After running a few deep -learning programs, I started to calm down to study the algorithm and code, because I just started to contact TensorFlow. The first way to block the use of various APIs, I don’t want to swallow jujube. , Each break.

Today is mainly the introduction of the TF.KERAS interface, and the preservation and loading of more models recently encountered.

1. Introduction to TF.KERAS

The TF.KERAS interface is the Keras interface encapsulated in TensorFlow. The Keras interface is a high -rise neural network API written in Python language that can be run on TensorFlow, CNTK, and Theano, that is, it can use these three frameworks as the background. Keras has the following advantages:

(1) User friendship, modularization, and expansion;

(2) Can support convolutional neural networks and circulating neural networks;

(3) Can be seamlessly supported by CPU and GPU.

The keyflow interface in TensorFlow has a comprehensive implementation of Keras. As long as TensorFlow is installed, you can use all the interfaces of Keras.

The TF.KERAS interface contains the source code of many mature models, including Densenet, NASNET, MOBILENET, etc. We can easily use these source code to train them. The parameter value is assigned to the weight of the model source code, and the model after assignment can be directly used to predict.

The key pages on GitHub have many models of the source code, structure, parameters, etc. students who need to make model evaluation can be viewed through the following links. For example, we need to understand the parameters of the model some time ago. D.

GitHub - keras-team/keras-applications: Reference implementations of popular deep learning models.

2. Model preservation and loading

Save the model can use the following API:

model.save() 
#or 
tf.keras.models.save_model()

Model loading:

tf.keras.models.load_model()

The model preservation format can be the SaveDmodel format of TensorFlow, or the Keras H5 format. Savedmodel format is the default format of Model.save (). If you want to save it in H5 format, there are two ways to achieve:

(1) When calling the save () function, pass the save_format = 'h5';

(2) Pass the file name ending with .h5 or .keras to the save () function.

2.1 Savedmodel model

Savedmodel example is as follows, the example comes fromOfficial tutorial。

def get_model():
    # Create a simple model.
    inputs = keras.Input(shape=(32,))
    outputs = keras.layers.Dense(1)(inputs)
    model = keras.Model(inputs, outputs)
    model.compile(optimizer="adam", loss="mean_squared_error")
    return model


model = get_model()

# Train the model.
test_input = np.random.random((128, 32))
test_target = np.random.random((128, 1))
model.fit(test_input, test_target)

# Calling `save('my_model')` creates a SavedModel folder `my_model`.
model.save("my_model")

# It can be used to reconstruct the model identically.
reconstructed_model = keras.models.load_model("my_model")

# Let's check:
np.testing.assert_allclose(
    model.predict(test_input), reconstructed_model.predict(test_input)
)

# The reconstructed model is already compiled and has retained the optimizer
# state, so training can resume:
reconstructed_model.fit(test_input, test_target)

transfer model.save('my_model')Will create a name calledmy_modelThe folder contains the following content:

Model architecture and training configuration (including optimizers, losses and indicators) stored insaved_model.pbIn the middle, the weight is preservedvariables/ Under contents.

2.2 H5 model

Keras also supports saving a single HDF5 file, which contains the architecture, weight value andcompile() information. It is a lightweight alternative option for SaveDModel.

The official tutorial examples are as follows:

model = get_model()

# Train the model.
test_input = np.random.random((128, 32))
test_target = np.random.random((128, 1))
model.fit(test_input, test_target)

# Calling `save('my_model.h5')` creates a h5 file `my_model.h5`.
model.save("my_h5_model.h5")

# It can be used to reconstruct the model identically.
reconstructed_model = keras.models.load_model("my_h5_model.h5")

# Let's check:
np.testing.assert_allclose(
    model.predict(test_input), reconstructed_model.predict(test_input)
)

# The reconstructed model is already compiled and has retained the optimizer
# state, so training can resume:
reconstructed_model.fit(test_input, test_target)

The saved file is a separate My_H5_model.h5 file.

Time relationship, write here today, and then update more content according to the learning situation.