batch_tfms=aug_transforms(size=224)
)

# 3. Create learner with pre-trained ResNet34
learn = vision_learner(dls, resnet34, metrics=error_rate)

# 4. Fine-tune!
learn.fine_tune(3)

What fine_tune() does automatically: (1) Freezes the pre-trained backbone, (2) trains only the new head for 1 epoch, (3) unfreezes everything, (4) trains the entire model with discriminative learning rates for your specified number of epochs. All best practices baked in!

The DataBlock API

For more control over data loading, use the flexible DataBlock API:

Python

dblock = 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)
)

dls = dblock.dataloaders(path/'images')
dls.show_batch(max_n=9)  # Visualize a batch

Data Augmentation

FastAI applies GPU-accelerated augmentation via aug_transforms():

Python

# Default augmentations (rotation, flip, zoom, warp, lighting)
batch_tfms = aug_transforms(size=224)

# Customize augmentations
batch_tfms = aug_transforms(
    size=224,
    mult=2.0,           # Increase augmentation intensity
    do_flip=True,
    flip_vert=False,     # Only horizontal flips
    max_rotate=15.0,
    max_zoom=1.2,
    max_lighting=0.3,
    max_warp=0.2,
    p_affine=0.75,
    p_lighting=0.75
)

Model Interpretation

FastAI includes powerful tools to understand what your model learned and where it makes mistakes:

Python

# Create interpretation object
interp = ClassificationInterpretation.from_learner(learn)

# Confusion matrix
interp.plot_confusion_matrix(figsize=(12, 12))

# Show top losses (images the model got most wrong)
interp.plot_top_losses(9, figsize=(15, 10))

# Most confused categories
interp.most_confused(min_val=5)

Choosing Architectures

Architecture	Accuracy	Speed	Best For
resnet18	Good	Very fast	Quick experiments, limited GPU
resnet34	Better	Fast	Good default choice
resnet50	Great	Moderate	Production models
convnext_tiny	Excellent	Moderate	Modern architecture
vit_small_patch16_224	Excellent	Slower	Vision Transformer

Next Up: Tabular Data

Learn how to apply FastAI's magic to structured/tabular data with TabularDataLoaders.

Next: Tabular Data →

← Installation Tabular Data →