Keras CNN 的输入维度错误：预期为 4-dim，但发现为 3-dim

Question

我有一个关于使用 Keras 编码 CNN 的问题。

输入数据(adj)的形状是(20000, 50, 50)； 20000 是样本数，50 x 50 是二维数据（如图像）。批量大小为 100。 （其实有两个输入：adj=(20000, 50, 50), features=(20000, 50, 52)。

下发部分如下：

from keras.layers import Conv2D, MaxPool2D, Flatten

adj_visible1 = Input(shape=(50, 50, 1))
conv11 = Conv2D(16, kernel_size=5, activation='relu')(adj_visible1)
pool11 = MaxPool2D(pool_size=(2, 2))(conv11)
conv12 = Conv2D(8, kernel_size=5, activation='relu')(pool11)
pool12 = MaxPool2D(pool_size=(2, 2))(conv12)
flat1 = Flatten()(pool12)

但出现如下错误消息：

ValueError: Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3

我发现打印相同消息的类似案例，但是，大多数原因是他们没有考虑像 (50, 50) 这样的过滤器，而不是 (50, 50, "1") 输入形状。

在我的例子中，我使用了形状 (50, 50, 1) 而不是 (50, 50)。但是，它仍然会打印相同的错误消息。

我该怎么办？

我附上完整代码如下：

from sklearn.cross_validation import train_test_split
from keras.models import Sequential
from keras.layers.core import Dense, Dropout
from keras.optimizers import RMSprop, Adam, Adadelta
from keras.utils import plot_model
from keras.models import Model
from keras.layers import Input, Flatten, MaxPool2D
from keras.layers.convolutional import Conv2D
from keras.layers.merge import concatenate
from keras.callbacks import CSVLogger

#Settings
epoch = 100
batch_size = 100
test_size = 10000

# Load data
adj = np.load('adj.npy')   #(20000, 50, 50)
features = np.load('features.npy')    #(20000, 50, 52)
Prop = np.load('Properties.npy')     #(20000, 1)

database = np.dstack((adj, features))    #(20000, 50, 102)


#Train/Test split
X_tr, X_te, Y_tr, Y_te = train_test_split(database, Prop, test_size=test_size)

X_tr_adj, X_tr_features = X_tr[:, :, 0:50], X_tr[:, :, 50:]
X_te_adj, X_te_features = X_te[:, :, 0:50], X_te[:, :, 50:]


def create_model():

    # first input model
    adj_visible1 = Input(shape=(50, 50, 1))
    conv11 = Conv2D(16, kernel_size=5, activation='relu')(adj_visible1)
    pool11 = MaxPool2D(pool_size=(2, 2))(conv11)
    conv12 = Conv2D(8, kernel_size=5, activation='relu')(pool11)
    pool12 = MaxPool2D(pool_size=(2, 2))(conv12)
    flat1 = Flatten()(pool12)

    # second input model
    features_visible2 = Input(shape=(50, 52, 1))
    conv21 = Conv2D(16, kernel_size=5, activation='relu')(features_visible2)
    pool21 = MaxPool2D(pool_size=(2, 2))(conv21)
    conv22 = Conv2D(8, kernel_size=5, activation='relu')(pool21)
    pool22 = MaxPool2D(pool_size=(2, 2))(conv22)
    flat2 = Flatten()(pool22)

    # merge input models
    merge = concatenate([flat1, flat2])

    # interpretation model
    hidden1 = Dense(128, activation='relu')(merge)
    hidden2 = Dense(32, activation='relu')(hidden1)
    output = Dense(1, activation='linear')(hidden2)
    model = Model(inputs=[adj_visible1, features_visible2], outputs=output)
    model.compile(loss='mean_squared_error', optimizer=Adam())
    # summarize layers
    print(model.summary())

    return model

def train_model(batch_size = 100, nb_epoch = 20):
    model = create_model()
    csv_logger = CSVLogger('CNN trial.csv')
    history = model.fit([X_tr_adj, X_tr_features], Y_tr,
                        batch_size=batch_size,
                        epochs=nb_epoch,
                        verbose=1,
                        validation_data=([X_te_adj, X_te_features], Y_te),
                        callbacks=[csv_logger])

    predictions_valid = model.predict(X_te_adj, X_te_features, batch_size=batch_size, verbose=1)

    return model

train_model(nb_epoch = epoch)

我参考以下材料编写代码：https://machinelearningmastery.com/keras-functional-api-deep-learning/

Answer 1

您必须使用 conv1D 和 MaxPool1D 而不是 conv2D 和 MaxPool2D 因为您的数据集是单一的通道图像而不是 3 通道图像。在 conv1D 层中，期望输入的格式为 Batch X Height X Width，而在 conv2D 中，它期望输入为维度 = 4，即 Batch X Height X Width X Channels。

from sklearn.model_selection import train_test_split
from keras.models import Sequential
from keras.layers.core import Dense, Dropout
from keras.optimizers import RMSprop, Adam, Adadelta
from keras.utils import plot_model
from keras.models import Model
from keras.layers import Input, Flatten, MaxPool1D
from keras.layers.convolutional import Conv1D
from keras.layers.merge import concatenate
from keras.callbacks import CSVLogger
import numpy as np
epoch = 100
batch_size = 100
test_size = 10000


adj = np.random.randint(0,high=100, size=(20000, 50, 50))      #(20000, 50, 50)
features = np.random.randint(0,high=100, size=(20000, 50, 52))    #(20000, 50, 52)
Prop = np.random.randint(0,high=100, size=(20000,))     #(20000, 1)

database = np.dstack((adj, features))    #(20000, 50, 102)

print( " shape of database :", database.shape)


t
X_tr, X_te, Y_tr, Y_te = train_test_split(database, Prop, test_size=test_size)

X_tr_adj, X_tr_features = X_tr[:, :, 0:50], X_tr[:, :, 50:]
X_te_adj, X_te_features = X_te[:, :, 0:50], X_te[:, :, 50:]


def create_model():

    # first input model
    adj_visible1 = Input(shape=(50, 50))
    conv11 = Conv1D(16, kernel_size=5, activation='relu')(adj_visible1)
    pool11 = MaxPool1D(pool_size=2)(conv11)
    conv12 = Conv1D(8, kernel_size=5, activation='relu')(pool11)
    pool12 = MaxPool1D(pool_size=2)(conv12)
    flat1 = Flatten()(pool12)

    # second input model
    features_visible2 = Input(shape=(50, 52))
    conv21 = Conv1D(16, kernel_size=5, activation='relu')(features_visible2)
    pool21 = MaxPool1D(pool_size=2)(conv21)
    conv22 = Conv1D(8, kernel_size=5, activation='relu')(pool21)
    pool22 = MaxPool1D(pool_size=2)(conv22)
    flat2 = Flatten()(pool22)

    # merge input models
    merge = concatenate([flat1, flat2])

    # interpretation model
    hidden1 = Dense(128, activation='relu')(merge)
    hidden2 = Dense(32, activation='relu')(hidden1)
    output = Dense(1, activation='linear')(hidden2)
    model = Model(inputs=[adj_visible1, features_visible2], outputs=output)
    model.compile(loss='mean_squared_error', optimizer=Adam())
    # summarize layers
    print(model.summary())

    return model

def train_model(batch_size = 100, nb_epoch = 20):
    model = create_model()
    csv_logger = CSVLogger('CNN trial.csv')
    history = model.fit([X_tr_adj, X_tr_features], Y_tr,
                        batch_size=batch_size,
                        epochs=nb_epoch,
                        verbose=1,
                        validation_data=([X_te_adj, X_te_features], Y_te),
                        callbacks=[csv_logger])

    return model

train_model(nb_epoch = 10)