目标不匹配的 Tensorflow 形状 (cifar10)

Question

I am currently using tensorflow and cifar10 to develop a model.

the input dimensions are loaded from cifar10.

(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.cifar10.load_data()
train_images, test_images = train_images / 255.0, test_images / 255.0
nb_classes = len(numpy.unique(train_labels))
train_labels = tf.one_hot(train_labels, nb_classes)
test_labels = tf.one_hot(test_labels, nb_classes) 

The input shapes are (10000, 32, 32, 3) and test shapes (10000, 1, 10).

I am receiving an error with my final line of the model's layers.

__________________________________________________________________________________________________

Layer (type)                    Output Shape         Param #     Connected to

==================================================================================================

input_1 (InputLayer)            [(None, 32, 32, 3)]  0

__________________________________________________________________________________________________

conv2d (Conv2D)                 (None, 32, 32, 32)   2432        input_1[0][0]

__________________________________________________________________________________________________

conv2d_1 (Conv2D)               (None, 32, 32, 32)   25632       conv2d[0][0]

__________________________________________________________________________________________________

conv2d_2 (Conv2D)               (None, 32, 32, 32)   25632       conv2d_1[0][0]

__________________________________________________________________________________________________

add (Add)                       (None, 32, 32, 32)   0           conv2d_2[0][0]

                                                                 conv2d[0][0]

__________________________________________________________________________________________________

conv2d_4 (Conv2D)               (None, 32, 32, 64)   51264       add[0][0]

__________________________________________________________________________________________________

conv2d_5 (Conv2D)               (None, 32, 32, 64)   102464      conv2d_4[0][0]

__________________________________________________________________________________________________

conv2d_3 (Conv2D)               (None, 32, 32, 64)   51264       conv2d[0][0]

__________________________________________________________________________________________________

add_1 (Add)                     (None, 32, 32, 64)   0           conv2d_5[0][0]

                                                                 conv2d_3[0][0]

__________________________________________________________________________________________________

flatten (Flatten)               (None, 65536)        0           add_1[0][0]

__________________________________________________________________________________________________

dense (Dense)                   (None, 512)          33554944    flatten[0][0]

__________________________________________________________________________________________________

dense_1 (Dense)                 (None, 10)           5130        dense[0][0]

==================================================================================================

I'm getting the error message "ValueError: Shapes (None, 1, 10) and (None, 10) are incompatible.

I'm unsure how to solve this error.

Answer 1

This happens because the CIFAR-10 labels are nested in the shape (None, 1) instead of just (None) .

>>> a = [[0], [1], [2], [3]]
>>> tf.one_hot(a, 4)
<tf.Tensor: shape=(4, 1, 4), dtype=float32, numpy=
array([[[1., 0., 0., 0.]],
       [[0., 1., 0., 0.]],
       [[0., 0., 1., 0.]],
       [[0., 0., 0., 1.]]], dtype=float32)>

>>> b = [0, 1, 2, 3]
>>> tf.one_hot(a, 4)
<tf.Tensor: shape=(4, 4), dtype=float32, numpy=
array([[1., 0., 0., 0.],
       [0., 1., 0., 0.],
       [0., 0., 1., 0.],
       [0., 0., 0., 1.]], dtype=float32)>

You can see that nested labels produces nested one-hot encoded labels. In your case, you can just flatten or reshape the labels before one-hot encoding:

...
nb_classes = len(numpy.unique(train_labels))
train_labels = train_labels.flatten()
test_labels = test_labels.flatten()

train_labels = tf.one_hot(train_labels, nb_classes)
test_labels = tf.one_hot(test_labels, nb_classes)

Now the labels' shape should correspond to the model's output shape.

Answer 2

your train_labels original shape is (50000,1) you need to remove the 1 using numpy.squeeze) which will make the shape(50000) same deal for test_labels

(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.cifar10.load_data()
train_images, test_images = train_images / 255.0, test_images / 255.0
nb_classes = len(numpy.unique(train_labels))
test_labels=numpy.squeeze(test_labels)
train_labels=numpy.squeeze(train_labels)
train_labels = tf.one_hot(train_labels, nb_classes)
test_labels = tf.one_hot(test_labels, nb_classes)