-
Notifications
You must be signed in to change notification settings - Fork 0
/
policy_gradient.py
128 lines (92 loc) · 5.43 KB
/
policy_gradient.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
import numpy as np
import tensorflow as tf
import gym
n_inputs = 4 #x, xdot, theta, thetadot
n_hidden = 4 #hidden nodes
n_outputs = 1 #probab
learning_rate = 0.01
initializer = tf.contrib.layers.variance_scaling_initializer()
X = tf.placeholder(tf.float32, shape=[None, n_inputs])
hidden_layer = tf.layers.dense(X, n_hidden, activation=tf.nn.relu, kernel_initializer=initializer)
logits = tf.layers.dense(hidden_layer, n_outputs)
outputs = tf.nn.sigmoid(logits) # P(Left)
probabilties = tf.concat(axis=1, values=[outputs, 1 - outputs])
action = tf.multinomial( probabilties, num_samples=1) #random sampling
y = 1. - tf.to_float(action) #tensor to float
cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=logits)
optimizer = tf.train.AdamOptimizer(learning_rate)
gradients_and_variables = optimizer.compute_gradients(cross_entropy) #compute gradients to multiply gradients with discount, not minimizing optimizer directly
gradients = []
gradient_placeholders = []
grads_and_vars_feed = []
for gradient, variable in gradients_and_variables:
gradients.append(gradient)
gradient_placeholder = tf.placeholder(tf.float32, shape=gradient.get_shape()) #single
gradient_placeholders.append(gradient_placeholder)
grads_and_vars_feed.append((gradient_placeholder, variable)) #list of tuples
training_operation = optimizer.apply_gradients(grads_and_vars_feed) #final feed fed
init = tf.global_variables_initializer()
saver = tf.train.Saver() #to save model later on
def helper_discount_rewards(rewards, discount_rate): #helper function takes in rewards and applies discount rate, can be 0.95-0.99 etc
discounted_rewards = np.zeros(len(rewards)) #len(rewards) to 0 in reverse
cumulative_rewards = 0
for step in reversed(range(len(rewards))):
cumulative_rewards = rewards[step] + cumulative_rewards*discount_rate
discounted_rewards[step] = cumulative_rewards
return discounted_rewards
def discount_and_normalize_rewards(all_rewards, discount_rate): #takes in all rewards, applies helper_discount function and then normalizes using mean and std deviation
all_discounted_rewards = []
for rewards in all_rewards:
all_discounted_rewards.append(helper_discount_rewards(rewards,discount_rate))
flattened_rewards = np.concatenate(all_discounted_rewards)
reward_mean = flattened_rewards.mean()
reward_std = flattened_rewards.std()
return [(discounted_rewards - reward_mean)/reward_std for discounted_rewards in all_discounted_rewards]
#Training
env = gym.make("CartPole-v0") #make the environment
n_game_rounds = 10
max_game_steps = 1000 #time steps before we have to manually day it's done
n_iterations = 500
discount_rate = 0.9
with tf.Session() as sess:
sess.run(init)
for iteration in range(n_iterations):
print("Currently on iteration: {} \n".format(iteration))
all_rewards = []
all_gradients = []
for game in range(n_game_rounds): #play n game rounds
current_rewards = []
current_gradients = []
obs = env.reset() #reset environment to default state, that's our first observation
for step in range(max_game_steps): #manual cutoff
cart_pos, cart_vel, pole_ang, ang_vel = obs
action_val, gradients_val = sess.run([action, gradients], feed_dict={X: obs.reshape(1, n_inputs)}) #get actions and gradients, reshape comma
obs, reward, done, info = env.step(action_val[0][0]) #perform action by passing into step function, get observations, reward, boolean indicating whether environment needs to be reset, debug info
current_rewards.append(reward) #get current rewards and gradients
current_gradients.append(gradients_val)
if done: #pole fell over (game ended)
break
all_rewards.append(current_rewards) #append to list of all rewards
all_gradients.append(current_gradients)
all_rewards = discount_and_normalize_rewards(all_rewards,discount_rate) #applying helper and normalizing
feed_dict = {}
for var_index, gradient_placeholder in enumerate(gradient_placeholders): #enumerate gives back index locations
mean_gradients = np.mean([reward * all_gradients[game_index][step][var_index] #multiplying correct reward with the correct gradient
for game_index, rewards in enumerate(all_rewards)
for step, reward in enumerate(rewards)], axis=0)
feed_dict[gradient_placeholder] = mean_gradients
sess.run(training_operation, feed_dict=feed_dict)
print('Saving graph and session')
meta_graph_def = tf.train.export_meta_graph(filename='/models/my-policy-gradient-model.meta') #exporting graphs if we need it in another file
saver.save(sess, '/models/my-policy-gradient-model') #saving actual session
#Running this trained model on environment
env = gym.make('CartPole-v0') #make the environment
obs = env.reset()
with tf.Session() as sess:
new_saver = tf.train.import_meta_graph('/models/my-policy-gradient-model.meta')
new_saver.restore(sess,'/models/my-policy-gradient-model')
for x in range(500):
env.render()
action_val, gradients_val = sess.run([action, gradients], feed_dict={X: obs.reshape(1, n_inputs)})
obs, reward, done, info = env.step(action_val[0][0])
#end