環境

Tensorflow 1.12.0

やりたいこと

MNISTを学習させたモデルを使って、予測結果を出力したい。

学習に使ったモデル

訓練データはKaggleのMNISTのデータを使っています。

python
1# 使う関数を定義
2def shuffle_batch(X, y, batch_size):
3    rnd_idx = np.random.permutation(len(X))
4    n_batches = len(X) // batch_size
5    for batch_idx in np.array_split(rnd_idx, n_batches):
6        X_batch, y_batch = X[batch_idx], y[batch_idx]
7        yield X_batch, y_batch
8        
9def leaky_relu(z, name=None):
10    return tf.maximum(0.01 * z, z, name=name)
11
12# モデル構築
13n_inputs = 28 * 28  # MNIST
14n_hidden1 = 300
15n_hidden2 = 100
16n_outputs = 10
17
18X = tf.placeholder(tf.float32, shape=(None, n_inputs), name="X")
19y = tf.placeholder(tf.int32, shape=(None), name="y")
20
21with tf.name_scope("dnn"):
22    hidden1 = tf.layers.dense(X, n_hidden1, activation=leaky_relu, name="hidden1")
23    hidden2 = tf.layers.dense(hidden1, n_hidden2, activation=leaky_relu, name="hidden2")
24    logits = tf.layers.dense(hidden2, n_outputs, name="outputs")
25with tf.name_scope("loss"):
26    xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=logits)
27    loss = tf.reduce_mean(xentropy, name="loss")
28    
29learning_rate = 0.01
30with tf.name_scope("train"):
31    optimizer = tf.train.GradientDescentOptimizer(learning_rate)
32    training_op = optimizer.minimize(loss)
33with tf.name_scope("eval"):
34    correct = tf.nn.in_top_k(logits, y, 1)
35    accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
36    
37init = tf.global_variables_initializer()
38saver = tf.train.Saver()
39
40# 訓練
41n_epochs = 40
42batch_size = 50
43
44with tf.Session() as sess:
45    init.run()
46    for epoch in range(n_epochs):
47        for X_batch, y_batch in shuffle_batch(X_train, y_train, batch_size):
48            sess.run(training_op, feed_dict={X: X_batch, y: y_batch})
49        if epoch % 5 == 0:
50            acc_batch = accuracy.eval(feed_dict={X: X_batch, y: y_batch})
51            acc_valid = accuracy.eval(feed_dict={X: X_test, y: y_test})
52            print(epoch, "Batch accuracy:", acc_batch, "Validation accuracy:", acc_valid)
53
54    save_path = saver.save(sess, "./my_model_final.ckpt")

訓練で得られたパラメータは、チェックポイントとして保存しました。

予測

python
1tf.reset_default_graph()
2
3X = tf.placeholder(tf.float32, shape=(None, n_inputs), name="X")
4y = tf.placeholder(tf.int32, shape=(None), name="y")
5
6with tf.name_scope("dnn"):
7    hidden1 = tf.layers.dense(X, n_hidden1, activation=leaky_relu, name="hidden1")
8    hidden2 = tf.layers.dense(hidden1, n_hidden2, activation=leaky_relu, name="hidden2")
9    logits = tf.layers.dense(hidden2, n_outputs, name="outputs")
10
11saver = tf.train.Saver()
12
13with tf.Session() as sess:
14    saver.restore(sess, './my_model_final.ckpt')
15    pred = logits.eval(feed_dict={X: X_train})
16
17result = np.argmax(pred, axis=1)