Keras 数字数据集错误：预期 conv2d_input 有 4 个维度，但得到的数组形状为 (60000, 28, 28)

Question

Error in Keras Digit Dataset: ValueError: Error when checking input: expected conv2d_input to have 4 dimensions, but got array with shape (60000, 28, 28). I'm not sure what's going on, why do

I need 4 dimensions? I am positive they are 28x28 images. What shall I do?

   (train_images, train_labels), (test_images, test_labels) = mnist.load_data()
train_images = train_images/255
test_images = test_images/255

classes = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']

model = keras.Sequential([
    tf.keras.layers.Conv2D(56, (3, 3), activation='relu', input_shape=(28, 28)),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Conv2D(56, (3, 3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(784, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.fit(train_images, train_labels, epochs=5)

test_loss, test_acc = model.evaluate(test_images, test_labels)

Answer 1

Add the fourth dimension which is channel dimension to your input data using np.expand_dims , before you pass it into the network. Using this way, you will be able to use Conv2D .

train_images = np.expand_dims(train_images, axis=-1)
test_images = np.expand_dims(test_images, axis=-1)

print(train_images.shape()) # (60000, 28, 28, 1)
print(test_images.shape()) # (10000, 28, 28, 1)

Answer 2

Thanks, @bit01 for pointing my mistake I have updated my answer accordingly -

you cannot use 2D methods on a no depth or channel images. Your image doesn't contain any channel, having 28 X 28 shape, Conv2D is defined for images with channels like grey_scale(256,256,1) or RGB(256,256,3) or png images with alpha channel (256,256,4). It won't work for the 2D matrix without any depth.

so you have two ways to solve this problem:

1. switch from 2D method to 1D methods for Convolution and max-pooling layers.

Like this -

(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
train_images = train_images/255
test_images = test_images/255

classes = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']

model = keras.Sequential([
    tf.keras.layers.Conv1D(56, 3, activation='relu', input_shape=(28, 28)),
    tf.keras.layers.MaxPooling1D(2, 2),
    tf.keras.layers.Conv1D(56, 3, activation='relu'),
    tf.keras.layers.MaxPooling1D(2, 2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(784, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.fit(train_images, train_labels, epochs=5)

test_loss, test_acc = model.evaluate(test_images, test_labels)

2. add one extra dimention to your dataset (make it 28 X 28 -> 28 X 28 X1).

Like this -

(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

train_images = np.expand_dims(train_images, axis=-1)  # (60000, 28, 28, 1)
test_images = np.expand_dims(test_images, axis=-1) # (10000, 28, 28, 1)

train_images = train_images/255
test_images = test_images/255

#train_images.shape, test_images.shape
classes = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']


input_layer = keras.layers.Input(shape=(28, 28, 1), name='image_input')
conv1 = keras.layers.Conv2D(56, (3,3), activation='relu')(input_layer)
pool1 = keras.layers.MaxPooling2D(2, 2)(conv1)
conv2 = keras.layers.Conv2D(56, (3,3), activation='relu')(pool1)
pool2 = keras.layers.MaxPooling2D(2, 2)(conv2)
flatten = keras.layers.Flatten()(pool2)
dense1 = keras.layers.Dense(784, activation='relu')(flatten)
output_layer = keras.layers.Dense(10, activation='softmax')(dense1)

model = keras.models.Model(inputs=input_layer, outputs=output_layer, name='my_model')
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.fit(train_images, train_labels, epochs=5)

test_loss, test_acc = model.evaluate(test_images, test_labels)

Answer 3

You need to make two tiny changes. Firstly, you need to convert your image to a 4D tensor with a format of NHWC (batch size, height, width, channel), which is expected by Conv2D. However, you don't have channel dimension in current image dataset. The channel, 4th, dimension can be added like follows

train_images = np.expand_dims(train_images, axis=-1)  # (60000, 28, 28, 1)
test_images = np.expand_dims(test_images, axis=-1) # (10000, 28, 28, 1)

Secondly, you need to change the input_shape to the model specifying the no. of channel of your input image. Since your image have single channel input image shape should be (28, 28, 1)

tf.keras.layers.Conv2D(56, (3, 3), activation='relu', input_shape=(28, 28, 1))