Quiz tổng hợp

1import tensorflow as tf
2from tensorflow.keras import layers, models
3
4model = models.Sequential([
5    # Block 1
6    layers.Conv2D(32, (3,3), padding='same', activation='relu', input_shape=(32,32,3)),
7    layers.BatchNormalization(),
8    layers.Conv2D(32, (3,3), padding='same', activation='relu'),
9    layers.BatchNormalization(),
10    layers.MaxPooling2D((2,2)),
11    layers.Dropout(0.25),
12
13    # Block 2
14    layers.Conv2D(64, (3,3), padding='same', activation='relu'),
15    layers.BatchNormalization(),
16    layers.Conv2D(64, (3,3), padding='same', activation='relu'),
17    layers.BatchNormalization(),
18    layers.MaxPooling2D((2,2)),
19    layers.Dropout(0.25),
20
21    # Block 3
22    layers.Conv2D(128, (3,3), padding='same', activation='relu'),
23    layers.BatchNormalization(),
24    layers.MaxPooling2D((2,2)),
25    layers.Dropout(0.25),
26
27    # Classifier
28    layers.Flatten(),
29    layers.Dense(256, activation='relu'),
30    layers.BatchNormalization(),
31    layers.Dropout(0.5),
32    layers.Dense(10, activation='softmax')
33])
34
35model.compile(
36    optimizer='adam',
37    loss='sparse_categorical_crossentropy',
38    metrics=['accuracy']
39)
40
41model.summary()
42# Tổng ~600K params, đạt ~90% accuracy trên CIFAR-10

1from tensorflow.keras.applications import ResNet50
2from tensorflow.keras import layers, models
3
4# 1. Load pre-trained ResNet50 (không lấy lớp FC cuối)
5base_model = ResNet50(weights='imagenet', include_top=False, input_shape=(224,224,3))
6
7# 2. Đóng băng các lớp pre-trained
8base_model.trainable = False
9
10# 3. Thêm classifier mới
11model = models.Sequential([
12    base_model,
13    layers.GlobalAveragePooling2D(),
14    layers.Dense(256, activation='relu'),
15    layers.Dropout(0.5),
16    layers.Dense(4, activation='softmax')
17])
18
19# 4. Train classifier mới (feature extraction)
20model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
21model.fit(train_data, epochs=10)
22
23# 5. Fine-tuning: mở khóa một số lớp cuối của ResNet50
24base_model.trainable = True
25for layer in base_model.layers[:-20]:  # Đóng băng tất cả trừ 20 lớp cuối
26    layer.trainable = False
27
28model.compile(
29    optimizer=tf.keras.optimizers.Adam(1e-5),  # Learning rate nhỏ
30    loss='sparse_categorical_crossentropy',
31    metrics=['accuracy']
32)
33model.fit(train_data, epochs=10)

1import torch
2import torch.onnx
3
4# 1. Export sang ONNX
5dummy_input = torch.randn(1, 3, 224, 224)
6torch.onnx.export(
7    model,
8    dummy_input,
9    "model.onnx",
10    opset_version=13,
11    input_names=['input'],
12    output_names=['output'],
13    dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}}
14)
15
16# 2. Post-training Quantization (PyTorch)
17quantized_model = torch.quantization.quantize_dynamic(
18    model,
19    {torch.nn.Linear},  # Quantize Linear layers
20    dtype=torch.qint8
21)
22
23# 3. So sánh kích thước
24import os
25torch.save(model.state_dict(), "original.pth")
26torch.save(quantized_model.state_dict(), "quantized.pth")
27print(f"Original: {os.path.getsize('original.pth') / 1e6:.1f} MB")
28print(f"Quantized: {os.path.getsize('quantized.pth') / 1e6:.1f} MB")
29
30# 4. ONNX Runtime inference
31import onnxruntime as ort
32session = ort.InferenceSession("model.onnx")
33result = session.run(None, {"input": dummy_input.numpy()})