Keras-输入尺寸为[？，4,80,64]，[5,5,64,64]的'conv2d_5 / convolution'（op：'Conv2D'）的4中减去5引起的负尺寸大小

Question

I have a similar model to the one below, but after modifying the architecture, I keep getting the following error:

Negative dimension size caused by subtracting 5 from 4 for 'conv2d_5/convolution' (op: 'Conv2D') with input shapes: [?,4,80,64], [5,5,64,64].

I am still new to machine learning so I couldn't make much sense of the parameters. Any help?

model_img = Sequential(name="img")
    # Cropping
    model_img.add(Cropping2D(cropping=((124,126),(0,0)), input_shape=(376,1344,3)))
    # Normalization
    model_img.add(Lambda(lambda x: (2*x / 255.0) - 1.0))
    model_img.add(Conv2D(16, (7, 7), activation="relu", strides=(2, 2)))
    model_img.add(Conv2D(32, (7, 7), activation="relu", strides=(2, 2)))
    model_img.add(Conv2D(32, (5, 5), activation="relu", strides=(2, 2)))
    model_img.add(Conv2D(64, (5, 5), activation="relu", strides=(2, 2)))
    model_img.add(Conv2D(64, (5, 5), activation="relu", strides=(2, 2)))
    model_img.add(Conv2D(128, (3, 3), activation="relu"))
    model_img.add(Conv2D(128, (3, 3), activation="relu"))
    model_img.add(Flatten())
    model_img.add(Dense(100))
    model_img.add(Dense(50))
    model_img.add(Dense(10))

    model_lidar = Sequential(name="lidar")
    model_lidar.add(Dense(32, input_shape=(360,)))
    model_lidar.add(Dropout(0.1))
    model_lidar.add(Dense(10))

    model_imu = Sequential(name='imu')
    model_imu.add(Dense(32, input_shape=(10, )))
    model_imu.add(Dropout(0.1))
    model_imu.add(Dense(10))

    merged = Merge([model_img, model_lidar, model_imu], mode="concat")
    model = Sequential()
    model.add(merged)
    model.add(Dense(16))
    model.add(Dropout(0.2))
    model.add(Dense(1))

Answer: I couldn't complete the training because of issues with sensor but the model works fine now thanks to the 2 answers below

Answer 1

Here is the output shapes of each layer in your model

(?, 376, 1344, 3) - Input
(?, 126, 1344, 3) - Cropping2D
(?, 126, 1344, 3) - Lambda
(?, 60, 669, 16)  - Conv2D 1
(?, 27, 332, 32)  - Conv2D 2
(?, 12, 164, 32)  - Conv2D 3
(?, 4, 80, 64)    - Conv2D 4

By the time the inputs have passed through the 4th Conv2D layer the output shape is already (4,80) . You cannot apply another Conv2D layer with filter size (5, 5) since the first dimension of your output is less than the filter size.

Answer 2

Your stack of convolutional layers reduces the image size quite fast. Therefore, once its size along one dimension is only 4, you cannot apply a 5x5-convolution anymore.

Without padding the output dimensions of a convolutional layer is (input_dimension - kernel_size)/strides . Substracting 7 (or 5) multiple times is not that important, but reducing the size by a factor of two gets the dimension down to 4 quite fast.

The solution is either not to use strides (after the first some layers). Adding padding helps against loosing size due to the kernel, but not due to strides.