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

Training neural networks on structured tabular data:

from fastai.tabular.all import *

df = pd.read_csv('customers.csv')

cat_names = ['workclass', 'education', 'marital-status', 'occupation']
cont_names = ['age', 'fnlwgt', 'education-num', 'hours-per-week']
procs = [Categorify, FillMissing, Normalize]

splits = RandomSplitter(valid_pct=0.2)(range_of(df))
to = TabularPandas(df, procs=procs, cat_names=cat_names, cont_names=cont_names, y_names='salary', splits=splits)

dls = to.dataloaders(bs=128)
learn = tabular_learner(dls, layers=[200, 100], metrics=accuracy)
learn.fit_one_cycle(5, 1e-2)

Exporting and Serving Model for Inference

Serializing the learner pipeline into a production artifact:

# Export model and all preprocessing pipelines
learn.export('classifier_model.pkl')

# In production server:
inference_learn = load_learner('classifier_model.pkl')
prediction, pred_idx, probabilities = inference_learn.predict('test_image.jpg')

print(f"Prediction: {prediction}, Confidence: {probabilities[pred_idx]:.4f}")

Common Patterns

Learning Rate Finder and Discriminative Learning Rates

Problem: Picking arbitrary learning rates leads to divergence or slow convergence.

Solution: Use lr_find() and fine-tune with slice learning rates:

# 1. Find optimal learning rate
lr_valley = learn.lr_find().valley

# 2. Unfreeze model and train with discriminative learning rates
learn.unfreeze()
learn.fit_one_cycle(4, slice(1e-5, lr_valley))

Best Practices

Do:

• Always run learn.lr_find() before training and use learn.fine_tune() for pre-trained weights.
• Use aug_transforms() with resize-presizing (item_tfms=Resize(460), batch_tfms=aug_transforms(size=224)) to minimize blur.
• Call learn.export() to package architecture, weights, and pre-processing transforms together.
• Inspect misclassified instances using ClassificationInterpretation.from_learner(learn).plot_top_losses().

Don't:

• Train pre-trained models with standard fit(); use fine_tune() to preserve backbone weights.
• Process single inference images in raw PyTorch without applying fastai's exported transform pipeline.
• Ignore class imbalance; supply custom weights to CrossEntropyLossFlat.

Troubleshooting

| Error | Cause | Solution | | :------------------------------------------------- | :-------------------------------------------------------------- | :------------------------------------------------------------------- | | CUDA out of memory in DataLoader | Batch size too large for GPU VRAM. | Reduce batch size: dls = ... bs=32 or bs=16. | | Can't get attribute '...' on <module '__main__'> | Custom labelling function not pickled during learn.export(). | Define functions in importable module or export before script exits. | | RuntimeError: Expected 4D tensor but got 3D | Single image passed without batch dimension to learn.predict. | Use learn.predict(img) directly; it handles batch expansion. |

References

• fast.ai