This is a well-structured Python script for classifying dog breeds using EfficientNetB0 and Keras. This README file provides an overview of the code and its functionalities.
1. Imports:
- The code imports necessary libraries for image processing, deep learning, model building, evaluation, and visualization.
2. Constants:
IMG_SIZE: Defines the image size for pre-processing (224x224 pixels).NUM_CLASSES: Specifies the number of dog breeds you aim to classify (120 in this case).BATCH_SIZE: Defines the batch size for training (32 images per batch).EPOCHS: Sets the number of training epochs (iterations) (30, can be increased for better accuracy).
3. Paths:
TRAIN_DIR: Path to the directory containing your training data (dog breed images).TEST_DIR: Path to the directory containing your test data (dog breed images for evaluation).WEIGHTS_PATH: Path to the pre-trained EfficientNetB0 weights file (without the top classification layer).
4. Data Augmentation and Normalization:
train_datagen: Defines an ImageDataGenerator for training data. This performs data augmentation techniques like rotation, shifting, zooming, flipping, etc., to increase the size and diversity of training data and prevent overfitting.test_datagen: Defines an ImageDataGenerator for test data. This only performs rescaling (normalization) to a range of 0-1 for consistency with the model's input.
5. Loading and Augmenting Data:
train_generator: Usestrain_datagento load, augment, and prepare training data in batches for model training.test_generator: Usestest_datagento load, rescale, and prepare test data in batches for model evaluation.
6. Load Pre-trained Model:
base_model: Loads the pre-trained EfficientNetB0 model without the final classification layer (weights loaded fromWEIGHTS_PATH). EfficientNetB0 is a powerful convolutional neural network (CNN) pre-trained on a large image dataset.
7. Build the Model:
- The model takes the output of the pre-trained EfficientNetB0 as input.
- A GlobalAveragePooling2D layer reduces the spatial dimensions of the features.
- A Dense layer with 256 units and ReLU activation adds non-linearity.
- BatchNormalization helps improve training stability.
- Dropout (0.3) helps prevent overfitting.
- A final Dense layer with
NUM_CLASSESunits and softmax activation provides the probabilities for each dog breed class.
8. Compile the Model:
- Adam optimizer with a learning rate of 0.0001 is used.
- Categorical cross-entropy loss is used for multi-class classification.
- Accuracy is used as the evaluation metric.
9. Train the Model:
historystores the training and validation loss during each epoch.
10. Plot Training and Validation Loss:
- Visualizes the training and validation loss curves to monitor the learning process.
11. Evaluate the Model:
- Evaluates the model on the test data and prints the test loss and accuracy.
12. Predict Classes: Sample Output
- Predicts the class probabilities for each image in the test data.
- Chooses the class with the highest probability as the predicted class.
13. Generate Confusion Matrix:
- Creates a confusion matrix to visualize the performance of the model on each dog breed class.
14. Generate Classification Report:
- Prints a classification report summarizing the model's performance, including precision, recall, F1-score, and support for each class.