In TensorFlow, a SavedModel is basically a serialized format for storing a complete TensorFlow program. The tf.saved_model.save() function in TensorFlow can be used to export a SavedModel. A trained model and its related variables are saved to disc in the SavedModel format by this function. It includes weights and computation graphs. This model can be used for deploying models for inference, transfer learning, or other purposes.
An illustration of code that shows how to export a SavedModel in Tensorflow is shown below:
Here are the following steps to export a SavedModel in TensorFlow:
Now we can take an example of exporting a Softmax Regression model in TensorFlow:
Step 1:
Create a Softmax Regression model and train it using the MNIST dataset.
Python
# Import the necessary libraries import tensorflow as tf # Load MNIST dataset mnist = tf.keras.datasets.mnist (x_train, y_train),(x_test, y_test) = mnist.load_data() # Normalize the data x_train, x_test = x_train / 255.0, x_test / 255.0 # Define the Softmax Regression model model = tf.keras.models.Sequential([ tf.keras.layers.Flatten(input_shape=(28, 28)), tf.keras.layers.Dense(10, activation='softmax') ]) # Compile the model model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) # Train the model model.fit(x_train, y_train, epochs=5, validation_data=(x_test, y_test)) |
Output:
Epoch 1/5 1875/1875 [==============================] - 3s 1ms/step - loss: 0.4726 - accuracy: 0.8764 - val_loss: 0.3060 - val_accuracy: 0.9154 Epoch 2/5 1875/1875 [==============================] - 2s 1ms/step - loss: 0.3038 - accuracy: 0.9158 - val_loss: 0.2808 - val_accuracy: 0.9222 Epoch 3/5 1875/1875 [==============================] - 3s 1ms/step - loss: 0.2831 - accuracy: 0.9209 - val_loss: 0.2758 - val_accuracy: 0.9235 Epoch 4/5 1875/1875 [==============================] - 3s 1ms/step - loss: 0.2734 - accuracy: 0.9229 - val_loss: 0.2688 - val_accuracy: 0.9256 Epoch 5/5 1875/1875 [==============================] - 2s 1ms/step - loss: 0.2661 - accuracy: 0.9257 - val_loss: 0.2675 - val_accuracy: 0.9267
Step 2:
Evaluate the model accuracy:
Python3
loss, accuracy = model.evaluate(x_test, y_test) print('The accuracy of the model is :',accuracy) |
Output:
313/313 [==============================] - 0s 1ms/step - loss: 0.2675 - accuracy: 0.9267 The accuracy of the model is : 0.9266999959945679
Step 3:
Save the model using the tf.saved_model.save function.
Python
# Save the model tf.saved_model.save(model, 'softmax_regression_model') |
Step 4:
Verify that the model has been saved successfully by inspecting the contents of the SavedModel directory.
Python
# Inspect the contents of the SavedModel directory !ls softmax_regression_model |
Output :
assets fingerprint.pb saved_model.pb variables
The saved_model.pb file contains the computation graph and the variables folder contains the trained weights of the model.
Step 5:
Optionally, you can add version numbers to the SavedModel directory to keep track of different versions of the model.
Python
# Save the model with version number tf.saved_model.save(model, 'softmax_regression_model/1') |
Step 6:
Verify that the versioned model has been saved successfully.
Python
# Inspect the contents of the SavedModel directory with version number !ls softmax_regression_model/1 |
Output:
assets fingerprint.pb saved_model.pb variables
The contents of the versioned directory are the same as of the original directory.
Step 7:
Let’s load the model using the TensorFlow hub
Python3
# import the necessary libraries import tensorflow_hub as hub import numpy as np #Load the model layer = hub.KerasLayer("softmax_regression_model/") # create custom input Input = np.linspace(0,1,784).reshape(28,28) Input = tf.expand_dims(Input, axis=0) print(layer(Input)) layer.trainable = Trueprint(layer.trainable_weights) |
Output:
tf.Tensor(
[[2.9126030e-16 5.6060693e-19 5.5827466e-03 6.9878437e-03 2.6784150e-19
9.8738652e-01 1.5881575e-17 4.2873238e-05 3.4423101e-15 5.1269178e-10]], shape=(1, 10), dtype=float32)
[<tf.Variable 'dense/kernel:0' shape=(784, 10) dtype=float32, numpy=
array([[-0.02441423, 0.04307493, -0.00332485, ..., 0.03308415,
0.07554208, -0.00533313],
[ 0.01357713, -0.00179666, 0.08344419, ..., -0.07203246,
0.04122115, -0.08428719],
[-0.06156562, -0.08187176, -0.06699241, ..., -0.05588092,
-0.04787212, -0.05234763],
...,
[-0.02957132, 0.07599085, -0.07504247, ..., 0.03821079,
0.04094885, -0.03252703],
[ 0.01038309, -0.01061375, 0.03050219, ..., -0.02180967,
0.07942755, 0.01886416],
[-0.02210779, 0.06026974, -0.00152316, ..., -0.02114672,
-0.06010536, -0.05444371]], dtype=float32)>, <tf.Variable 'dense/bias:0' shape=(10,) dtype=float32, numpy=
array([-0.37059996, 0.48123115, 0.09208864, -0.29430756, 0.09193437,
0.9103967 , -0.08532815, 0.5119505 , -1.0492268 , -0.1949552 ],
dtype=float32)>]
Here are some of the different ways related to exporting a SavedModel in TensorFlow:
- Exporting a SavedModel with custom signatures: You can directly specify custom signatures for your SavedModel that can define the inputs and the outputs of the model. This can be useful for serving the model using various interfaces, such as the REST APIs.
- Exporting a SavedModel for TensorFlow Serving: TensorFlow Serving is basically a system for serving machine learning models in production environments. To export a model for the TensorFlow Serving, first, you need to include a serving_default signature that can specify the inputs and outputs of the model.
- Exporting a SavedModel for TensorFlow Lite: TensorFlow Lite is a framework used for deploying machine learning
