损失函数正在返回 nan tensorflow

Question

我在这里编写了一个简单的 tensorflow 程序，它读取特征列表并尝试预测类别。

with tf.Session() as sess:
        sess.run(tf.initialize_all_variables())

        for epoch in range (hm_epochs):
            epoch_loss = 0
            itere = int(X_train.shape[0]/batch_size)
            last = 0
            add = 1
            for start in range(itere):
                x_train_epoch = X_train[last: ((start + add) * batch_size),:]
                y_train_epoch = y_1Hot_train.eval()[last: ((start + add) * batch_size),:]
#                 print("shape of x", x_train_epoch.shape, "shape of y", y_train_epoch.shape)
                _, c = sess.run([optimizer, cost], feed_dict = {x: x_train_epoch, y: y_train_epoch})
                epoch_loss += c
                last = start * batch_size
                add = 0
            print('Epoch', epoch, 'completed out of', hm_epochs, 'loss', epoch_loss )
        correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
        accuracy = tf.reduce_mean(tf.cast(correct, 'float'))
        print('Accuracy:', accuracy.eval( {x: X_test, y: y_1Hot_test.eval() }))

链接：https://gist.github.com/makark/79af6ca53ca27d51abb1d87c9b9bac07

数据文件：https://gist.github.com/makark/eb859f50237edb9343f3ca32aeb3be2b

但是，当我运行我的代码时，我一直在丢失“nan”。我不确定发生了什么...任何帮助将不胜感激！

WARNING:tensorflow:From <ipython-input-149-0164f4af7d52>:46: initialize_all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02.
Instructions for updating:
Use `tf.global_variables_initializer` instead.
Epoch 0 completed out of 10 loss nan
Epoch 1 completed out of 10 loss nan
Epoch 2 completed out of 10 loss nan
Epoch 3 completed out of 10 loss nan
Epoch 4 completed out of 10 loss nan
Epoch 5 completed out of 10 loss nan
Epoch 6 completed out of 10 loss nan
Epoch 7 completed out of 10 loss nan
Epoch 8 completed out of 10 loss nan
Epoch 9 completed out of 10 loss nan
Accuracy: 0.589097

Answer 1

• 输入有 nan，请通过 X[np.isnan(X)] = 0 修复它。

• 输入未缩放，请使用 sklearn 的 StandardScaler 标准化您的输入。

• 在 random_normal 中使用 stddev 将权重设置为较小的初始值。

• 修复输出计算错误：output = tf.add(tf.matmul(l3, output_layer['weights']),output_layer['biases'] )。