keras ConvLSTM2D 的简单应用

示例1： 返回各时刻状态 

import tensorflow as tf import numpy as np import keras from keras.layers import ConvLSTM2D lstm_input = np.random.random((4,6,30,30,3)).astype(np.float32) lstm_input = tf.convert_to_tensor(lstm_input) lstm_out1 = ConvLSTM2D(filters=1,kernel_size=[5,5],strides=(1,1),padding='valid',activation='relu', input_shape=(6,30,30,3),return_sequences=True)(lstm_input) lstm_out2 = ConvLSTM2D(filters=2,kernel_size=[5,5],strides=(1,1),padding='valid',activation='relu', return_sequences=True)(lstm_out1) lstm_out3 = ConvLSTM2D(filters=3,kernel_size=[5,5],strides=(1,1),padding='valid',activation='relu', return_sequences=True)(lstm_out2) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) lstm_out1_,lstm_out2_,lstm_out3_ = sess.run([lstm_out1,lstm_out2,lstm_out3]) print(lstm_out1_.shape) print(lstm_out2_.shape) print(lstm_out3_.shape) """ 返回： (4, 6, 26, 26, 1) (4, 6, 22, 22, 2) (4, 6, 18, 18, 3) """

备注：

return_sequences: 默认是False，控制LSTM的输出：

False: 仅返回最后一个时刻的状态（hidden_state），shape = [B,H,W,C]，是一个4维张量True: 返回所有时刻的输出（hidden_state），shape = [B,Clip_len,H,W,C]，是一个5维张量

示例2： 同时返回各时刻状态与最后一个时刻的state(包含state.h,state.c)

用return_state=True控制 

import tensorflow as tf import numpy as np import keras from keras.layers import ConvLSTM2D lstm_input = np.random.random((4,6,30,30,3)).astype(np.float32) lstm_input = tf.convert_to_tensor(lstm_input) lstm_out,state_h,state_c = ConvLSTM2D(filters=1,kernel_size=[5,5],strides=(1,1),padding='valid',activation='relu', batch_input_shape=(-1,6,30,30,3),return_sequences=False,return_state=True)(lstm_input) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) lstm_out_,state_h_,state_c_= sess.run([lstm_out,state_h,state_c]) print(lstm_out_==state_h_) print(lstm_out_.shape) print(state_h_.shape) print(state_c_.shape) """ 返回： [ True]]]] (4, 26, 26, 1) (4, 26, 26, 1) (4, 26, 26, 1) """

综上：

return_sequences: 决定是否返回所有时刻的状态

return_state:决定是否返回最后一个时刻的cell状态，由示例2结果可见，最后一个时刻的state = [h,c]

注意：在keras 中文文档中，在介绍ConvLSTM2D时，没有介绍 return_state 参数，该参数在LSTM的介绍中介绍，但是在ConvLSTM2D中通用。。。自己差点以为ConvLSTM2D中没有这个功能。。。