ValueError：层顺序需要 1 个输入，但它收到 239 个输入张量

Question

I am trying to test a simple convolutional layer where the input images (with 1 band= grayscale) are numpy arrays stored in a list and targets are stored in a pandas dataframe. The size of input images is 16x16. The output for the model.fit is an error of "Layer sequential expects 1 input(s), but it received 239 input tensors". I also checked this link but still I couldn't find the answer. Can anyone help me to resolve this error?

(trainY, testY, trainX, testX) = train_test_split(df, images, test_size=0.20, random_state=42)
print (np.shape(trainY),np.shape(testY),np.shape(trainX),np.shape(testX))

result: (239, 1) (60, 1) (239, 16, 16) (60, 16, 16)

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(16, 16, 1),name='Layer1'))
model.add(layers.Flatten())
model.add(layers.Dense(16, activation='relu',name='layer2'))
model.add(layers.Dense(1,activation='linear',name='Layer3'))
model.summary()

model.compile(optimizer='adam',
              loss='mean_squared_error',
              metrics=['mae'])

history = model.fit(trainX, trainY, epochs=10, 
                    validation_split=.2, batch_size=4)

Answer 1

The input expected by the model as defined in you model architecture is

input_shape=(16, 16, 1)

So while training you can send a batch of 16X16 single channel ( X1 ) images. However your data shape is (239, 16, 16) . ie you have a batch of 239 images of 16X16 . All you have to do is reshape 16X16 to 16X16X1 . Since your data is in numpy array you can do this using expand_dim .

trainX = np.expand_dims(trainX, -1)

Fixed code

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', 
                        input_shape=(16, 16, 1),name='Layer1'))
model.add(layers.Flatten())
model.add(layers.Dense(16, activation='relu',name='layer2'))
model.add(layers.Dense(1,activation='linear',name='Layer3'))

model.compile(optimizer='adam',
              loss='mean_squared_error',
              metrics=['mae'])

trainY = np.random.randint(0,10, (239, 1))
trainX = np.random.randn(239, 16, 16)

trainX = np.expand_dims(trainX, -1)
history = model.fit(trainX, trainY, epochs=10, validation_split=.2, batch_size=4)