Tensorflow：logits和标签的第一维必须相同

Question

我是tensoflow的新手，我想用我自己的数据（40x40的图像）改编MNIST教程https://www.tensorflow.org/tutorials/layers。这是我的模型函数：

def cnn_model_fn(features, labels, mode):
        # Input Layer
        input_layer = tf.reshape(features, [-1, 40, 40, 1])

        # Convolutional Layer #1
        conv1 = tf.layers.conv2d(
                inputs=input_layer,
                filters=32,
                kernel_size=[5, 5],
                #  To specify that the output tensor should have the same width and height values as the input tensor
                # value can be "same" ou "valid"
                padding="same",
                activation=tf.nn.relu)

        # Pooling Layer #1
        pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)

        # Convolutional Layer #2 and Pooling Layer #2
        conv2 = tf.layers.conv2d(
                inputs=pool1,
                filters=64,
                kernel_size=[5, 5],
                padding="same",
                activation=tf.nn.relu)
        pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)

        # Dense Layer
        pool2_flat = tf.reshape(pool2, [-1, 10 * 10 * 64])
        dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)
        dropout = tf.layers.dropout(
                inputs=dense, rate=0.4, training=mode == tf.estimator.ModeKeys.TRAIN)

        # Logits Layer
        logits = tf.layers.dense(inputs=dropout, units=2)

        predictions = {
            # Generate predictions (for PREDICT and EVAL mode)
            "classes":       tf.argmax(input=logits, axis=1),
            # Add `softmax_tensor` to the graph. It is used for PREDICT and by the
            # `logging_hook`.
            "probabilities": tf.nn.softmax(logits, name="softmax_tensor")
        }

        if mode == tf.estimator.ModeKeys.PREDICT:
            return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)

        # Calculate Loss (for both TRAIN and EVAL modes)
        loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)

        # Configure the Training Op (for TRAIN mode)
        if mode == tf.estimator.ModeKeys.TRAIN:
            optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
            train_op = optimizer.minimize(
                    loss=loss,
                    global_step=tf.train.get_global_step())
            return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op)

        # Add evaluation metrics (for EVAL mode)
        eval_metric_ops = {
            "accuracy": tf.metrics.accuracy(
                    labels=labels, predictions=predictions["classes"])}
        return tf.estimator.EstimatorSpec(
                mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)

我在标签和logit之间有形状尺寸错误：

InvalidArgumentError（回溯，请参见上文）：logits和标签的第一维必须相同，logits形状为[3,2]，标签形状为[1]

filenames_array是16个字符串的数组

["file1.png", "file2.png", "file3.png", ...]

labels_array是16个整数的数组

[0,0,1,1,0,1,0,0,0,...]

主要功能是：

# Create the Estimator
mnist_classifier = tf.estimator.Estimator(model_fn=cnn_model_fn, model_dir="/tmp/test_convnet_model")

# Train the model
cust_train_input_fn = lambda: train_input_fn_custom(
        filenames_array=filenames, labels_array=labels, batch_size=1)

mnist_classifier.train(
        input_fn=cust_train_input_fn,
        steps=20000,
        hooks=[logging_hook])

我试图重塑logit，但没有成功：

logits = tf.reshape（logits，[1，2]）

我需要你的帮助，谢谢

编辑

经过更多时间搜索后，在模型函数的第一行

input_layer = tf.reshape(features, [-1, 40, 40, 1])

表示将动态计算batch_size尺寸的“ -1”在此处具有值“ 3”。与我的错误相同的“ 3”：
logits和标签的第一维必须相同，logits形状为[3,2]，标签形状为[1]

如果我将值强制设置为“ 1”，则会出现此新错误：

要重塑的输入是具有4800值的张量，但请求的形状具有1600

也许我的功能有问题？

编辑2：

完整的代码在这里：https
:
//gist.github.com/geoffreyp/cc8e97aab1bff4d39e10001118c6322e

编辑3

我更新了要点

logits = tf.layers.dense(inputs=dropout, units=1)

https://gist.github.com/geoffreyp/cc8e97aab1bff4d39e10001118c6322e

但是我不完全理解您对批次大小的回答，这里的批次大小如何为3，而我选择的批次大小为1？

如果我选择batch_size = 3， 则会 出现此错误：
logits和标签的第一维必须相同，logits的形状为[9,1]，标签的形状为[3]

我试图重塑标签：

labels = tf.reshape(labels, [3, 1])

我更新了功能和标签结构：

    filenames_train = [['blackcorner-data/1.png', 'blackcorner-data/2.png', 'blackcorner-data/3.png',
                   'blackcorner-data/4.png', 'blackcorner-data/n1.png'],
                   ['blackcorner-data/n2.png',
                   'blackcorner-data/n3.png', 'blackcorner-data/n4.png',
                   'blackcorner-data/11.png', 'blackcorner-data/21.png'],
                   ['blackcorner-data/31.png',
                   'blackcorner-data/41.png', 'blackcorner-data/n11.png', 'blackcorner-data/n21.png',
                   'blackcorner-data/n31.png']
                   ]

labels = [[0, 0, 0, 0, 1], [1, 1, 1, 0, 0], [0, 0, 1, 1, 1]]

但没有成功…

Answer 1

问题出在您的目标形状上，并且与正确选择适当的损失函数有关。您有两种可能性：

1.可能性 ：如果您具有一维整数编码目标，则可以sparse_categorical_crossentropy用作损失函数

n_class = 3
n_features = 100
n_sample = 1000

X = np.random.randint(0,10, (n_sample,n_features))
y = np.random.randint(0,n_class, n_sample)

inp = Input((n_features,))
x = Dense(128, activation='relu')(inp)
out = Dense(n_class, activation='softmax')(x)

model = Model(inp, out)
model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
history = model.fit(X, y, epochs=3)

2.可能性
：如果您已对目标进行一次热编码以具有2D形状（n_samples，n_class），则可以使用categorical_crossentropy

n_class = 3
n_features = 100
n_sample = 1000

X = np.random.randint(0,10, (n_sample,n_features))
y = pd.get_dummies(np.random.randint(0,n_class, n_sample)).values

inp = Input((n_features,))
x = Dense(128, activation='relu')(inp)
out = Dense(n_class, activation='softmax')(x)

model = Model(inp, out)
model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
history = model.fit(X, y, epochs=3)