时间戳

首页 > 解决方案 > ValueError：传递了与“cell.state_size”不兼容的“initial_state”

python-3.x - ValueError：传递了与“cell.state_size”不兼容的“initial_state”

问题描述

我试图在我的模型中添加一个自定义层，但我遇到了这个不合逻辑的错误，我需要改变什么？如果需要自定义层的代码，请告诉我，如果您帮助我了解我做错了什么，这对我来说意味着世界。以及如何解决它？

from keras.models import Sequential , Model
from keras.layers.convolutional_recurrent import ConvLSTM2D
from keras.layers.normalization import BatchNormalization
from keras.layers import Input ,Activation, Dense , Flatten ,RNN
import tensorflow as tf 
from layers.se3 import SE3CompositeLayer  # this is the layer that i want to add 

batch_size = 500

seq = Sequential()

seq.add(ConvLSTM2D(filters=128, kernel_size=(3, 3),
                   batch_input_shape=(500,4,34, 17, 768),
                   padding='same', return_sequences=True,stateful=True,dropout=0.2))
seq.add(BatchNormalization())
seq.add(Activation('relu'))

seq.add(ConvLSTM2D(filters=32, kernel_size=(3, 3),padding='same',dropout=0.2))
seq.add(BatchNormalization())
seq.add(Activation('relu'))

output1 = Flatten()(seq.layers[11].output)

branch1= Dense(64,activation ='relu')(output1)
branch1= Dense(4)(branch1)

branch2= Dense(64,activation='relu')(output1)
branch2= Dense(3)(branch2)

output_conv_lstm =tf.concat([branch2,branch1] ,1)
output_conv_lstm = tf.expand_dims(output_conv_lstm , axis=0)


init_s=tf.placeholder_with_default(tf.eye(4, batch_shape=[batch_size]) , shape=(None, 4 ,4))

l_se3comp = SE3CompositeLayer()

se3_outputs, se3_state =keras.layers.RNN(cell=l_se3comp , dtype=tf.float32 ,unroll=True)(output_conv_lstm ,initial_state=init_s)

model = Model(inputs= seq.layers[0].input ,outputs=se3_outputs )
model.summary()


 File "main.py", line 59, in <module>
    se3_outputs, se3_state =keras.layers.RNN(cell=l_se3comp , dtype=tf.float32 ,unroll=True)(output_conv_lstm ,initial_state=init_s)
  File "/home/ridha/.local/lib/python3.6/site-packages/keras/layers/recurrent.py", line 574, in __call__
    return super(RNN, self).__call__(inputs, **kwargs)
  File "/home/ridha/.local/lib/python3.6/site-packages/keras/engine/base_layer.py", line 431, in __call__
    self.build(unpack_singleton(input_shapes))
  File "/home/ridha/.local/lib/python3.6/site-packages/keras/layers/recurrent.py", line 508, in build
    '{}'.format(self.state_spec, self.cell.state_size))
ValueError: An `initial_state` was passed that is not compatible with `cell.state_size`. Received `state_spec`=[InputSpec(shape=(None, 4, 4), ndim=3)]; however `cell.state_size` is (?, 4, 4) ```

标签： python-3.xtensorflowkerasdeep-linkingrecurrent-neural-network

解决方案

不是真正的答案，但可能是一种解决方法：tensorflow 初始化程序据说可以是张量或可调用对象，但我只有每个可调用对象都可以工作。

例如，如果我想将 RNN 中的内核初始化为所有 7 的 2D 向量，

kernel_initializer = tf.constant([7.,7.])

从来没有为我工作过，但是

init_state_getter = lambda shape, dtype: 7*np.ones(*shape)
...
kernel_initializer = init_state_getter

没关系。

推荐阅读