Implementation:Fastai Fastbook Fine Tune

Knowledge Sources	fastbook fastai docs Learner.fine_tune
Domains	Deep_Learning, Transfer_Learning, Optimization
Last Updated	2026-02-09 17:00 GMT

Overview

Concrete tools for training a transfer-learning model provided by Learner.fine_tune, Learner.fit_one_cycle, and Learner.unfreeze from fastai.callback.schedule.

Description

The fine_tune method implements the complete two-phase transfer learning training protocol in a single call: it freezes the body, trains the head for a specified number of warmup epochs, then unfreezes the body and trains the full model with discriminative learning rates for the main epochs.

For practitioners who need more control, fit_one_cycle and unfreeze can be used separately to implement custom training schedules with explicit learning rate ranges.

Usage

Call learn.fine_tune(epochs) as the primary training method after creating a cnn_learner. Use fit_one_cycle and unfreeze when you need to customize the number of frozen epochs, the discriminative learning rate range, or when you want to interleave lr_find calls between phases.

Code Reference

Source Location

Repository: fastbook
File: translations/cn/02_production.md (lines 402-414), translations/cn/05_pet_breeds.md (lines 648-715)

Signature

# All-in-one fine-tuning
Learner.fine_tune(
    epochs,              # Number of unfrozen training epochs
    base_lr=None,        # Base learning rate (auto-detected if None)
    freeze_epochs=1,     # Number of frozen epochs for head warmup
    lr_mult=100,         # LR multiplier between lowest and highest layer group
    pct_start=0.3,       # Percentage of training spent in warmup phase
    div=5.0,             # Initial LR = base_lr / div during warmup start
    div_final=1e5,       # Final LR = base_lr / div_final at end of annealing
    wd=None,             # Weight decay
    moms=None,           # Momentum range
    cbs=None             # Additional callbacks
)

# Manual one-cycle training
Learner.fit_one_cycle(
    n_epoch,             # Number of epochs
    lr_max=None,         # Maximum learning rate (or slice for discriminative LRs)
    div=25.0,            # LR at start = lr_max / div
    div_final=1e5,       # LR at end = lr_max / div_final
    pct_start=0.25,      # Fraction of training spent in warmup
    wd=None,             # Weight decay
    moms=None,           # Momentum range (default: (0.95, 0.85, 0.95))
    cbs=None             # Additional callbacks
)

# Unfreeze all model parameters
Learner.unfreeze()

Import

from fastai.vision.all import Learner, cnn_learner
# fine_tune, fit_one_cycle, and unfreeze are methods on Learner

I/O Contract

Inputs

Name	Type	Required	Description
epochs	int	Yes	Number of epochs to train the unfrozen model (fine_tune) or total epochs (fit_one_cycle)
base_lr	float	No	Base learning rate; auto-detected from lr_find if None
freeze_epochs	int	No	Number of epochs to train with body frozen (default: 1)
lr_mult	int	No	Ratio between the highest and lowest layer group learning rates (default: 100)
lr_max	float or slice	No	Maximum LR for fit_one_cycle; use slice(low, high) for discriminative rates
n_epoch	int	Yes	Number of training epochs for fit_one_cycle

Outputs

Name	Type	Description
trained model	Learner (mutated in place)	The Learner with updated model weights after training
training log	printed table	Epoch-by-epoch table showing train_loss, valid_loss, and specified metrics

Usage Examples

Basic Usage: fine_tune (Chapter 2)

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)
learn.fine_tune(4)
# Output:
# epoch  train_loss  valid_loss  error_rate  time
# 0      0.163754    0.021388    0.005413    00:14  (frozen)
# 0      0.058871    0.025816    0.009472    00:19  (unfrozen)
# 1      0.038910    0.018753    0.006766    00:19
# 2      0.024956    0.016284    0.004060    00:19
# 3      0.017429    0.016641    0.005413    00:19

Manual Two-Phase Training (Chapter 5)

from fastai.vision.all import *

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

# Phase 1: Train only the head
learn.fit_one_cycle(3, 3e-3)

# Phase 2: Unfreeze and train with discriminative LRs
learn.unfreeze()

# Find appropriate LR for unfrozen model
lr_min, lr_steep = learn.lr_find()

# Train with lower LR for body, higher for head
learn.fit_one_cycle(6, lr_max=slice(1e-6, 1e-4))

Custom Freeze Epochs and Learning Rate

from fastai.vision.all import *

learn = cnn_learner(dls, resnet50, metrics=[accuracy, error_rate])

# Use lr_find to determine the base learning rate
lr_min, lr_steep = learn.lr_find()

# Fine-tune with 2 frozen warmup epochs and custom base LR
learn.fine_tune(5, base_lr=lr_min, freeze_epochs=2)

Related Pages

Implements Principle

Principle:Fastai_Fastbook_Fine_Tuning

Requires Environment

Uses Heuristic

Page Connections

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

Principle

Implementation

Heuristic

Environment