tf.layers denseを使ったNNの重みWの取り出し方がわからない

tf.layers denseを使用してNNを実装しました。学習済みの重みWを取り出して回帰線を引きたいのですが、重みWの取り出し方がわかりません。コードは以下の通りです。

Python
1
2from collections import namedtuple
3def build_neural_network(hidden_units1=100, hidden_units2=50, hidden_units3=10):
4    tf.reset_default_graph()
5    inputs = tf.placeholder(tf.float32, shape=[None, 2])
6    labels = tf.placeholder(tf.float32, shape=[None, 1])
7    learning_rate = tf.placeholder(tf.float32)
8    is_training = tf.Variable(True, dtype=tf.bool)
9    
10    initializer = tf.contrib.layers.xavier_initializer()
11    fc = tf.layers.dense(inputs, hidden_units1, activation=None, kernel_initializer=initializer)
12
13    fc=tf.nn.relu(fc)
14    
15    fc = tf.layers.dense(fc, hidden_units2, activation=None, kernel_initializer=initializer)
16
17    fc = tf.nn.relu(fc)
18    
19    fc = tf.layers.dense(fc, hidden_units3, activation=None, kernel_initializer=initializer)
20    fc = tf.layers.batch_normalization(fc, training=is_training)
21    fc = tf.nn.relu(fc)
22    
23    logits = tf.layers.dense(fc, 1, activation=None)
24    cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels, logits=logits)
25    cost = tf.reduce_mean(cross_entropy)
26    
27    with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
28        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
29        
30        predicted = tf.nn.sigmoid(logits)
31        correct_pred = tf.equal(tf.round(predicted), labels)
32        accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
33        
34        export_nodes = ['inputs', 'labels', 'learning_rate', 'is_training', 'logits', 'cost', 'optimizer', 'predicted', 'accuracy']
35        Graph = namedtuple('Graph', export_nodes)
36        local_dict = locals()
37        graph = Graph(*[local_dict[each] for each in export_nodes])
38
39    return graph
40
41model = build_neural_network()
42
43
44def get_batch(data_x,data_y,batch_size=32):
45    batch_n=len(data_x)//batch_size
46    for i in range(batch_n):
47        batch_x=data_x[i*batch_size:(i+1)*batch_size]
48        batch_y=data_y[i*batch_size:(i+1)*batch_size]
49        
50        yield batch_x,batch_y
51
52
53epochs = 50
54train_collect = 50
55train_print=train_collect*2
56
57learning_rate_value = 0.001
58batch_size=16
59
60x_collect = []
61train_loss_collect = []
62train_acc_collect = []
63valid_loss_collect = []
64valid_acc_collect = []
65
66saver = tf.train.Saver()
67with tf.Session() as sess:
68    sess.run(tf.global_variables_initializer())
69    iteration=0
70    for e in range(epochs):
71        for batch_x,batch_y in get_batch(train_x,train_y,batch_size):
72            iteration+=1
73            feed = {model.inputs: train_x,
74                    model.labels: train_y,
75                    model.learning_rate: learning_rate_value,
76                    model.is_training:True
77                   }
78
79            train_loss, _, train_acc = sess.run([model.cost, model.optimizer, model.accuracy], feed_dict=feed)
80            
81            if iteration % train_collect == 0:
82                x_collect.append(e)
83                train_loss_collect.append(train_loss)
84                train_acc_collect.append(train_acc)
85
86                if iteration % train_print==0:
87                     print("Epoch: {}/{}".format(e + 1, epochs),
88                      "Train Loss: {:.4f}".format(train_loss),
89                      "Train Acc: {:.4f}".format(train_acc))
90                        
91                feed = {model.inputs: valid_x,
92                        model.labels: valid_y,
93                        model.is_training:False
94                       }
95                val_loss, val_acc = sess.run([model.cost, model.accuracy], feed_dict=feed)
96                valid_loss_collect.append(val_loss)
97                valid_acc_collect.append(val_acc)
98                
99                if iteration % train_print==0:
100                    print("Epoch: {}/{}".format(e + 1, epochs),
101                      "Validation Loss: {:.4f}".format(val_loss),
102                      "Validation Acc: {:.4f}".format(val_acc))
103        saver.save(sess, "./ex_nn.ckpt")