Skip to contents

Transfer Learning with Keras Applications

Source: vignettes/applications.Rmd
applications.Rmd

Introduction

Transfer learning is a powerful technique where a model developed for one task is reused as the starting point for a model on a second task. It is especially popular in computer vision, where pre-trained models like ResNet50, which were trained on the massive ImageNet dataset, can be used as powerful, ready-made feature extractors.

The kerasnip package makes it easy to incorporate these pre-trained Keras Applications directly into a tidymodels workflow. This vignette will demonstrate how to:

  1. Define a kerasnip model that uses a pre-trained ResNet50 as a frozen base layer.
  2. Add a new, trainable classification “head” on top of the frozen base.
  3. Tune the hyperparameters of the new classification head using a standard tidymodels workflow.

Setup

First, we load the necessary packages.

library(kerasnip)
library(tidymodels)
#> ── Attaching packages ────────────────────────────────────── tidymodels 1.5.0 ──
#> ✔ broom        1.0.13     ✔ recipes      1.3.3 
#> ✔ dials        1.4.4      ✔ rsample      1.3.2 
#> ✔ dplyr        1.2.1      ✔ tailor       0.1.0 
#> ✔ ggplot2      4.0.3      ✔ tidyr        1.3.2 
#> ✔ infer        1.1.0      ✔ tune         2.1.0 
#> ✔ modeldata    1.5.1      ✔ workflows    1.3.0 
#> ✔ parsnip      1.6.0      ✔ workflowsets 1.1.1 
#> ✔ purrr        1.2.2      ✔ yardstick    1.4.0
#> ── Conflicts ───────────────────────────────────────── tidymodels_conflicts() ──
#> ✖ purrr::discard() masks scales::discard()
#> ✖ dplyr::filter()  masks stats::filter()
#> ✖ dplyr::lag()     masks stats::lag()
#> ✖ recipes::step()  masks stats::step()
library(keras3)
#> 
#> Attaching package: 'keras3'
#> The following object is masked from 'package:yardstick':
#> 
#>     get_weights
#> The following object is masked from 'package:infer':
#> 
#>     generate

Data Preparation

We’ll use the CIFAR-10 dataset, which consists of 60,000 32x32 color images in 10 classes. keras3 provides a convenient function to download it.

The ResNet50 model was pre-trained on ImageNet, which has a different set of classes. Our goal is to fine-tune it to classify the 10 classes in CIFAR-10.

# Load CIFAR-10 dataset
cifar10 <- dataset_cifar10()

# Separate training and test data
x_train <- cifar10$train$x
y_train <- cifar10$train$y
x_test <- cifar10$test$x
y_test <- cifar10$test$y

# Rescale pixel values from [0, 255] to [0, 1]
x_train <- x_train / 255
x_test <- x_test / 255

# Convert outcomes to factors for tidymodels
y_train_factor <- factor(y_train[, 1])
y_test_factor <- factor(y_test[, 1])

# For tidymodels, it's best to work with data frames.
# We'll use a list-column to hold the image arrays.
train_df <- tibble::tibble(
  x = lapply(seq_len(nrow(x_train)), function(i) x_train[i, , , , drop = TRUE]),
  y = y_train_factor
)

test_df <- tibble::tibble(
  x = lapply(seq_len(nrow(x_test)), function(i) x_test[i, , , , drop = TRUE]),
  y = y_test_factor
)

# Use a smaller subset for faster vignette execution
train_df_small <- train_df[1:500, ]
test_df_small <- test_df[1:100, ]

Functional API with a Pre-trained Base

The standard approach for transfer learning is to use the Keras Functional API. We will define a model where: 1. The base is a pre-trained ResNet50, with its final classification layer removed (include_top = FALSE). 2. The weights of the base are frozen (trainable = FALSE) so that only our new layers are trained. 3. A new classification “head” is added, consisting of a flatten layer and a dense output layer.

Define Layer Blocks 

# Input block: shape is determined automatically from the data
input_block <- function(input_shape) {
  layer_input(shape = input_shape)
}

# ResNet50 base block
resnet_base_block <- function(tensor) {
  # The base model is not trainable; we use it for feature extraction.
  resnet_base <- application_resnet50(
    weights = "imagenet",
    include_top = FALSE
  )
  resnet_base$trainable <- FALSE
  resnet_base(tensor)
}

# New classification head
flatten_block <- function(tensor) {
  tensor |> layer_flatten()
}

output_block_functional <- function(tensor, num_classes) {
  tensor |> layer_dense(units = num_classes, activation = "softmax")
}

Create the kerasnip Specification

We connect these blocks using create_keras_functional_spec().

create_keras_functional_spec(
  model_name = "resnet_transfer",
  layer_blocks = list(
    input = input_block,
    resnet_base = inp_spec(resnet_base_block, "input"),
    flatten = inp_spec(flatten_block, "resnet_base"),
    output = inp_spec(output_block_functional, "flatten")
  ),
  mode = "classification"
)

Fit and Evaluate the Model

Now we can use our new resnet_transfer() specification within a tidymodels workflow.

spec_functional <- resnet_transfer(
  fit_epochs = 5,
  fit_validation_split = 0.2
) |>
  set_engine("keras")

rec_functional <- recipe(y ~ x, data = train_df_small)

wf_functional <- workflow() |>
  add_recipe(rec_functional) |>
  add_model(spec_functional)

fit_functional <- fit(wf_functional, data = train_df_small)
#> Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/resnet/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5
#>        0/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0s/step  540672/94765736 ━━━━━━━━━━━━━━━━━━━━ 8s 0us/step 7888896/94765736 ━━━━━━━━━━━━━━━━━━━━ 1s 0us/step21512192/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step30916608/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step43696128/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step54009856/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step66289664/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step77144064/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step89997312/94765736 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step94765736/94765736 ━━━━━━━━━━━━━━━━━━━━ 1s 0us/step
#> 16/16 - 3s - 183ms/step

# Evaluate on the test set
predictions <- predict(fit_functional, new_data = test_df_small)
#> 4/4 - 2s - 577ms/step
bind_cols(predictions, test_df_small) |>
  accuracy(truth = y, estimate = .pred_class)
#> # A tibble: 1 × 3
#>   .metric  .estimator .estimate
#>   <chr>    <chr>          <dbl>
#> 1 accuracy multiclass      0.18

Even with a small dataset and few epochs, the pre-trained features from ResNet50 give us a reasonable starting point for accuracy.

Conclusion

This vignette demonstrated how kerasnip bridges the world of pre-trained Keras applications with the structured, reproducible workflows of tidymodels.

The Functional API is the most direct way to perform transfer learning by attaching a new head to a frozen base model.

This approach allows you to leverage the power of deep learning models that have been trained on massive datasets, significantly boosting performance on smaller, domain-specific tasks.