关于tfkan中Conv3DKAN函数的使用

[JunXinVan]

您好！在使用Conv3DKAN函数的过程中，我发现五维的输入经过函数处理后变为了四维，请问这是为什么？您辛苦了！

[ZPZhou-lab]

I'm not sure how you configure your Conv3DKAN layer.

The following is my code:

import tensorflow as tf
from tfkan.layers import Conv3DKAN

# x has (batch, spatial_1, spatial_2, spatial_3, channels) shape
x = tf.random.normal(shape=(1, 10, 32, 32, 3))
layer = Conv3DKAN(filters=8, kernel_size=3, strides=1)
y = layer(x)
print(y.shape)

and the output is

TensorShape([1, 8, 30, 30, 8])

which has 5D

[JunXinVan]

当我运行您的示例时，我能得到同样的结果，但当我自己编写代码时，却总会遇到维度下降的问题：

def TokKANBlock(input_tensor, filters, kernel_size=(3, 3, 3), strides=(1, 1, 1)):
    # 第一层标准 3D 卷积
    x = tf.keras.layers.Conv3D(filters=filters, kernel_size=kernel_size, strides=strides, padding='same')(input_tensor)
    print(x.shape)

    # 第二层使用 KAN 的 3D 卷积
    x = Conv3DKAN(filters=filters, kernel_size=3, strides=1, padding="same")(x)
    print(x.shape)

    x = tf.keras.layers.Conv3D(filters=filters, kernel_size=kernel_size, strides=strides, padding='same')(x)
    print(x.shape)

    return x


def autoKAN(n_filters=16, n_classes=1, dropout_rate=0, grid_size=3, spline_order=3):
    inputs = tf.keras.layers.Input(shape=[None, None, None, 1])
    kan = TokKANBlock(inputs, filters=n_filters * 8, kernel_size=(3, 3, 3), strides=(1, 1, 1))


model1 = autoKAN()

#以下为运行结果

(None, None, None, None, 128)
(None, None, None, 128)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-40-9423babc311f> in <cell line: 21>()
     19 
     20 
---> 21 model1 = autoKAN()

3 frames
/usr/local/lib/python3.10/dist-packages/keras/src/engine/input_spec.py in assert_input_compatibility(input_spec, inputs, layer_name)
    251             ndim = x.shape.rank
    252             if ndim is not None and ndim < spec.min_ndim:
--> 253                 raise ValueError(
    254                     f'Input {input_index} of layer "{layer_name}" '
    255                     "is incompatible with the layer: "
    
ValueError: Input 0 of layer "conv3d_40" is incompatible with the layer: expected min_ndim=5, found ndim=4. Full shape received: (None, None, None, 128)

[ZPZhou-lab]

I have find where the problem is, we use the API of tensorflow to extract patches when processing 3D inputs(with (spatial_1, spatial_2, spatial_3) as follows:

patches = tf.extract_volume_patches(
    inputs,
    ksizes=[1, *self.kernel_size, 1],
    strides=[1, *self.strides, 1],
    padding=self.padding
)

I found that this API reduces one dimension when encountering undetermined tensors (virtual inputs created through tf.keras.layers.Input), resulting in the final output dimension not meeting expectations.😫

I will consider how to fix this issue. I'm sorry, before that, you can only use tensors of a certain size to initialize your model.

[JunXinVan]

感谢您百忙中抽出时间进行回答，我理解了，您辛苦了！谢谢！