Data augmentation importance in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Data augmentation importance

Problem:We want to train a model to recognize handwritten digits using the MNIST dataset. The current model trains well on the training data but performs poorly on new images.

Current Metrics:Training accuracy: 98%, Validation accuracy: 85%, Validation loss: 0.45

Issue:The model is overfitting. It learns the training data too well but does not generalize to new data.

Your Task

Use data augmentation to reduce overfitting and improve validation accuracy to above 90% while keeping training accuracy below 95%.

You can only add data augmentation techniques during training.

Do not change the model architecture or optimizer.

Keep training epochs and batch size the same.

Hint 1

Hint 2

Hint 3

Solution

Computer Vision

import tensorflow as tf
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Load data
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# Normalize data
X_train = X_train.astype('float32') / 255.0
X_test = X_test.astype('float32') / 255.0

# Reshape for model input
X_train = X_train.reshape(-1, 28, 28, 1)
X_test = X_test.reshape(-1, 28, 28, 1)

# Define simple model
model = Sequential([
    Flatten(input_shape=(28,28,1)),
    Dense(128, activation='relu'),
    Dense(10, activation='softmax')
])

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Data augmentation setup
datagen = ImageDataGenerator(
    rotation_range=10,
    width_shift_range=0.1,
    height_shift_range=0.1,
    zoom_range=0.1
)

datagen.fit(X_train)

# Train model with augmentation
batch_size = 64
epochs = 10

history = model.fit(
    datagen.flow(X_train, y_train, batch_size=batch_size),
    epochs=epochs,
    validation_data=(X_test, y_test),
    steps_per_epoch=len(X_train) // batch_size
)

Added ImageDataGenerator with rotation, width/height shifts, and zoom augmentation.

Used datagen.flow to feed augmented images during training.

Kept model architecture and training parameters unchanged.

Results Interpretation

Before augmentation: Training accuracy was 98%, validation accuracy was 85%, showing overfitting.

After augmentation: Training accuracy dropped to 93%, validation accuracy improved to 91%, and validation loss decreased, indicating better generalization.

Data augmentation helps the model see more varied examples, reducing overfitting and improving performance on new data.

Bonus Experiment

Try adding more augmentation types like horizontal flips or brightness changes and observe the effect on validation accuracy.

💡 Hint

Be careful with flips on digits as some digits may look different when flipped; test carefully.

Practice

1. Why is data augmentation important in training computer vision models?

easy

A. It increases the variety of training images to help the model generalize better.

B. It reduces the size of the training dataset to speed up training.

C. It removes noisy images from the dataset automatically.

D. It guarantees 100% accuracy on the training data.

Solution

1. Step 1: Understand data augmentation purpose

   Data augmentation creates new images by slightly changing existing ones to increase variety.

2. Step 2: Connect augmentation to model learning

   More variety helps the model learn features that work on new, unseen images, improving generalization.

3. Final Answer:

   It increases the variety of training images to help the model generalize better. -> Option A

4. Quick Check:

   Data augmentation = better generalization [OK]

Hint: Think: more image variety means better learning [OK]

Common Mistakes:

• Confusing augmentation with data reduction
• Believing augmentation removes bad images
• Assuming augmentation guarantees perfect accuracy

2. Which of the following is a correct way to apply horizontal flip augmentation using Python's torchvision library?

easy

A. transforms.FlipHorizontal(prob=0.5)

B. transforms.HorizontalFlip(0.5)

C. transforms.RandomHorizontalFlip(p=0.5)

D. transforms.RandomFlipHorizontal()

Solution

1. Step 1: Recall torchvision syntax for horizontal flip

   The correct transform is RandomHorizontalFlip with a probability parameter p.

2. Step 2: Check each option's correctness

   Only transforms.RandomHorizontalFlip(p=0.5) matches the correct syntax and parameter name.

3. Final Answer:

   transforms.RandomHorizontalFlip(p=0.5) -> Option C

4. Quick Check:

   Correct torchvision flip syntax = transforms.RandomHorizontalFlip(p=0.5) [OK]

Hint: Look for 'RandomHorizontalFlip' with parameter p= [OK]

Common Mistakes:

• Using wrong class names like HorizontalFlip
• Incorrect parameter names like prob instead of p
• Missing the probability parameter

3. What will be the output shape of the augmented image after applying the following PyTorch transform?

transform = transforms.Compose([
  transforms.Resize((128, 128)),
  transforms.RandomRotation(30),
  transforms.ToTensor()
])
augmented_image = transform(original_image)

medium

A. [128, 3, 128]

B. [128, 128, 3]

C. [1, 128, 128]

D. [3, 128, 128]

Solution

1. Step 1: Analyze the transform steps

   Resize changes image to 128x128 pixels. RandomRotation keeps size same. ToTensor converts image to tensor with channels first.

2. Step 2: Determine tensor shape format

   PyTorch tensors from images have shape [channels, height, width]. For RGB images, channels=3.

3. Final Answer:

   [3, 128, 128] -> Option D

4. Quick Check:

   PyTorch image tensor shape = [channels, height, width] [OK]

Hint: PyTorch image tensors are channels first: [3, H, W] [OK]

Common Mistakes:

• Confusing channel order with height and width
• Assuming rotation changes image size
• Mixing up tensor shape formats

4. You wrote this augmentation code but get an error:

transform = transforms.Compose([
  transforms.RandomRotation(45),
  transforms.RandomHorizontalFlip(0.3),
  transforms.ToTensor()
])

What is the likely cause?

medium

A. RandomHorizontalFlip expects a keyword argument p, not a positional float.

B. RandomRotation requires integer degrees, not float.

C. ToTensor must come before RandomRotation.

D. Compose cannot combine these transforms.

Solution

1. Step 1: Check RandomHorizontalFlip usage

   RandomHorizontalFlip requires the probability parameter as a keyword argument p=, not a positional argument.

2. Step 2: Verify other transform usages

   RandomRotation accepts float degrees, ToTensor can be last, Compose supports these transforms.

3. Final Answer:

   RandomHorizontalFlip expects a keyword argument p, not a positional float. -> Option A

4. Quick Check:

   RandomHorizontalFlip(p=0.3) correct syntax [OK]

Hint: Check if transform params use correct keywords [OK]

Common Mistakes:

• Passing probability as positional argument
• Thinking rotation degrees must be integer
• Misordering transforms in Compose

5. You have a small dataset of 100 images for a classification task. Which data augmentation strategy will most likely improve your model's ability to recognize objects in new photos?

hard

A. Only resize images to a fixed size without any other changes.

B. Apply random flips, rotations up to 30 degrees, and brightness changes during training.

C. Add Gaussian noise to all images without any geometric transforms.

D. Train without augmentation but increase model layers.

Solution

1. Step 1: Consider dataset size and augmentation needs

   Small datasets benefit from augmentations that create varied views of images to prevent overfitting.

2. Step 2: Evaluate augmentation types

   Random flips, rotations, and brightness changes simulate real-world variations, improving generalization better than noise alone or no augmentation.

3. Final Answer:

   Apply random flips, rotations up to 30 degrees, and brightness changes during training. -> Option B

4. Quick Check:

   Varied augmentations = better generalization on small data [OK]

Hint: Use varied simple transforms for small datasets [OK]

Common Mistakes:

• Ignoring augmentation on small datasets
• Using only noise without geometric changes
• Relying on bigger models instead of data variety