critic_network2.py

import tensorflow as tf
import tensorflow.contrib as tc
import numpy as np
import math
from configuration import *

class CriticNetwork:
    """docstring for CriticNetwork"""

    def __init__(self, sess, state_dim, action_dim, config):
        self.layer1_size = config.conf['critic-layer-size'][0]
        self.layer2_size = config.conf['critic-layer-size'][1]
        self.learning_rate = config.conf['critic-lr']
        self.tau = config.conf['tau']
        self.l2 = config.conf['critic-l2-reg']
        self.is_prioritized_replay = config.conf['prioritized-exp-replay']
        self.is_layer_norm = config.conf['critic-layer-norm']
        self.is_observation_norm = config.conf['critic-observation-norm']
        if config.conf['critic-activation-fn'] == "relu":
            self.activation_fn = tf.nn.relu
        elif config.conf['critic-activation-fn'] == "elu":
            self.activation_fn = tf.nn.elu
        else:
            self.activation_fn = None


        self.time_step = 0
        self.sess = sess
        # create q network
        self.critic_network = {}
        self.state_input, \
        self.action_input, \
        self.q_value_output, \
        self.critic_network['vars'], \
        self.critic_network['trainable_vars'], \
        self.critic_network['perturbable_vars'],\
        self.is_training = self.create_q_network(state_dim, action_dim, 'critic_network')

        # create target q network (the same structure with q network)
        self.target_critic_network = {}
        self.target_state_input, \
        self.target_action_input, \
        self.target_q_value_output, \
        self.target_critic_network['vars'], \
        self.target_critic_network['trainable_vars'], \
        self.target_critic_network['perturbable_vars'],\
        self.target_is_training = self.create_q_network(state_dim, action_dim, 'target_critic_network')

        self.create_training_method()
        self.setup_target_network_updates()
        self.sess.run(tf.global_variables_initializer())

        self.sess.run(self.target_init_updates)
        #self.load_network()

    def create_training_method(self):
        # Define training optimizer
        self.y_input = tf.placeholder("float", [None, 1])
        weight_decay = tf.add_n([self.l2 * tf.nn.l2_loss(var) for var in self.critic_network['trainable_vars']])
        if self.is_prioritized_replay == True:
            self.ISWeights = tf.placeholder(tf.float32, [None, 1], name='IS_weights')#importance sampling for experience replay
            self.abs_errors = tf.reduce_sum(tf.abs(self.y_input - self.q_value_output), axis=1)  # TD_error for updating Sumtree
            self.cost = tf.reduce_mean(self.ISWeights * tf.square(self.y_input - self.q_value_output)) + weight_decay
            #self.cost = tf.reduce_mean(self.ISWeights * tf.squared_difference(self.y_input, self.q_value_output)) + weight_decay
        else:
            self.cost = tf.reduce_mean(tf.square(self.y_input - self.q_value_output)) + weight_decay

        self.optimizer = tf.train.AdamOptimizer(self.learning_rate).minimize(self.cost)
        self.action_gradients = tf.gradients(self.q_value_output, self.action_input)

    def create_q_network(self, state_dim, action_dim, name):
        # the layer size could be changed
        layer1_size = self.layer1_size
        layer2_size = self.layer2_size
        with tf.variable_scope(name) as scope:
            state_input = tf.placeholder("float", [None, state_dim])
            action_input = tf.placeholder("float", [None, action_dim])
            is_training = tf.placeholder(tf.bool, None)

            W1 = self.variable([state_dim, layer1_size], state_dim)
            b1 = self.variable([layer1_size], state_dim)
            W2 = self.variable([layer1_size, layer2_size], layer1_size + action_dim)
            W2_action = self.variable([action_dim, layer2_size], layer1_size + action_dim)
            b2 = self.variable([layer2_size], layer1_size + action_dim)
            W3 = tf.Variable(tf.random_uniform([layer2_size, 1], -3e-3, 3e-3))
            b3 = tf.Variable(tf.random_uniform([1], -3e-3, 3e-3))

            if self.is_layer_norm == True:
                layer1 = tf.matmul(state_input, W1) + b1
                layer1_norm = tc.layers.layer_norm(layer1, center=True, scale=True)#, activation_fn=tf.nn.relu)#normalize layer before activation function?
                #layer1_norm = tf.nn.relu(layer1_norm)
                layer1_norm = self.activation_fn(layer1_norm)
                layer2 = (tf.matmul(layer1_norm, W2) + tf.matmul(action_input, W2_action) + b2)
                layer2_norm = tc.layers.layer_norm(layer2, center=True, scale=True)#, activation_fn=tf.nn.relu)
                #layer2_norm = tf.nn.relu(layer2_norm)
                layer2_norm = self.activation_fn(layer2_norm)
                q_value_output = tf.identity(tf.matmul(layer2_norm, W3) + b3)
            else:
                layer1 = tf.matmul(state_input, W1) + b1
                layer1 = self.activation_fn(layer1)#tf.nn.relu(layer1)
                layer2 = (tf.matmul(layer1, W2) + tf.matmul(action_input, W2_action) + b2)
                layer2 = self.activation_fn(layer2)#tf.nn.relu(layer2)
                q_value_output = tf.identity(tf.matmul(layer2, W3) + b3)

        vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=name)
        trainable_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=name)
        perturbable_vars = [var for var in trainable_vars if 'LayerNorm' not in var.name]
        return state_input, action_input, q_value_output, vars, trainable_vars, perturbable_vars, is_training

    def update_target(self):
        self.sess.run(self.target_soft_updates)

    def train(self, y_batch, state_batch, action_batch, ISWeights):
        self.time_step += 1
        abs_errors = []
        if self.is_prioritized_replay == True:
            _, abs_errors = self.sess.run([self.optimizer, self.abs_errors], feed_dict={
                self.y_input: y_batch,
                self.state_input: state_batch,
                self.action_input: action_batch,
                self.ISWeights: ISWeights,
                self.is_training: True
            })
        else:
            self.sess.run(self.optimizer, feed_dict={
                self.y_input: y_batch,
                self.state_input: state_batch,
                self.action_input: action_batch,
                self.is_training: True
            })
        return abs_errors

    def gradients(self, state_batch, action_batch):
        return self.sess.run(self.action_gradients, feed_dict={
            self.state_input: state_batch,
            self.action_input: action_batch,
            self.is_training: False
        })[0]

    def q_value_target(self, state_batch, action_batch):
        return self.sess.run(self.target_q_value_output, feed_dict={
            self.target_state_input: state_batch,
            self.target_action_input: action_batch,
            self.target_is_training: False
        })

    def q_value(self, state_batch, action_batch):
        return self.sess.run(self.q_value_output, feed_dict={
            self.state_input: state_batch,
            self.action_input: action_batch,
            self.is_training: False
        })

    def get_target_updates(self, vars, target_vars, tau):
        soft_updates = []
        init_updates = []
        assert len(vars) == len(target_vars)
        for var, target_var in zip(vars, target_vars):
            init_updates.append(tf.assign(target_var, var))
            soft_updates.append(tf.assign(target_var, (1. - tau) * target_var + tau * var))
        assert len(init_updates) == len(vars)
        assert len(soft_updates) == len(vars)
        return tf.group(*init_updates), tf.group(*soft_updates)

    def setup_target_network_updates(self):
        actor_init_updates, actor_soft_updates = self.get_target_updates(self.critic_network['vars'], self.target_critic_network['vars'], self.tau)
        self.target_init_updates = actor_init_updates
        self.target_soft_updates = actor_soft_updates

    def batch_norm_layer(self, x, training_phase, scope_bn, activation=None):
        return tf.cond(training_phase,
                       lambda: tc.layers.batch_norm(x, activation_fn=activation, center=True, scale=True,
                                                            updates_collections=None, is_training=True, reuse=None,
                                                            scope=scope_bn, decay=0.9, epsilon=1e-5),
                       lambda: tc.layers.batch_norm(x, activation_fn=activation, center=True, scale=True,
                                                            updates_collections=None, is_training=False, reuse=True,
                                                            scope=scope_bn, decay=0.9, epsilon=1e-5))

    # f fan-in size
    def variable(self, shape, f):
        return tf.Variable(tf.random_uniform(shape, -1 / math.sqrt(f), 1 / math.sqrt(f)))
        #return tf.Variable(tf.random_normal(shape))

    def add_layer(inputs, in_size, out_size, normalization=None, activation_fn=None, dropout=None):
        Weights = tf.Variable(tf.random_normal([in_size, out_size]))
        # biases = tf.Variable(tf.zeros([1, out_size]) + 0.1)
        biases = tf.Variable(tf.random_normal([1, out_size]))
        Wx_plus_b = tf.matmul(inputs, Weights) + biases
        if activation_fn is None:
            outputs = Wx_plus_b
        else:
            outputs = activation_fn(Wx_plus_b)

        return outputs

    def load_network(self, dir_path):
        self.saver = tf.train.Saver()
        checkpoint = tf.train.get_checkpoint_state(dir_path + '/saved_critic_networks')
        if checkpoint and checkpoint.model_checkpoint_path:
            self.saver.restore(self.sess, checkpoint.model_checkpoint_path)
            print("Successfully loaded:"+checkpoint.model_checkpoint_path)
        else:
            print("Could not find old network weights")

    def save_network(self,time_step, dir_path):
        print('save critic-network...'+str(time_step))
        self.saver.save(self.sess, dir_path + '/saved_critic_networks/' + 'critic-network')#, global_step = time_step)
        #self.saver.save(self.sess, 'critic-network')