Ứng Dụng CNN và Transfer Learning

1import tensorflow as tf
2from tensorflow.keras import layers
3
4# Data Augmentation layer
5data_augmentation = tf.keras.Sequential([
6    layers.RandomFlip("horizontal"),
7    layers.RandomRotation(0.1),
8    layers.RandomZoom(0.1),
9])
10
11# Demo: 1 ảnh → 4 ảnh khác nhau
12import matplotlib.pyplot as plt
13
14# Tạo ảnh giả
15image = tf.random.uniform((1, 224, 224, 3), 0, 255)
16
17plt.figure(figsize=(10, 3))
18for i in range(4):
19    augmented = data_augmentation(image)
20    plt.subplot(1, 4, i+1)
21    plt.imshow(augmented[0].numpy().astype("uint8"))
22    plt.title(f"Augment {i+1}")
23    plt.axis("off")
24plt.show()

1import tensorflow as tf
2from tensorflow import keras
3from tensorflow.keras import layers
4
5# Load data từ thư mục
6# Cấu trúc: data/train/cat/, data/train/dog/
7train_ds = keras.utils.image_dataset_from_directory(
8    'data/train',
9    image_size=(224, 224),
10    batch_size=32,
11    validation_split=0.2,
12    subset='training',
13    seed=42
14)
15
16val_ds = keras.utils.image_dataset_from_directory(
17    'data/train',
18    image_size=(224, 224),
19    batch_size=32,
20    validation_split=0.2,
21    subset='validation',
22    seed=42
23)
24
25class_names = train_ds.class_names
26print(f"Classes: {class_names}")

1from tensorflow.keras.applications import ResNet50
2
3# Load ResNet50 (không lấy lớp classification cuối)
4base_model = ResNet50(
5    weights='imagenet',      # Pretrained trên ImageNet
6    include_top=False,       # Bỏ lớp cuối
7    input_shape=(224, 224, 3)
8)
9
10# FREEZE base model!
11base_model.trainable = False
12
13print(f"Base model layers: {len(base_model.layers)}")
14print(f"Trainable: {base_model.trainable}")

1# Data augmentation
2data_augmentation = keras.Sequential([
3    layers.RandomFlip("horizontal"),
4    layers.RandomRotation(0.1),
5    layers.RandomZoom(0.1),
6])
7
8# Xây model
9model = keras.Sequential([
10    # Input
11    layers.Input(shape=(224, 224, 3)),
12    
13    # Augmentation (tự động tắt khi inference)
14    data_augmentation,
15    
16    # Preprocessing cho ResNet
17    layers.Rescaling(1./127.5, offset=-1),
18    
19    # Pretrained base
20    base_model,
21    
22    # Custom classifier
23    layers.GlobalAveragePooling2D(),
24    layers.Dropout(0.5),
25    layers.Dense(len(class_names), activation='softmax')
26])
27
28model.summary()

1# Compile
2model.compile(
3    optimizer='adam',
4    loss='sparse_categorical_crossentropy',
5    metrics=['accuracy']
6)
7
8# Train
9history = model.fit(
10    train_ds,
11    validation_data=val_ds,
12    epochs=10,
13    callbacks=[
14        keras.callbacks.EarlyStopping(patience=3, restore_best_weights=True)
15    ]
16)

1# Sau khi train Feature Extraction xong...
2
3# Bước 1: Unfreeze top layers của base model
4base_model.trainable = True
5
6# Chỉ train 30 layers cuối
7for layer in base_model.layers[:-30]:
8    layer.trainable = False
9
10# Bước 2: Re-compile với learning rate RẤT NHỎ!
11model.compile(
12    optimizer=keras.optimizers.Adam(learning_rate=1e-5),  # 100x nhỏ hơn
13    loss='sparse_categorical_crossentropy',
14    metrics=['accuracy']
15)
16
17# Bước 3: Train tiếp
18history_fine = model.fit(
19    train_ds,
20    validation_data=val_ds,
21    epochs=5
22)

1# 🚀 QUICK START: Image Classification
2from tensorflow import keras
3from tensorflow.keras.applications import ResNet50
4
5# 1. Load data
6train_ds = keras.utils.image_dataset_from_directory(
7    'data/train', image_size=(224, 224))
8
9# 2. Build model
10base = ResNet50(weights='imagenet', include_top=False,
11                input_shape=(224, 224, 3))
12base.trainable = False  # FREEZE!
13
14model = keras.Sequential([
15    base,
16    keras.layers.GlobalAveragePooling2D(),
17    keras.layers.Dense(len(train_ds.class_names), activation='softmax')
18])
19
20# 3. Train
21model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', 
22              metrics=['accuracy'])
23model.fit(train_ds, epochs=10)
24
25# 4. Save
26model.save('my_classifier.keras')