convLSTM网络中的参数含义及计算问题

参数代表的含义

先贴一段模型代码代码

from keras.layers import (Input,ConvLSTM2D)
from keras.models import Model
from keras.models import Sequential

def Seq():
    '''
    input_shape为(time_steps, map_height, map_width, channels)
     time_steps 就是将一个样例分为多少个时间点读入，x1,x2...,xt，的t
     map_height, map_width, channels分别为输入图像的长、宽、高
     return_sequences为True时每一个时间点都有输出
     return_sequences为False时，只有最后一个时间点有输出
    '''
    seq = Sequential()
    seq.add(ConvLSTM2D(filters=30, kernel_size=(3, 3),input_shape=(15, 40, 40, 3),  \
                       padding='same', return_sequences=True,data_format='channels_last'))
	seq.add(ConvLSTM2D(filters=50, kernel_size=(3, 3),   \
                       padding='same', return_sequences=True,data_format='channels_last'))
	seq.add(ConvLSTM2D(filters=60, kernel_size=(3, 3),  \
                       padding='same', return_sequences=True,data_format='channels_last'))
    seq.add(ConvLSTM2D(filters=70, kernel_size=(3, 3),  \
    				  padding='same', return_sequences=False,data_format='channels_last'))
    seq.summary()
    
def main():

	'''	
     模型的另一种搭建形式
    '''
    Inputs=[]
    Outputs=[]
    input = Input(shape=(15, 40, 40, 3))
    Inputs.append(input)
    convlstm1 = ConvLSTM2D(filters=30, kernel_size=(3,3),padding='same',
     	 				   return_sequences=True,data_format='channels_last')(input)
    convlstm2 = ConvLSTM2D(filters=50, kernel_size=(3,3),padding='same',
    					   return_sequences=True, data_format='channels_last')(convlstm1)
    convlstm3 = ConvLSTM2D(filters=60, kernel_size=(3, 3),padding='same',
    					   return_sequences=True, data_format='channels_last')(convlstm2)
    convlstm4 = ConvLSTM2D(filters=70, kernel_size=(3, 3),padding='same',
                           return_sequences=False, data_format='channels_last')(convlstm3)
    Outputs.append(convlstm4)
    model =Model(inputs=input, outputs=convlstm4)
    model.summary()
if __name__ == '__main__':
    Seq()
    main()

运行结果如下所示

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv_lst_m2d_1 (ConvLSTM2D)  (None, 15, 40, 40, 30)    35760     
_________________________________________________________________
conv_lst_m2d_2 (ConvLSTM2D)  (None, 15, 40, 40, 50)    144200    
_________________________________________________________________
conv_lst_m2d_3 (ConvLSTM2D)  (None, 15, 40, 40, 60)    237840    
_________________________________________________________________
conv_lst_m2d_4 (ConvLSTM2D)  (None, 40, 40, 70)        327880    
=================================================================
Total params: 745,680
Trainable params: 745,680
Non-trainable params: 0
_________________________________________________________________
Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 15, 40, 40, 3)     0         
_________________________________________________________________
conv_lst_m2d_5 (ConvLSTM2D)  (None, 15, 40, 40, 30)    35760     
_________________________________________________________________
conv_lst_m2d_6 (ConvLSTM2D)  (None, 15, 40, 40, 50)    144200    
_________________________________________________________________
conv_lst_m2d_7 (ConvLSTM2D)  (None, 15, 40, 40, 60)    237840    
_________________________________________________________________
conv_lst_m2d_8 (ConvLSTM2D)  (None, 40, 40, 70)        327880    
=================================================================
Total params: 745,680
Trainable params: 745,680
Non-trainable params: 0
_________________________________________________________________

LSTM模型的图如下所示

如上图所示，在LSTM的计算流程中，只有图中所示的4个部分需要参数，取其中之一展开，最后参数量$\times 4$即可。

上图是第一层卷积的示意图，卷积核$filter$个数是30。
由$8940\times 4=35760$，即得到模型的$summary$输出结果中的35760。

设置第一层的$retur{n}_{s}equences=True$，可以将第一层的输出作为第二层的convLSTM的输入，输入的形状为$40\times 40\times 30$，卷积核$filter=40$

由$36050\times 4=144200$，即得到模型的$summary$输出结果中的144200。
后面的层的计算以此类推即可。