具有约束权重的keras层在初始化模型时会增加错误

Question

我正在训练一个在actor网络中具有约束层的actor-critic模型。约束强制权重为对角线。下面是一个最小可复制的示例

import tensorflow.compat.v1 as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense, Input, Conv1D, Concatenate, BatchNormalization, Reshape
from tensorflow.keras.constraints import Constraint
from tensorflow.python.keras.utils.vis_utils import plot_model
from tensorflow.keras.optimizers import Adam
from tensorflow.python.keras.backend import set_session
from tensorflow.python.keras import backend as K

import numpy as np
import random
from collections import deque
tf.disable_v2_behavior() 
# For more repetitive results
np.random.seed(1)
random.seed(1)
class DiagonalWeight(Constraint):
"""Constrains the weights to be diagonal.
"""
def __call__(self, w):
N = K.int_shape(w)[-1]
m = K.eye(N)
return w*m
state_shape = (10,3)
class AC():
def __init__(self, sess, LRA, LRC):
self.sess = sess     #session

self.LRA = LRA                                   #learning rate for actor
self.LRC = LRC                                   #learning rate for critic

self.graph = tf.get_default_graph()
set_session(self.sess)
# training actor and target actor 
self.actor, self.input_actor = self.create_actor()
plot_model(self.actor, to_file='Ac_architecture.png', show_shapes=True, show_layer_names=True)
self.target_actor, _ = self.create_actor()
###initialize the weights of the target with the weights of training actor
self.target_actor.set_weights(self.actor.get_weights())
#training critic and target critic
self.critic, self.critic_state_input, self.critic_action_input = self.create_critic()
self.target_critic, _, _  = self.create_critic()
plot_model(self.critic, to_file='Cr_architecture.png', show_shapes=True, show_layer_names=True)
###initialize the weights of the target with the weights of training critic
self.target_critic.set_weights(self.critic.get_weights())

######################## Actor/Critic Grads ########################################################
self.actor_critic_grad = tf.placeholder(tf.float32, [None, state_shape[0], state_shape[0]+2])   
actor_weights = self.actor.trainable_weights
self.actor_grads = tf.gradients(self.actor.output, actor_weights, -self.actor_critic_grad) 
grads = zip(self.actor_grads, actor_weights)
self.optimize = tf.train.AdamOptimizer(self.LRA).apply_gradients(grads) 
self.critic_grads = tf.gradients(self.critic.output, self.critic_action_input)
# Initialize for later gradient calculations
self.sess.run(tf.global_variables_initializer())
#######################################################################################################################
def create_actor(self):
actor_input = Input(shape=state_shape, name='state_input')     
h3 = Conv1D(128, 3, padding='same', activation='relu', name='h3')(actor_input)
h3 = BatchNormalization(name='h3_BN')(h3)
matrix = Conv1D(state_shape[0], 3, padding='same', activation='relu', name='matrix')(h3)
vect0= Conv1D(1, 3, padding='same', activation='relu', name='vect0')(h3)
vect0_resh = Reshape((1, state_shape[0]))(vect0)
vect1 = Dense(state_shape[0], activation='relu', name='vect1',
use_bias=False, kernel_constraint=DiagonalWeight())(vect0_resh) #, kernel_constraint=DiagonalWeight()
vect1 = Reshape((state_shape[0],1))(vect1)
actor_output = Concatenate(axis=-1)([vect0, vect1, matrix]) 
model = Model(actor_input, actor_output)  
adam  = Adam(lr=self.LRA)
model.compile(loss="mse", optimizer=adam)
return model, actor_input
def create_critic(self):
state_input = Input(shape=state_shape, name='state_input')     
action_input = Input(shape=(state_shape[0],state_shape[0]+2), name='action_input')                                           
critic_input = Concatenate(axis=-1)([state_input, action_input]) 
h3 = Conv1D(128, 3, padding='same', name='h3')(critic_input)
h3 = BatchNormalization(name='h3_BN')(h3)
Q = Conv1D(state_shape[0]+2, 3, padding='same', name='Q')(h3)
model = Model([state_input, action_input], Q)
adam  = Adam(lr=self.LRC)
model.compile(loss="mse", optimizer=adam)
return model, state_input, action_input

sess = tf.compat.v1.Session()
K.set_session(sess)
agent = AC(sess, 0.01, 0.001)

当我从密集层中移除内核约束时，一切都很好，但当我添加它时，我会得到以下三个错误之一：

tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_19' with dtype float and shape [?,10,12]
[[node Placeholder_19 (defined at AC.py:59) ]]

或

tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'state_input' with dtype float and shape [?,10,3]
[[node state_input (defined at AC.py:73) ]]

或

tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable beta1_power from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost/beta1_power/class tensorflow::Var does not exist.
[[node Adam/update_vect1/kernel/ResourceApplyAdam/ReadVariableOp (defined at AC.py:64) ]]

稠密层中的核是一个正方形矩阵，因此应用对角约束应该不会产生问题。

Answer 1

对于任何对解决方案感兴趣的人：在class DiagonalWeight中，我用m = tf.eye(N)更改了线路m = K.eye(N)

我不知道这到底是怎么回事，但我认为这与会议有关。Tensorflow需要存储操作，以便将来在培训中使用。