如何在多图像输入中使用Conv2D？

Question

我想使用多个图像作为网络的输入。我想添加Conv2D图层，类似这样的：from tensorflow.keras.layers import *from tensorflow.keras.models import Sequentialmodel = Sequential([    Input(shape=(1, 128, 128, 1)),    Conv2D(32, 3),    Flatten(),])但这段代码会引发错误：Input 0 of layer conv2d_40 is incompatible with the layer: expected ndim=4, found ndim=5. Full shape received: [None, 1, 128, 128, 1]但下面的代码工作正常：model = Sequential([    Input(shape=(1, 512, 512, 1)),    Dense(32),    Flatten(),])我知道，我可以添加多个Input图层，但我想知道有没有办法做到这样？我的意思是我想使用输入形状的数据[NUMBER_OF_IMAGES, WIDTH, HEIGHT, N_CHANNELS]并且NUMBER_OF_IMAGES不是所有图像的数量。这是当前输入的金额

Answer 1

Conv2D需要 4D 输入，您无法更改它。我不太确定你想要完成什么，但你可以使用Conv3D：

from tensorflow.keras.layers import *

from tensorflow.keras.models import Sequential

import tensorflow as tf

model = Sequential([

    Input(shape=(None, 128, 128, 1)),

    Conv3D(32, kernel_size=(1, 3, 3)),

    Flatten()

])

multiple_images = tf.random.uniform((10, 10, 128, 128, 1), dtype=tf.float32)

model(multiple_images)

<tf.Tensor: shape=(10, 5080320), dtype=float32, numpy=

array([[-0.26742983, -0.09689523, -0.12120364, ..., -0.02987139,

         0.05515741,  0.12026916],

       [-0.18898709,  0.12448274, -0.17439063, ...,  0.23424357,

        -0.06001307, -0.13852882],

       [-0.14464797,  0.26356792, -0.34748033, ...,  0.07819699,

        -0.11639086,  0.10701762],

       ...,

       [-0.1536693 ,  0.13642962, -0.18564   , ...,  0.07165999,

        -0.0173855 , -0.04348694],

       [-0.32320747,  0.09207243, -0.22274591, ...,  0.11940736,

        -0.02635285, -0.1140241 ],

       [-0.21126074, -0.00094431, -0.10933039, ...,  0.06002581,

        -0.09649743,  0.09335127]], dtype=float32)>