Load model Keras từ SavedModel format Python

Mô tả

SavedModel là format mặc định của TensorFlow (thư mục chứa .pb, variables/, assets/). Phù hợp cho production serving (TF Serving, TFLite, TF.js).

Lưu SavedModel

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

model = keras.Sequential([
    layers.Dense(64, activation='relu', input_shape=(10,)),
    layers.Dense(1)
])
model.compile(optimizer='adam', loss='mse')

# Dummy train
import numpy as np
model.fit(np.random.rand(100, 10), np.random.rand(100, 1), epochs=1, verbose=0)

# LƯU SAVEDMODEL (thư mục)
model.save('saved_model/my_model')  # Không có extension
print("SavedModel saved to saved_model/my_model/")

Cấu trúc thư mục SavedModel

saved_model/my_model/
├── assets/
├── variables/
│   ├── variables.data-00000-of-00001
│   └── variables.index
├── saved_model.pb          # Graph definition (protobuf)
├── keras_metadata.pb       # Keras metadata
└── fingerprint.pb          # Fingerprint

Load SavedModel

# Cách 1: Load as Keras model (full functionality)
new_model = tf.keras.models.load_model('saved_model/my_model')
print("Loaded as Keras model:")
new_model.summary()

# Cách 2: Load as TensorFlow SavedModel (low-level, cho serving)
imported = tf.saved_model.load('saved_model/my_model')
print("\nLoaded as TF SavedModel:")
print(f"Signatures: {list(imported.signatures.keys())}")

# Gọi signature serving_default
infer = imported.signatures['serving_default']
input_tensor = tf.ones((1, 10))
output = infer(input_tensor)
print(f"Output: {output}")

Đầu ra

Loaded as Keras model:
Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 dense (Dense)               (None, 64)                704
 dense_1 (Dense)             (None, 1)                 65
=================================================================
Total params: 769

Loaded as TF SavedModel:
Signatures: ['serving_default', '__saved_model_init_op']
Output: {'dense_1': }

So sánh HDF5 vs SavedModel

Khi nào dùng SavedModel?

• Deploy với TensorFlow Serving / Docker.
• Convert sang TFLite (mobile), TF.js (web).
• Model có custom layers/objects phức tạp.
• Production pipeline.