Comparing Linear and Convolutional Models with TensorFlow in Python

Rate this post

💡 Problem Formulation: Today’s deep learning landscape offers various model architectures, and choosing the right one for your dataset can be pivotal. Imagine you have an image dataset and want to predict a numerical value related to each image. You are undecided between a simple linear regression model and a more complex convolutional neural network (CNN). This article discusses how to implement and compare these models using TensorFlow in Python, aiming to guide you towards a decision based on performance metrics.

Method 1: Setting Up the Experiment

This method involves setting up a controlled comparison experiment. Specify the dataset, preprocess the inputs for both models, and set aside a test set. Ensure that both models are trained and tested under the same conditions for fair comparison.

Here’s an example:

import tensorflow as tf
from sklearn.model_selection import train_test_split

# Load and preprocess your dataset
(images, targets) = load_data()
images = images / 255.0  # Normalize images

# Split the dataset
train_images, test_images, train_targets, test_targets = train_test_split(
    images, targets, test_size=0.2, random_state=42

Output is not applicable for the setup phase.

The above code is responsible for loading your image dataset, normalizing it by scaling pixel values to the [0,1] range, and then splitting it into training and testing sets. This setup is essential for fair comparison, as both models will be trained and evaluated on the same data.

Method 2: Training and Evaluating the Linear Model

In this method, we’ll define, compile, train, and evaluate a simple linear model using TensorFlow. The focus is on the model’s architecture—specifically a single dense layer—as well as using appropriate loss and metrics for regression.

Here’s an example:

linear_model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(image_height, image_width)),

linear_model.compile(optimizer='adam', loss='mse', metrics=['mae'])

history =, train_targets, epochs=10, validation_split=0.2)

linear_model.evaluate(test_images, test_targets)

The output would be the mean squared error (MSE) and mean absolute error (MAE) on the test set, indicating the model’s performance.

This snippet outlines the steps to create a simple linear model with TensorFlow’s Keras API, compile it with an optimizer and loss function suited for regression, and then train and evaluate it using the training and test data. The evaluation metrics directly inform us about the model’s regression performance.

Method 3: Training and Evaluating the Convolutional Model

Now, we’ll define a convolutional neural network, which is more suited for image data. We will compile, train, and evaluate this model. Focus on the convolutional layers and the architecture’s complexity, and note that the same loss and metrics will be used as in the linear model.

Here’s an example:

cnn_model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(32, (3, 3), activation='relu', input_shape=(image_height, image_width, 3)),
    tf.keras.layers.MaxPooling2D((2, 2)),
    tf.keras.layers.Dense(64, activation='relu'),

cnn_model.compile(optimizer='adam', loss='mse', metrics=['mae'])

history =, train_targets, epochs=10, validation_split=0.2)

cnn_model.evaluate(test_images, test_targets)

The output would also be the MSE and MAE on the test set, providing a direct performance comparison to the linear model.

This code snippet sets up a CNN with convolutional and pooling layers, followed by dense layers, then compiles, trains, and evaluates it using the same dataset. By comparing the resulting metrics with those of the linear model, we can determine which model performs better on the image regression task.

Bonus One-Liner Method 4: Using TensorFlow’s Built-in Metrics for Direct Comparison

TensorFlow offers a range of built-in metrics that we can use to compare models. We specifically look at the ‘mae’ for regression tasks, as it is easy to interpret and allows for direct model-to-model comparison post-training.

Here’s an example:

# Assume linear_model and cnn_model are already trained

linear_mae = linear_model.evaluate(test_images, test_targets, verbose=0)[1]
cnn_mae = cnn_model.evaluate(test_images, test_targets, verbose=0)[1]

print(f"Linear Model MAE: {linear_mae}")
print(f"CNN Model MAE: {cnn_mae}")

The output will give us the mean absolute error for both models, like so:

Linear Model MAE: 0.1234
CNN Model MAE: 0.0567

This quick comparison utilizes TensorFlow’s evaluation method to retrieve the mean absolute error for the linear model and the CNN, allowing for a brief and direct performance comparison in terms of error.


  • Method 1: Setting Up the Experiment: This foundational setup ensures fair testing conditions. However, it does not involve model performance comparison.
  • Method 2: Training and Evaluating the Linear Model: The linear model is easy to implement and quickly offers a benchmark for performance. Its simplistic nature, though, often underperforms on complex data such as images.
  • Method 3: Training and Evaluating the Convolutional Model: CNNs can outperform linear models on image data due to their ability to capture spatial hierarchies. Yet, they require more computation and are a heavier model structure.
  • Bonus One-Liner Method 4: Direct Comparison: Directly comparing metrics is quick and convenient. However, it relies on models being already trained and does not account for model complexity or training time.