总的来说，我对 CNN 和神经网络架构还很陌生，我正在尝试从 2D 图像创建回归模型。我对数据进行了重新整形，以便 xpix 表示图像的高度，ypix 表示宽度。
xpix = 322
ypix = 76
x_train = x_train.reshape(-1, 332, 76, 1)
x_val = x_val.reshape(-1, 332, 76, 1)
x_test = x_test.reshape(-1, 332, 76, 1)
print ('Training images shape: ', x_train.shape)
print ('Validation images shape: ', x_val.shape)
print ('Test Images shape:', x_test.shape)
Training images shape: (1387, 332, 76, 1)
Validation images shape: (463, 332, 76, 1)
Test Images shape: (617, 332, 76, 1)
我试过的模型架构是这样的
Regressor = Sequential()
Regressor.add(Conv2D(64, kernel_size=(3, 3), activation = 'relu', padding='same', input_shape=(xpix,ypix, 1)))
Regressor.add(MaxPooling2D(pool_size=(2, 2),padding='same'))
Regressor.add(BatchNormalization())
Regressor.add(Flatten())
Regressor.add(Dense(64, activation='relu'))
Regressor.add(Dropout(0.1))
Regressor.add(Dense(1, activation='linear'))
Adam= optimizers.Adam(lr=0.0001)
Regressor.compile(loss = 'mean_squared_error', optimizer=Adam)
History = Regressor.fit(x_train, y_train, batch_size=16, epochs=35, validation_data = (x_val, y_val))
但我收到以下错误消息
Epoch 1/35
WARNING:tensorflow:Model was constructed with shape (None, 322, 76, 1) for input Tensor("conv2d_3_input:0", shape=(None, 322, 76, 1), dtype=float32), but it was called on an input with incompatible shape (None, 332, 76, 1)
摘要如下
Regressor.summary()
Model: "sequential_5"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_4 (Conv2D) (None, 322, 76, 64) 640
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 161, 38, 64) 0
_________________________________________________________________
batch_normalization_3 (Batch (None, 161, 38, 64) 256
_________________________________________________________________
flatten_3 (Flatten) (None, 391552) 0
_________________________________________________________________
dense_5 (Dense) (None, 64) 25059392
_________________________________________________________________
dropout_2 (Dropout) (None, 64) 0
_________________________________________________________________
dense_6 (Dense) (None, 1) 65
=================================================================
Total params: 25,060,353
Trainable params: 25,060,225
Non-trainable params: 128
希望对此有任何评论或反馈！