如何保存 model，使用 Tensorflow 上保存的 model 进行加载和预测？

Question

I'm new to python and neural networks with Tensorflow, so I have no idea how to restore a saved model and make predictions. I have collected data from three cameras (X_test) of a line follower robot and It's wheels speed (left and right) set by simulator for each camera view (y_test). I need to train a neural network to predict the wheels speed based on cameras views. That's the code:

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

x_train = np.zeros((2392,3072))
y_train = np.zeros((2392,2))
x_test = np.zeros((2392,3072))
y_test = np.zeros((2392,2))

X_1_train = np.loadtxt('X_Treinamento_3.txt')
X_2_train = np.loadtxt('X_Treinamento_4.txt')
y_1_train = np.loadtxt('Vl_Vr_Treinamento_3.txt')
y_2_train = np.loadtxt('Vl_Vr_Treinamento_4.txt')

X_1_test = np.loadtxt('X_Teste_3.txt')
X_2_test = np.loadtxt('X_Teste_4.txt')
y_1_test = np.loadtxt('Vl_Vr_Teste_3.txt')
y_2_test = np.loadtxt('Vl_Vr_Teste_4.txt')

x_train = np.concatenate((X_1_train[4:1200,:],X_2_train[4:1200,:]))/255
x_test = np.concatenate((X_1_test[4:1200,:],X_2_test[4:1200,:]))/255
y_train = np.concatenate((y_1_train[4:1200,:],y_2_train[4:1200,:]))
y_test = np.concatenate((y_1_test[4:1200,:],y_2_test[4:1200,:]))

n_inputs = 32*32*3 #3 images 32x32 pixels
n_hidden_1 = 300 # 1st layer number of neurons
n_hidden_2 = 100 # 2nd layer number of neurons
n_outputs = 2

tf.compat.v1.disable_eager_execution()
X = tf.compat.v1.placeholder("float",[None,n_inputs])
Y = tf.compat.v1.placeholder("float",[None,n_outputs])

weights = {
    'h1': tf.Variable(tf.random.normal([n_inputs, n_hidden_1],stddev=0.1)),#4 inputs 10  nodes in h1 layer
    'h2': tf.Variable(tf.random.normal([n_hidden_1, n_hidden_2],stddev=0.1)),# 10 nodes in h2 layer
    'out': tf.Variable(tf.random.normal([n_hidden_2, n_outputs],stddev=0.1))# 1 ouput label
}
biases = {
    'b1': tf.Variable(tf.random.normal([n_hidden_1])),
    'b2': tf.Variable(tf.random.normal([n_hidden_2])),
    'out': tf.Variable(tf.random.normal([n_outputs]))
}

def neural_net(x):
    #hidden layer 1
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])
    layer_1 = tf.nn.tanh(layer_1)#activation
    #hideen layer 2
    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])
    layer_2 = tf.nn.tanh(layer_2)#activation
    # output layer
    out_layer = tf.matmul(layer_2, weights['out']) + biases['out']
    return (out_layer)

Y_hat=neural_net(X)
loss_op=tf.losses.mean_squared_error(Y,Y_hat)#loss function
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=1e-3)  # define optimizer # play around with learning rate
train_op = optimizer.minimize(loss_op)  # minimize losss

init = tf.compat.v1.global_variables_initializer()
epoch=2000

with tf.compat.v1.Session() as sess:
    sess.run(init)
    saver = tf.compat.v1.train.Saver()
    for i in range(0,epoch):
        sess.run(train_op,feed_dict={X:x_train,Y:y_train})
        loss=sess.run(loss_op,feed_dict={X:x_train,Y:y_train})
        if(i%100==0):
            print("epoch no "+str(i),(loss))
        pred=sess.run(Y_hat,feed_dict={X:x_test})
        saver.save(sess, "output")

So, I need your help to check if I'm saving the model correctly, if not, how to save it and after how to load and predict with the saved model.

Answer 1

You are saving it correctly, but if you want save more checkpoints ( for example: maybe last checkpoint is overfitting and previous checkpoint will give better results with test data), than you need add global_step ( your checkpoints will be saved something like output-0, output-2000 and so on), for more info look here docs

saver.save(sess, "output", global_step = i)

Loading and making prediction in your case will look something like this:

with tf.compat.v1.Session() as sess:
    sess.run(init)
    saver = tf.compat.v1.train.Saver()
    saver.restore(sess, "output")
    pred=sess.run(Y_hat,feed_dict={X:x_test})

"output" is path to your saved checkpoint

Answer 2

Thank you a lot, it works !