对于连续的卷积层，filter 的维度是跟输入图像的维度一致
model = Sequential([
Conv2D(8, 3, input_shape=(28, 28, 1), use_bias=False),
Conv2D(16, 3, use_bias=False)
])
model.summary()
输出
Model: "sequential_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d_3 (Conv2D) (None, 26, 26, 8) 72 _________________________________________________________________ conv2d_4 (Conv2D) (None, 24, 24, 16) 1152 ================================================================= Total params: 1,224 Trainable params: 1,224 Non-trainable params: 0 _________________________________________________________________
其中：
第一层的filter为 3x3x1x8=72（原始数据是 28x28x1，得到数据 26x26x8）
第二层的filter为 3x3x8x16=1152（上一个数据是 26x26x8，得到数据 24x24x16）
原文：https://www.cnblogs.com/alex-bn-lee/p/14901087.html