120 Dog Breeds, more than 10,000 Images: Deep Learning Tutorial for dogs classification

# This section imports the necessary libraries for the script, which include numpy, cv2 for image processing, and pandas for data handling.
import numpy as np  # Import numpy for array handling
import cv2          # Import OpenCV for image processing
import pandas as pd # Import pandas for handling CSV files and dataframes

# Defines global constants for image size and full-size dimensions to be used when resizing and processing images.
IMAGE_SIZE = (331, 331)       # Target size for resizing images
IMAGE_FULL_SIZE = (331, 331, 3)  # Full-size dimension including color channels

# Sets paths to the image folder and the labels CSV file which contains image IDs and corresponding dog breed labels.
trainMyImageFolder = "E:/Data-sets/Dog Breed Identification/train"  # Path to folder containing training images

# Loads the CSV file containing image IDs and dog breeds, then performs some basic data exploration to understand its contents.
df = pd.read_csv('E:/Data-sets/Dog Breed Identification/labels.csv')  # Load CSV file containing image IDs and labels
print("head of labels:")
print("================")
print(df.head())            # Print the first few rows of the DataFrame to check data format
print(df.describe())        # Print summary statistics of the DataFrame

print("Group by labels: ")
grouplables = df.groupby("breed")["id"].count()  # Group data by breed and count occurrences
print(grouplables.head(10))  # Display the counts of the top 10 breeds

# Reads and displays a sample image using OpenCV to verify the image path and loading process.
imgPath = "E:/Data-sets/Dog Breed Identification/train/00a366d4b4a9bbb6c8a63126697b7656.jpg"  # Path to a sample image
img = cv2.imread(imgPath)  # Read image using OpenCV
# cv2.imshow("img", img)    # Uncomment to display the image in a window
# cv2.waitKey(0)            # Wait for any key press to close the window

# This section loops through the DataFrame to load each image and its label, resizing images and storing them in lists.
allImages = []  # Initialize list to store image data
allLabes = []   # Initialize list to store breed labels

import os       # Import os for handling file paths

# Loop through each row in the DataFrame, process each image and label
for ix, (image_name, breed) in enumerate(df[['id', 'breed']].values):
    img_dir = os.path.join(trainMyImageFolder, image_name + '.jpg')  # Construct full image path
    print(img_dir)    # Print the image path for debugging

    img = cv2.imread(img_dir)    # Read image from the directory
    resized = cv2.resize(img, IMAGE_SIZE, interpolation=cv2.INTER_AREA)  # Resize image to the specified size
    allImages.append(resized)    # Append resized image to the list
    allLabes.append(breed)       # Append label to the list

print(len(allImages))  # Print total number of images loaded
print(len(allLabes))   # Print total number of labels loaded


# This section saves the lists of images and labels as .npy files for future use in model training.
print("Save the data:")
np.save("e:/temp/allDogsImages.npy", allImages)  # Save all images as a numpy array
np.save("e:/temp/allDogsLables.npy", allLabes)   # Save all labels as a numpy array
print("Finished saving the data...")

# This section imports the necessary libraries for the script, which include numpy, cv2 for image processing, and pandas for data handling. import numpy as np  # Import numpy for array handling import cv2          # Import OpenCV for image processing import pandas as pd # Import pandas for handling CSV files and dataframes  # Defines global constants for image size and full-size dimensions to be used when resizing and processing images. IMAGE_SIZE = (331, 331)       # Target size for resizing images IMAGE_FULL_SIZE = (331, 331, 3)  # Full-size dimension including color channels  # Sets paths to the image folder and the labels CSV file which contains image IDs and corresponding dog breed labels. trainMyImageFolder = "E:/Data-sets/Dog Breed Identification/train"  # Path to folder containing training images  # Loads the CSV file containing image IDs and dog breeds, then performs some basic data exploration to understand its contents. df = pd.read_csv('E:/Data-sets/Dog Breed Identification/labels.csv')  # Load CSV file containing image IDs and labels print("head of labels:") print("================") print(df.head())            # Print the first few rows of the DataFrame to check data format print(df.describe())        # Print summary statistics of the DataFrame  print("Group by labels: ") grouplables = df.groupby("breed")["id"].count()  # Group data by breed and count occurrences print(grouplables.head(10))  # Display the counts of the top 10 breeds  # Reads and displays a sample image using OpenCV to verify the image path and loading process. imgPath = "E:/Data-sets/Dog Breed Identification/train/00a366d4b4a9bbb6c8a63126697b7656.jpg"  # Path to a sample image img = cv2.imread(imgPath)  # Read image using OpenCV # cv2.imshow("img", img)    # Uncomment to display the image in a window # cv2.waitKey(0)            # Wait for any key press to close the window  # This section loops through the DataFrame to load each image and its label, resizing images and storing them in lists. allImages = []  # Initialize list to store image data allLabes = []   # Initialize list to store breed labels  import os       # Import os for handling file paths  # Loop through each row in the DataFrame, process each image and label for ix, (image_name, breed) in enumerate(df[['id', 'breed']].values):     img_dir = os.path.join(trainMyImageFolder, image_name + '.jpg')  # Construct full image path     print(img_dir)    # Print the image path for debugging      img = cv2.imread(img_dir)    # Read image from the directory     resized = cv2.resize(img, IMAGE_SIZE, interpolation=cv2.INTER_AREA)  # Resize image to the specified size     allImages.append(resized)    # Append resized image to the list     allLabes.append(breed)       # Append label to the list  print(len(allImages))  # Print total number of images loaded print(len(allLabes))   # Print total number of labels loaded   # This section saves the lists of images and labels as .npy files for future use in model training. print("Save the data:") np.save("e:/temp/allDogsImages.npy", allImages)  # Save all images as a numpy array np.save("e:/temp/allDogsLables.npy", allLabes)   # Save all labels as a numpy array print("Finished saving the data...")

# Import necessary libraries
import numpy as np  # For numerical operations and array handling

# Set image parameters
IMAGE_SIZE = (331, 331)  # Target image dimensions for resizing
IMAGE_FULL_SIZE = (331, 331, 3)  # Full image size including color channels
batchSize = 8  # Number of images per training batch

# Load preprocessed image and label data
allImages = np.load("e:/temp/allDogsImages.npy")  # Load images as numpy array
allLabes = np.load("e:/temp/allDogsLables.npy")   # Load labels as numpy array

# Display the shapes of loaded data to verify successful loading
print(allImages.shape)  # Print shape of image array
print(allLabes.shape)   # Print shape of labels array

# Convert labels from text to integer values for model processing
print(allLabes)  # Print original labels to see text values

from sklearn.preprocessing import LabelEncoder  # For encoding text labels into integers
le = LabelEncoder()  # Initialize the label encoder
integerLables = le.fit_transform(allLabes)  # Fit encoder and transform text labels to integers
print(integerLables)  # Display encoded integer labels

# Count the number of unique breeds (categories)
numOfCategories = len(np.unique(integerLables))  # Count unique integer labels, e.g., 120 breeds
print(numOfCategories)  # Print the total number of categories

# Convert integer labels to categorical format (one-hot encoding)
from tensorflow.keras.utils import to_categorical  # For one-hot encoding
allLablesForModel = to_categorical(integerLables, num_classes=numOfCategories)  # Convert to categorical
print(allLablesForModel)  # Display the one-hot encoded labels

# Normalize the image pixel values from 0-255 to 0-1
allImagesForModel = allImages / 255.0  # Scale image values to [0, 1]

# Split the data into training and testing sets
from sklearn.model_selection import train_test_split  # For splitting data

print("Before split train and test:")

# Split images and labels into training (70%) and testing (30%) sets
X_train, X_test, y_train, y_test = train_test_split(
    allImagesForModel, allLablesForModel, test_size=0.3, random_state=42)

# Print shapes of training and testing data arrays to verify split
print("X_train, X_test, y_train, y_test -----> shapes:")
print(X_train.shape)  # Shape of training images
print(X_test.shape)   # Shape of testing images
print(y_train.shape)  # Shape of training labels
print(y_test.shape)   # Shape of testing labels

# Free up memory by deleting unused variables
del allImages
del allLabes
del integerLables
del allImagesForModel

# Build the model using a pre-trained NASNetLarge model

from tensorflow.keras.layers import Dense, Flatten  # Import layers for model building
from tensorflow.keras.models import Model  # Import Model class
from tensorflow.keras.applications.nasnet import NASNetLarge  # Import NASNetLarge

# Load NASNetLarge model with pre-trained weights, excluding top layers
myModel = NASNetLarge(input_shape=IMAGE_FULL_SIZE, weights='imagenet', include_top=False)

# Freeze all layers of the NASNetLarge model to prevent re-training
for layer in myModel.layers:
    layer.trainable = False  # Set layer to non-trainable
    print(layer.name)  # Print layer name (optional)

# Add a flattening layer to prepare features for the final dense layer
plusFlattenLayer = Flatten()(myModel.output)  # Flatten the output from NASNetLarge

# Add a dense layer for final predictions with softmax activation for multi-class classification
prediction = Dense(numOfCategories, activation='softmax')(plusFlattenLayer)

# Create a new model with NASNetLarge as the base and custom dense layer on top
model = Model(inputs=myModel.input, outputs=prediction)

# Uncomment to print a summary of the model architecture
# print(model.summary())

from tensorflow.keras.optimizers import Adam  # Import Adam optimizer

# Set learning rate and initialize optimizer
lr = 1e-4  # Learning rate for optimizer
opt = Adam(lr)  # Initialize Adam optimizer with defined learning rate

# Compile the model with categorical crossentropy loss and accuracy metric
model.compile(
    loss='categorical_crossentropy',  # Loss function for multi-class classification
    optimizer=opt,  # Use Adam optimizer
    metrics=['accuracy']  # Track accuracy during training
)

# Calculate the number of steps per epoch and validation steps
stepsPerEpoch = np.ceil(len(X_train) / batchSize)  # Number of steps in each training epoch
validationSteps = np.ceil(len(X_test) / batchSize)  # Number of steps for validation

# Set up callbacks for early stopping and model checkpointing
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping

best_model_file = "e:/temp/dogs.h5"  # Path to save the best model

callbacks = [
    ModelCheckpoint(best_model_file, verbose=1, save_best_only=True),  # Save best model during training
    ReduceLROnPlateau(monitor='val_loss', patience=3, factor=0.1, verbose=1, min_lr=1e-6),  # Reduce LR on plateau
    EarlyStopping(monitor='val_accuracy', patience=7, verbose=1)  # Stop training if no improvement in val accuracy
]

# Train the model
r = model.fit(
    X_train, y_train,
    validation_data=(X_test, y_test),
    epochs=30,  # Number of training epochs
    batch_size=batchSize,  # Batch size for training
    steps_per_epoch=stepsPerEpoch,  # Steps per epoch based on training set size
    validation_steps=validationSteps,  # Steps for validation based on testing set size
    callbacks=[callbacks]  # Include defined callbacks
)

# Import necessary libraries import numpy as np  # For numerical operations and array handling  # Set image parameters IMAGE_SIZE = (331, 331)  # Target image dimensions for resizing IMAGE_FULL_SIZE = (331, 331, 3)  # Full image size including color channels batchSize = 8  # Number of images per training batch  # Load preprocessed image and label data allImages = np.load("e:/temp/allDogsImages.npy")  # Load images as numpy array allLabes = np.load("e:/temp/allDogsLables.npy")   # Load labels as numpy array  # Display the shapes of loaded data to verify successful loading print(allImages.shape)  # Print shape of image array print(allLabes.shape)   # Print shape of labels array  # Convert labels from text to integer values for model processing print(allLabes)  # Print original labels to see text values  from sklearn.preprocessing import LabelEncoder  # For encoding text labels into integers le = LabelEncoder()  # Initialize the label encoder integerLables = le.fit_transform(allLabes)  # Fit encoder and transform text labels to integers print(integerLables)  # Display encoded integer labels  # Count the number of unique breeds (categories) numOfCategories = len(np.unique(integerLables))  # Count unique integer labels, e.g., 120 breeds print(numOfCategories)  # Print the total number of categories  # Convert integer labels to categorical format (one-hot encoding) from tensorflow.keras.utils import to_categorical  # For one-hot encoding allLablesForModel = to_categorical(integerLables, num_classes=numOfCategories)  # Convert to categorical print(allLablesForModel)  # Display the one-hot encoded labels  # Normalize the image pixel values from 0-255 to 0-1 allImagesForModel = allImages / 255.0  # Scale image values to [0, 1]  # Split the data into training and testing sets from sklearn.model_selection import train_test_split  # For splitting data  print("Before split train and test:")  # Split images and labels into training (70%) and testing (30%) sets X_train, X_test, y_train, y_test = train_test_split(     allImagesForModel, allLablesForModel, test_size=0.3, random_state=42)  # Print shapes of training and testing data arrays to verify split print("X_train, X_test, y_train, y_test -----> shapes:") print(X_train.shape)  # Shape of training images print(X_test.shape)   # Shape of testing images print(y_train.shape)  # Shape of training labels print(y_test.shape)   # Shape of testing labels  # Free up memory by deleting unused variables del allImages del allLabes del integerLables del allImagesForModel  # Build the model using a pre-trained NASNetLarge model  from tensorflow.keras.layers import Dense, Flatten  # Import layers for model building from tensorflow.keras.models import Model  # Import Model class from tensorflow.keras.applications.nasnet import NASNetLarge  # Import NASNetLarge  # Load NASNetLarge model with pre-trained weights, excluding top layers myModel = NASNetLarge(input_shape=IMAGE_FULL_SIZE, weights='imagenet', include_top=False)  # Freeze all layers of the NASNetLarge model to prevent re-training for layer in myModel.layers:     layer.trainable = False  # Set layer to non-trainable     print(layer.name)  # Print layer name (optional)  # Add a flattening layer to prepare features for the final dense layer plusFlattenLayer = Flatten()(myModel.output)  # Flatten the output from NASNetLarge  # Add a dense layer for final predictions with softmax activation for multi-class classification prediction = Dense(numOfCategories, activation='softmax')(plusFlattenLayer)  # Create a new model with NASNetLarge as the base and custom dense layer on top model = Model(inputs=myModel.input, outputs=prediction)  # Uncomment to print a summary of the model architecture # print(model.summary())  from tensorflow.keras.optimizers import Adam  # Import Adam optimizer  # Set learning rate and initialize optimizer lr = 1e-4  # Learning rate for optimizer opt = Adam(lr)  # Initialize Adam optimizer with defined learning rate  # Compile the model with categorical crossentropy loss and accuracy metric model.compile(     loss='categorical_crossentropy',  # Loss function for multi-class classification     optimizer=opt,  # Use Adam optimizer     metrics=['accuracy']  # Track accuracy during training )  # Calculate the number of steps per epoch and validation steps stepsPerEpoch = np.ceil(len(X_train) / batchSize)  # Number of steps in each training epoch validationSteps = np.ceil(len(X_test) / batchSize)  # Number of steps for validation  # Set up callbacks for early stopping and model checkpointing from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping  best_model_file = "e:/temp/dogs.h5"  # Path to save the best model  callbacks = [     ModelCheckpoint(best_model_file, verbose=1, save_best_only=True),  # Save best model during training     ReduceLROnPlateau(monitor='val_loss', patience=3, factor=0.1, verbose=1, min_lr=1e-6),  # Reduce LR on plateau     EarlyStopping(monitor='val_accuracy', patience=7, verbose=1)  # Stop training if no improvement in val accuracy ]  # Train the model r = model.fit(     X_train, y_train,     validation_data=(X_test, y_test),     epochs=30,  # Number of training epochs     batch_size=batchSize,  # Batch size for training     steps_per_epoch=stepsPerEpoch,  # Steps per epoch based on training set size     validation_steps=validationSteps,  # Steps for validation based on testing set size     callbacks=[callbacks]  # Include defined callbacks )

# Import necessary libraries
import tensorflow as tf  # TensorFlow for loading and using the trained model
import os  # For file path management
import numpy as np  # For numerical operations
import cv2  # For image processing

# Load the pre-trained model from a file
modelFile = "e:/temp/dogs.h5"  # Path to the saved model
model = tf.keras.models.load_model(modelFile)  # Load the model

# Uncomment to print a summary of the model architecture
# print(model.summary())

# Set input image size for the model
inputShape = (331, 331)  # Image dimensions expected by the model

# Load labels to get the breed names
allLabes = np.load("e:/temp/allDogsLables.npy")  # Load labels from file
categories = np.unique(allLabes)  # Get unique breed names

# Function to preprocess the input image
def prepareImage(img):
    resized = cv2.resize(img, inputShape, interpolation=cv2.INTER_AREA)  # Resize image to inputShape
    imgResult = np.expand_dims(resized, axis=0)  # Add batch dimension for model input
    imgResult = imgResult / 255.0  # Normalize pixel values to [0, 1]
    return imgResult

# Path to the test image
testImagePath = "TensorFlowProjects/Dog Breed Identification - Pre-trained - Works !!/Irish_Water_Spaniel1.png"
# Alternative test image paths (can uncomment and use one of these)
# testImagePath = "TensorFlowProjects/Dog Breed Identification - Pre-trained - Works !!/soft-coated_wheaten_terrier.jpg"

# Load the test image
img = cv2.imread(testImagePath)  # Read the image from file
imageForModel = prepareImage(img)  # Prepare image for the model

# Predict the breed of the dog in the image
resultArray = model.predict(imageForModel, verbose=1)  # Get prediction from the model
answers = np.argmax(resultArray, axis=1)  # Find the class with the highest probability

# Get the breed name corresponding to the predicted class
print(answers)  # Print the integer label of the predicted breed
text = categories[answers[0]]  # Retrieve breed name from categories array
print(text)  # Print the breed name

# Display the breed name on the image
font = cv2.FONT_HERSHEY_COMPLEX  # Set font style
cv2.putText(img, text, (0, 20), font, 0.5, (209, 19, 77), 2)  # Add text to the image

# Show the image with the breed name overlayed
cv2.imshow('img', img)  # Display the image in a new window
cv2.waitKey(0)  # Wait for a key press to close the window

# Save the annotated image
cv2.imwrite("e:/temp/img1.jpg", img)  # Save the image with breed name
cv2.destroyAllWindows()  # Close all OpenCV windows

# Import necessary libraries import tensorflow as tf  # TensorFlow for loading and using the trained model import os  # For file path management import numpy as np  # For numerical operations import cv2  # For image processing  # Load the pre-trained model from a file modelFile = "e:/temp/dogs.h5"  # Path to the saved model model = tf.keras.models.load_model(modelFile)  # Load the model  # Uncomment to print a summary of the model architecture # print(model.summary())  # Set input image size for the model inputShape = (331, 331)  # Image dimensions expected by the model  # Load labels to get the breed names allLabes = np.load("e:/temp/allDogsLables.npy")  # Load labels from file categories = np.unique(allLabes)  # Get unique breed names  # Function to preprocess the input image def prepareImage(img):     resized = cv2.resize(img, inputShape, interpolation=cv2.INTER_AREA)  # Resize image to inputShape     imgResult = np.expand_dims(resized, axis=0)  # Add batch dimension for model input     imgResult = imgResult / 255.0  # Normalize pixel values to [0, 1]     return imgResult  # Path to the test image testImagePath = "TensorFlowProjects/Dog Breed Identification - Pre-trained - Works !!/Irish_Water_Spaniel1.png" # Alternative test image paths (can uncomment and use one of these) # testImagePath = "TensorFlowProjects/Dog Breed Identification - Pre-trained - Works !!/soft-coated_wheaten_terrier.jpg"  # Load the test image img = cv2.imread(testImagePath)  # Read the image from file imageForModel = prepareImage(img)  # Prepare image for the model  # Predict the breed of the dog in the image resultArray = model.predict(imageForModel, verbose=1)  # Get prediction from the model answers = np.argmax(resultArray, axis=1)  # Find the class with the highest probability  # Get the breed name corresponding to the predicted class print(answers)  # Print the integer label of the predicted breed text = categories[answers[0]]  # Retrieve breed name from categories array print(text)  # Print the breed name  # Display the breed name on the image font = cv2.FONT_HERSHEY_COMPLEX  # Set font style cv2.putText(img, text, (0, 20), font, 0.5, (209, 19, 77), 2)  # Add text to the image  # Show the image with the breed name overlayed cv2.imshow('img', img)  # Display the image in a new window cv2.waitKey(0)  # Wait for a key press to close the window  # Save the annotated image cv2.imwrite("e:/temp/img1.jpg", img)  # Save the image with breed name cv2.destroyAllWindows()  # Close all OpenCV windows