我目前正在进行一个项目，需要在 Keras 中创建一个新层。这个新层应该能够重新排列输入张量。我已经使用按预期功能的代码成功实现了自定义层。但是，当我将此自定义层集成到模型结构中时，遇到与尺寸相关的错误。
这是自定义图层代码：
# Coustom Layer
import tensorflow as tf
from tensorflow.keras.layers import Layer
import numpy as np
class RearrangeLayer(Layer):
def call(self, inputs):
num_rows, num_columns = tf.shape(inputs)[1], tf.shape(inputs)[2]
# Create indices for gathering columns
indices = tf.range(num_columns)
# Gather even and odd columns
even_columns = tf.gather(inputs, indices[::2], axis=2)
even_columns = tf.reshape(even_columns, (tf.shape(inputs)[0], -1, 2 * even_columns.shape[-1]))
odd_columns = tf.gather(inputs, indices[1::2], axis=2)
odd_columns = tf.reshape(odd_columns, (tf.shape(inputs)[0], -1, 2 * odd_columns.shape[-1]))
# Interleave the rows alternatively
output_data = tf.zeros((1, even_columns.shape[1] + odd_columns.shape[1], even_columns.shape[2]), dtype='float32')
output_data = np.array(output_data)
output_data[0, 0::2] = even_columns
output_data[0, 1::2] = odd_columns
return output_data
为了确保该层按我的预期工作，我输入了一个张量示例并收到了所需的重新排列的输出：
# Example usage
inputs = tf.constant([[
[1, 2, 3, 4, 5, 6],
[7, 8, 9, 10, 11, 12],
[13, 14, 15, 16, 17, 18],
[19, 20, 21, 22, 23, 24]]
], dtype=tf.float32)
rearrange_layer = RearrangeLayer()
output = rearrange_layer(inputs)
print(output)
[[[ 1. 3. 5. 7. 9. 11.]
[ 2. 4. 6. 8. 10. 12.]
[13. 15. 17. 19. 21. 23.]
[14. 16. 18. 20. 22. 24.]]]
现在，当我将自定义层集成到模型架构中时，遇到错误...
# The model architecture
from keras.layers import Input, Flatten, Conv1D, Conv2D, GlobalAveragePooling1D, Dense, Reshape
from keras.models import Model
#define input layer
INPUT = Input((100,60))
reshaped_input = Reshape((100, 60, 1))(INPUT)
conv = Conv2D(1, (2,2), (2,2), activation='relu')(reshaped_input)
conv = RearrangeLayer()(conv)
conv = Conv2D(16, (2,2), (2,1), activation='relu')(conv)
conv = Conv2D(32, (2,2), (2,1), activation='relu')(conv)
flatted = Flatten()(conv)
reshaped = Reshape((10752, 1))(flatted)
conv = Conv1D(32, 6, activation='relu')(reshaped)
pooling = GlobalAveragePooling1D()(conv)
output = Dense(1024, activation='relu')(pooling)
output = Dense(1)(output)
model = Model(inputs=INPUT, outputs=output)
# The error
TypeError Traceback (most recent call last)
Cell In[3], line 9
7 reshaped_input = Reshape((100, 60, 1))(INPUT)
8 conv = Conv2D(1, (2,2), (2,2), activation='relu')(reshaped_input)
----> 9 conv = RearrangeLayer()(conv)
10 conv = Conv2D(16, (2,2), (2,1), activation='relu')(conv)
11 conv = Conv2D(32, (2,2), (2,1), activation='relu')(conv)
File ~\AppData\Roaming\Python\Python310\site-packages\keras\utils\traceback_utils.py:70, in filter_traceback.<locals>.error_handler(*args, **kwargs)
67 filtered_tb = _process_traceback_frames(e.__traceback__)
68 # To get the full stack trace, call:
69 # `tf.debugging.disable_traceback_filtering()`
---> 70 raise e.with_traceback(filtered_tb) from None
71 finally:
72 del filtered_tb
File ~\AppData\Local\Temp\__autograph_generated_filen_2mzai2.py:16, in outer_factory.<locals>.inner_factory.<locals>.tf__call(self, inputs)
14 odd_columns = ag__.converted_call(ag__.ld(tf).gather, (ag__.ld(inputs), ag__.ld(indices)[1::2]), dict(axis=2), fscope)
15 odd_columns = ag__.converted_call(ag__.ld(tf).reshape, (ag__.ld(odd_columns), (ag__.converted_call(ag__.ld(tf).shape, (ag__.ld(inputs),), None, fscope)[0], -1, 2 * ag__.ld(odd_columns).shape[-1])), None, fscope)
---> 16 output_data = ag__.converted_call(ag__.ld(tf).zeros, ((1, ag__.ld(even_columns).shape[1] + ag__.ld(odd_columns).shape[1], ag__.ld(even_columns).shape[2]),), dict(dtype='float32'), fscope)
17 output_data = ag__.converted_call(ag__.ld(np).array, (ag__.ld(output_data),), None, fscope)
18 ag__.ld(output_data)[0, 0::2] = ag__.ld(even_columns)
TypeError: Exception encountered when calling layer "rearrange_layer_1" (type RearrangeLayer).
in user code:
File "C:\Users\dpc\AppData\Local\Temp\ipykernel_1928\1750308305.py", line 19, in call *
output_data = tf.zeros((1, even_columns.shape[1] + odd_columns.shape[1], even_columns.shape[2]), dtype='float32')
TypeError: unsupported operand type(s) for +: 'NoneType' and 'NoneType'
Call arguments received by layer "rearrange_layer_1" (type RearrangeLayer):
• inputs=tf.Tensor(shape=(None, 50, 30, 1), dtype=float32)
您遇到的错误是由于自定义 Keras 层中的维度不匹配造成的。 TypeError: unsupported operand type(s) for +: 'NoneType' and 'NoneType' 表示 Even_columns 和 odd_columns 的形状未正确推断，导致 NoneType 维度。这会导致 tf.zeros 函数中后续的加法操作失败。 要解决此问题，您需要确保正确推断出 Even_columns 和 odd_columns 的形状。您可以通过使用 tf.shape 函数显式获取张量的形状来实现此目的。这是自定义图层的修改后的调用方法：
class RearrangeLayer(Layer):
def call(self, inputs):
num_rows, num_columns = tf.shape(inputs)[1], tf.shape(inputs)[2]
# Create indices for gathering columns
indices = tf.range(num_columns)
# Gather even and odd columns
even_columns = tf.gather(inputs, indices[::2], axis=2)
even_columns = tf.reshape(even_columns, (tf.shape(inputs)[0], -1, 2 * tf.shape(even_columns)[2]))
odd_columns = tf.gather(inputs, indices[1::2], axis=2)
odd_columns = tf.reshape(odd_columns, (tf.shape(inputs)[0], -1, 2 * tf.shape(odd_columns)[2]))
# Interleave the rows alternatively
rearranged_output = tf.concat([even_columns, odd_columns], axis=1)
return rearranged_output
from keras.layers import Input, Flatten, Conv1D, Conv2D, GlobalAveragePooling1D, Dense, Reshape
from keras.models import Model
# The model architecture
from keras.layers import Input, Flatten, Conv1D, Conv2D, GlobalAveragePooling1D, Dense, Reshape
from keras.models import Model
#define input layer
INPUT = Input((100,60))
reshaped_input = Reshape((100, 60, 1))(INPUT)
conv = Conv2D(1, (2,2), (2,2), activation='relu')(reshaped_input)
conv = RearrangeLayer()(conv)
conv = Reshape((50, 30, 1))(conv) # Add Reshape layer to reshape output tensor
conv = Conv2D(16, (2,2), (2,1), activation='relu')(conv)
conv = Conv2D(32, (2,2), (2,1), activation='relu')(conv)
flatted = Flatten()(conv)
reshaped = Reshape((10752, 1))(flatted)
conv = Conv1D(32, 6, activation='relu')(reshaped)
pooling = GlobalAveragePooling1D()(conv)
output = Dense(1024, activation='relu')(pooling)
output = Dense(1)(output)
model = Model(inputs=INPUT, outputs=output)