Question

我是tensorflow的新手，现在我想通过自定义图层进行分类工作。我已经构建了一个模型，它将前几个DNN层输出连接到一个大张量，然后进行逻辑二进制分类。像这样（但没有RNN）： Model Structure

这是我的代码：

import tensorflow as tf
import numpy as np
import pandas as pd
from load_data import load_data_from_file

TRAIN_FILE = 'train.csv'
TEST_FILE = 'test.csv'
X_train, y_train = load_data_from_file(TRAIN_FILE)
y_train = y_train[np.newaxis]
y_train = y_train.transpose()
n_x_dnn = X_train.shape[1]
n_hidden_dnn = 50
n_merge_dnn1 = 5 * n_hidden_dnn
n_dnn6 = 50
hm_epochs = 10
batch_size = 50
X = tf.placeholder(tf.float32, [None, n_x_dnn])
y = tf.placeholder(tf.float32, [None, 1])

weights = {
    'dnn1': tf.Variable(tf.random_normal([n_x_dnn, n_hidden_dnn])),
    'dnn2': tf.Variable(tf.random_normal([n_hidden_dnn, n_hidden_dnn])),
    'dnn3': tf.Variable(tf.random_normal([n_hidden_dnn, n_hidden_dnn])),
    'dnn4': tf.Variable(tf.random_normal([n_hidden_dnn, n_hidden_dnn])),
    'dnn5': tf.Variable(tf.random_normal([n_hidden_dnn, n_hidden_dnn])),
    'dnn_merge': tf.Variable(tf.random_normal([n_merge_dnn1, n_dnn6])),
    'dnn6': tf.Variable(tf.random_normal([n_dnn6, 1]))
}
biases = {
    'dnn1_b': tf.Variable(tf.random_normal([n_hidden_dnn])),
    'dnn2_b': tf.Variable(tf.random_normal([n_hidden_dnn])),
    'dnn3_b': tf.Variable(tf.random_normal([n_hidden_dnn])),
    'dnn4_b': tf.Variable(tf.random_normal([n_hidden_dnn])),
    'dnn5_b': tf.Variable(tf.random_normal([n_hidden_dnn])),
    'dnn_merge_b': tf.Variable(tf.random_normal([n_dnn6])),
    'dnn6_b': tf.Variable(tf.random_normal([1])),
}


def define_layers():
    # dnn layer 1
    dnn_layer1 = tf.add(tf.matmul(X, weights['dnn1']), biases['dnn1_b'])
    dnn_layer1 = tf.nn.relu(dnn_layer1)

    # dnn layer 2
    dnn_layer2 = tf.add(tf.matmul(dnn_layer1, weights['dnn2']), biases['dnn2_b'])
    dnn_layer2 = tf.nn.relu(dnn_layer2)

    # dnn layer 3
    dnn_layer3 = tf.add(tf.matmul(dnn_layer2, weights['dnn3']), biases['dnn3_b'])
    dnn_layer3 = tf.nn.relu(dnn_layer3)

    # dnn layer 4
    dnn_layer4 = tf.add(tf.matmul(dnn_layer3, weights['dnn4']), biases['dnn4_b'])
    dnn_layer4 = tf.nn.relu(dnn_layer4)

    # dnn layer 5
    dnn_layer5 = tf.add(tf.matmul(dnn_layer4, weights['dnn5']), biases['dnn5_b'])
    dnn_layer5 = tf.nn.relu(dnn_layer5)

    # merge layer
    merged = tf.concat([dnn_layer1, dnn_layer2, dnn_layer3, dnn_layer4, dnn_layer5], 1)
    dnn_merge = tf.add(tf.matmul(merged, weights['dnn_merge']), biases['dnn_merge_b'])
    dnn_merge = tf.nn.relu(dnn_merge)

    # dnn layer 6
    dnn_layer6 = tf.add(tf.matmul(dnn_merge, weights['dnn6']), biases['dnn6_b'])
    dnn_layer6 = tf.nn.sigmoid(dnn_layer6)
    return dnn_layer6


logits = define_layers()
cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=y))
optimizer = tf.train.AdamOptimizer().minimize(cost)

saver = tf.train.Saver()
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    for epoch in range(hm_epochs):
        # epoch_loss = 0
        i = 0
        while i < len(X_train):
            start = i
            end = i + batch_size
            batch_x = np.array(X_train[start:end])
            batch_y = np.array(y_train[start:end])

            _, c, l = sess.run([optimizer, cost, logits], feed_dict={X: batch_x,
                                                                          y: batch_y})
            # epoch_loss += c
            i += batch_size

        # print(c, l)
        print('Epoch', epoch, 'completed out of', hm_epochs, 'loss:', c, 'logits:', l)

    save_path = saver.save(sess, "model/NF_dnn_and_rnn_model.ckpt")
    print("Model saved in file: %s" % save_path)

    X_test, y_test = load_data_from_file(TEST_FILE)
    predict = sess.run(logits, feed_dict={X: X_test})
    print(predict)

从输出结果可以看出我的费用没有变化：

Epoch 7 completed out of 10 loss: 1.27325 logits: [[ 1.]
Epoch 8 completed out of 10 loss: 1.27325 logits: [[ 1.]
Epoch 9 completed out of 10 loss: 1.27325 logits: [[ 1.]

Answer 1

您应该从最后一层dnn_layer6中移除sigmoid激活，因为sigmoid_cross_entropy_with_logits也会在内部应用sigmoid激活。

Tensorflow为什么我的物流成本不会改变？

1 个答案: