train_model.py

import os
import json
import numpy as np
import tensorflow as tf
from tensorflow.keras.preprocessing import image_dataset_from_directory
from tensorflow.keras.applications import MobileNetV2  # Lightweight & effective for small datasets
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D, Dropout, BatchNormalization, Rescaling
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
from tensorflow.keras.optimizers import Adam, SGD
import matplotlib.pyplot as plt

# === Paths ===
DATA_DIR = "data/train"
MODEL_SAVE_PATH = "src/model/dog_breed_classifier.h5"
CLASS_NAMES_PATH = "src/model/class_names.json"
IMG_SIZE = (224, 224)
BATCH_SIZE = 32
SEED = 42

# === Load dataset ===
print("[INFO] Loading dataset...")
train_ds = image_dataset_from_directory(
    DATA_DIR,
    validation_split=0.2,
    subset="training",
    seed=SEED,
    image_size=IMG_SIZE,
    batch_size=BATCH_SIZE
)

val_ds = image_dataset_from_directory(
    DATA_DIR,
    validation_split=0.2,
    subset="validation",
    seed=SEED,
    image_size=IMG_SIZE,
    batch_size=BATCH_SIZE
)

# Save class names for inference
class_names = train_ds.class_names
num_classes = len(class_names)
print(f"[INFO] Classes found: {num_classes}")

with open(CLASS_NAMES_PATH, "w") as f:
    json.dump(class_names, f)

# === Data preprocessing & augmentation ===
resize_and_rescale = Sequential([
    Rescaling(1./255)
])

data_augmentation = Sequential([
    tf.keras.layers.RandomFlip("horizontal"),
    tf.keras.layers.RandomRotation(0.15),
    tf.keras.layers.RandomZoom(0.1)
])

AUTOTUNE = tf.data.AUTOTUNE
train_ds = train_ds.map(lambda x, y: (resize_and_rescale(x), y))
train_ds = train_ds.map(lambda x, y: (data_augmentation(x, training=True), y))
train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)

val_ds = val_ds.map(lambda x, y: (resize_and_rescale(x), y))
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

# === Compute class weights (to handle class imbalance) ===
print("[INFO] Computing class weights...")
y_train = np.concatenate([y.numpy() for _, y in train_ds], axis=0)
class_counts = np.bincount(y_train)
total = len(y_train)
class_weights = {i: total / (num_classes * count) for i, count in enumerate(class_counts)}
print("[INFO] Class weights applied.")

# === Build model ===
print("[INFO] Building model...")
base_model = MobileNetV2(input_shape=IMG_SIZE + (3,), include_top=False, weights='imagenet')
base_model.trainable = False

x = base_model.output
x = GlobalAveragePooling2D()(x)
x = BatchNormalization()(x)
x = Dropout(0.4)(x)
output = Dense(num_classes, activation='softmax')(x)

model = Model(inputs=base_model.input, outputs=output)
model.compile(
    optimizer=Adam(learning_rate=1e-4),
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

model.summary()

# === Callbacks ===
os.makedirs(os.path.dirname(MODEL_SAVE_PATH), exist_ok=True)
checkpoint = ModelCheckpoint(MODEL_SAVE_PATH, monitor='val_loss', save_best_only=True, verbose=1)
earlystop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.3, patience=2, verbose=1)

# === Phase 1: Train head ===
print("[INFO] Training model (frozen base)...")
history = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=15,
    class_weight=class_weights,
    callbacks=[checkpoint, earlystop, reduce_lr]
)

# === Phase 2: Fine-tune full model ===
print("[INFO] Fine-tuning entire model...")
base_model.trainable = True

model.compile(
    optimizer=SGD(learning_rate=1e-4, momentum=0.9),
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

fine_tune_epochs = 10
total_epochs = len(history.history["loss"]) + fine_tune_epochs

fine_tune_history = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=total_epochs,
    initial_epoch=history.epoch[-1] + 1,
    class_weight=class_weights,
    callbacks=[checkpoint, earlystop, reduce_lr]
)

# === Merge histories ===
for key in fine_tune_history.history:
    history.history[key] += fine_tune_history.history[key]

# === Plot training results ===
plt.figure(figsize=(12, 4))

plt.subplot(1, 2, 1)
plt.plot(history.history['loss'], label='Train Loss')
plt.plot(history.history['val_loss'], label='Val Loss')
plt.title("Loss")
plt.legend()

plt.subplot(1, 2, 2)
plt.plot(history.history['accuracy'], label='Train Acc')
plt.plot(history.history['val_accuracy'], label='Val Acc')
plt.title("Accuracy")
plt.legend()

plt.savefig("training_curves.png")
plt.show()

print(f"[DONE] Model saved to {MODEL_SAVE_PATH}")