Question

我有一个用于预测CIFAR10类的CNN，我已经从MNIST数据集的工作解决方案中进行了改编。

该模型是 image-＆gt; conv-＆gt; pool-＆gt; conv-＆gt; pool-＆gt; FC-＆gt; dropout-＆gt; FC-＆gt; class_prediction ，我假设比完全连接的网络（我在做CS231 course时达到约50％）要好得多。

我收到了一些奇怪的结果，其中准确性和损失似乎并非真正相互关联：training accuracy and loss graph

如图所示，损失不断下降，这是一件好事，但即使在训练集上，它似乎也不会真正影响准确性。

很抱歉在这里全部转储，但在下面我已经提取了我认为相关的所有代码。如果我可以添加任何进一步的信息，请告诉我。

with tf.name_scope('Model'):
    # Construct model
    pred = conv_net(x, weights, biases, keep_prob)

with tf.name_scope('Loss'):
    # Define loss function and optimizer
    net = tf.nn.softmax_cross_entropy_with_logits(pred, y_true, name='Softmax_Loss')
    cost = tf.reduce_mean(net)

with tf.name_scope('Adam_Descent'):
    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, name='Adam_Descent').minimize(cost)   

# Evaluate the model
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y_true, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

#
# Code omitted
#

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

    step = 1
    # Train for a while
    while step * batch_size < training_iters:

        batch_x, batch_y = get_rand_train_batch(batch_size, 10000)

        # Run optimization        
        _, summary = sess.run([optimizer, merged_summary_op], feed_dict={x: batch_x, 
                                                                         y_true: batch_y, 
                                                                         keep_prob: 0.75})
        if step * index % display_step == 0:
            # Calculate loss and accuracy
            loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x, 
                                                              y_true: batch_y, 
                                                              keep_prob: 1})
            print_blue("Iter " + str(step*batch_size) + ", Minibatch Loss= " + \
                  "{:.6f}".format(loss) + ", Training Accuracy= " + \
                  "{:2.3f}%".format(acc*100))
        summary_writer.add_summary(summary, step * batch_size)
        step += 1

    # Calculate test accuracy
    print("\nTesting Accuracy:", \
    sess.run(accuracy, feed_dict={x: X_test[:1000],
                                  y_true: y_test[:1000],
                                  keep_prob: 1.}))

为什么我的损失和准确性似乎无关？

0 个答案: