Implementation:Fastai Fastbook Cnn Learner

Knowledge Sources	fastbook fastai docs vision_learner API
Domains	Computer_Vision, Deep_Learning, Transfer_Learning
Last Updated	2026-02-09 17:00 GMT

Overview

Concrete tool for creating a transfer-learning Learner with a pretrained CNN backbone and custom classification head provided by fastai.vision.learner.cnn_learner (also available as vision_learner in newer fastai versions).

Description

The cnn_learner function combines three operations in a single call:

Loads a pretrained architecture (e.g., ResNet34) from torchvision with ImageNet weights.
Removes the original classification head and replaces it with a new head sized for the target number of categories (inferred from the DataLoaders vocabulary).
Wraps everything into a Learner object that manages the training loop, loss function, optimizer, and metrics.

The resulting Learner has its body frozen by default, so initial training only updates the head weights.

Usage

Call cnn_learner after creating DataLoaders and before any training calls (lr_find, fine_tune, fit_one_cycle). Choose the architecture based on the complexity of your task: ResNet34 is the recommended default; use ResNet50 or larger for harder problems.

Code Reference

Source Location

Repository: fastbook
File: translations/cn/01_intro.md (line 291), translations/cn/05_pet_breeds.md (lines 628-830)

Signature

cnn_learner(
    dls,                # DataLoaders object
    arch,               # Model architecture (e.g., resnet34, resnet50)
    metrics=None,       # List of metrics to track (e.g., [accuracy, error_rate])
    pretrained=True,    # Whether to use pretrained ImageNet weights
    loss_func=None,     # Loss function (auto-detected from DataLoaders if None)
    opt_func=Adam,      # Optimizer function
    lr=0.001,           # Default learning rate
    splitter=None,      # Function to split model into parameter groups
    cbs=None,           # Additional callbacks
    path=None,          # Path for saving models
    model_dir='models', # Directory name for saved models
    wd=None,            # Weight decay
    wd_bn_bias=False,   # Apply weight decay to batch norm and bias params
    train_bn=True,      # Train batch norm layers even when frozen
    moms=(0.95, 0.85, 0.95),  # Momentum range for one-cycle training
    n_out=None          # Number of output classes (auto-detected if None)
)

Import

from fastai.vision.all import cnn_learner, resnet34, resnet50, accuracy, error_rate

I/O Contract

Inputs

Name	Type	Required	Description
dls	DataLoaders	Yes	DataLoaders object containing training and validation data streams
arch	callable	Yes	Torchvision model constructor (e.g., resnet34, resnet50, resnet18)
metrics	list	No	List of metric functions to compute after each epoch (e.g., [accuracy, error_rate])
pretrained	bool	No	Use ImageNet pretrained weights (default: True)
loss_func	callable	No	Loss function; auto-detected as CrossEntropyLoss for single-label classification
n_out	int	No	Number of output classes; auto-detected from DataLoaders vocabulary if None

Outputs

Name	Type	Description
learn	Learner	A Learner object with pretrained body (frozen) and randomly initialized head, ready for training

Usage Examples

Basic Usage: Cat vs Dog Classifier (Chapter 1)

from fastai.vision.all import *

path = untar_data(URLs.PETS) / 'images'

dls = ImageDataLoaders.from_name_func(
    path, get_image_files(path),
    valid_pct=0.2, seed=42,
    label_func=lambda x: x[0].isupper(),
    item_tfms=Resize(224)
)

learn = cnn_learner(dls, resnet34, metrics=error_rate)

Pet Breed Classifier (Chapter 5)

from fastai.vision.all import *

path = untar_data(URLs.PETS) / 'images'

pets = DataBlock(
    blocks=(ImageBlock, CategoryBlock),
    get_items=get_image_files,
    splitter=RandomSplitter(valid_pct=0.2, seed=42),
    get_y=using_attr(RegexLabeller(r'^(.+)_\d+.jpg$'), 'name'),
    item_tfms=Resize(460),
    batch_tfms=aug_transforms(size=224, min_scale=0.75)
)
dls = pets.dataloaders(path, bs=64)

learn = cnn_learner(dls, resnet34, metrics=[accuracy, error_rate])

# The learner is now ready for lr_find() or fine_tune()
print(f'Model architecture: {learn.arch.__name__}')
print(f'Number of classes: {learn.dls.c}')
print(f'Class names: {learn.dls.vocab}')

Using a Larger Architecture

from fastai.vision.all import *

# For more complex classification tasks, use ResNet50
learn = cnn_learner(dls, resnet50, metrics=[accuracy, error_rate])

# Inspect the model head
print(learn.model[-1])  # Shows the custom head layers

Related Pages

Implements Principle

Principle:Fastai_Fastbook_Transfer_Learning

Requires Environment

Uses Heuristic

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment