Tensorflow二进制分类器训练随机工作

Question

我已经修改了tensorflow MNIST示例，用我自己的照片创建二进制分类。 CNN分类器确实训练，但只有在我尝试运行该脚本几次之后。大多数时候，我只是在每一步都获得50％的准确度。

看起来权重初始化并不起作用。或许还有其他问题。

有什么想法吗？

with tf.Graph().as_default():
    with tf.Session() as sess:
        # Tensorflow placeholders
        input = tf.placeholder(
            tf.float32, shape=(None, 32, 32, 1))
        label = tf.placeholder(
            tf.float32, shape=(None, 2))


        def train(output, epochs=1000):
            loss = tf.reduce_mean(-tf.reduce_sum(label * tf.log(output), reduction_indices=[1]))
            correct_samples = tf.reduce_sum(tf.reduce_sum(label * output))

            train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)

            sess.run(tf.initialize_all_variables())

            for i in range(epochs):
                for j in range(int(math.floor(X_train.shape[0] / batch_size))):
                    batch_input = X_train[j*batch_size: (j*batch_size) + batch_size]
                    batch_label = y_train[j*batch_size: (j*batch_size) + batch_size]
                    sess.run(train_step, feed_dict={input: batch_input, label: batch_label})
                if i % 200 == 0:
                    print i, sess.run(correct_samples, feed_dict={input: X_train, label: y_train}) / X_train.shape[0]

        def convolution(flow, kernel_size, in_maps, out_maps):
            filter_ = tf.Variable(tf.random_normal([1,1,1,1], stddev=0.1))
            return tf.nn.conv2d(flow, filter_, strides=[1, 1, 1, 1], padding='VALID')

        def pooling(flow, kernel_size):
             return tf.nn.max_pool(flow, ksize=[1, kernel_size, kernel_size, 1], strides=[1, kernel_size, kernel_size, 1], padding='SAME')

        conv = convolution(input, 5, 32, 32)
        conv = tf.nn.relu(conv)
        conv = pooling(conv, 2)

        conv = convolution(conv, 5, 32, 32)
        conv = tf.nn.relu(conv)
        conv = pooling(conv, 2)

        conv = tf.reshape(conv, [-1, 8*8])

        W1 = tf.Variable(tf.truncated_normal(shape=[8*8, 200], stddev=0.1))
        b1 = tf.Variable(tf.constant(0.1, shape=[200]))

        W2 = tf.Variable(tf.truncated_normal(shape=[200, 2], stddev=0.1))
        b2 = tf.Variable(tf.constant(0.1, shape=[2]))

        layer1 = tf.matmul(conv, W1) + b1
        activation1 = tf.nn.relu(layer1)
        logits = tf.matmul(activation1, W2) + b2
        output = tf.nn.softmax(logits)

        import datetime
        print datetime.datetime.now()
        train(output, epochs=5000)
        print datetime.datetime.now()