-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
299 lines (285 loc) · 12.5 KB
/
train.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
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
import os
import gym
import random
import numpy as np
import tensorflow as tf
import logging
from tqdm import tqdm
from collections import deque
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, Dense, Flatten, BatchNormalization
def build_q_net(input_shape, action_space_size, learning_rate, game_id):
print('Initialize Q net with action space size {0} and state shape {1}'.format(
action_space_size, input_shape))
model = Sequential()
if game_id in ('Berzerk-v0', 'Skiing-v0', 'Assault-v0', 'Breakout-v0'):
model.add(Conv2D(16, (4, 4), strides=(2, 2), padding='same',
kernel_initializer='he_uniform',
activation='relu', input_shape=input_shape))
model.add(Conv2D(32, (4, 4), strides=(2, 2),
padding='same', kernel_initializer='he_uniform',
activation='relu'))
model.add(Conv2D(64, (4, 4), strides=(2, 2),
padding='same', kernel_initializer='he_uniform',
activation='relu'))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(512, activation='relu',
kernel_initializer='he_uniform'))
model.add(Dense(256, activation='relu',
kernel_initializer='he_uniform'))
model.add(Dense(128, activation='relu',
kernel_initializer='he_uniform'))
elif game_id in ('CartPole-v1', 'Breakout-ram-v0'):
model.add(Dense(512, input_shape=input_shape,
activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(256, activation='relu',
kernel_initializer='he_uniform'))
model.add(Dense(128, activation='relu',
kernel_initializer='he_uniform'))
else:
raise RuntimeError('Model for {0} not found'.format(game_id))
model.add(Dense(action_space_size, activation='linear'))
model.compile(optimizer=tf.optimizers.Adam(
learning_rate=learning_rate), loss='mse')
return model
def sanitize_state(state, game_id):
if game_id in ('Berzerk-v0', 'Breakout-v0'):
sanitized_state = np.array(state)
sanitized_state = sanitized_state[0:-28, :, :]
sanitized_state = np.expand_dims(
np.dot(sanitized_state, [0.2989, 0.5870, 0.1140]), axis=2)
sanitized_state = sanitized_state / 255.0
return sanitized_state
else:
return state
def sample_action(q_vals, epsilon, action_space, collect=False):
if np.random.random() < epsilon and collect:
return np.random.randint(0, action_space)
else:
return np.argmax(q_vals)
def eval(q_net, max_eps, env, action_space_size, max_steps, render, game_id, state_buffer_size):
total_reward = 0.0
total_normalized_reward = 0.0
state_buffer = deque(maxlen=state_buffer_size)
for _ in range(max_eps):
done = False
state = env.reset()
sanitized_initial_state = sanitize_state(state, game_id)
for _ in range(state_buffer_size):
state_buffer.append(sanitized_initial_state)
step = 0
while not done and step < max_steps:
if render:
env.render()
frames = np.swapaxes(list(state_buffer), 0, 3)
q_vals = q_net(tf.convert_to_tensor(
frames, dtype=tf.float32))
action = sample_action(
q_vals, 0.0, action_space_size, collect=False)
state, raw_reward, done, _ = env.step(action)
state_buffer.append(sanitize_state(state, game_id))
reward = normalize_reward(raw_reward, game_id)
total_reward += raw_reward
total_normalized_reward += reward
step += 1
avg_reward = total_reward / max_eps
avg_normalized_reward = total_normalized_reward / max_eps
return avg_reward, avg_normalized_reward
def normalize_reward(reward, game_id):
if game_id == 'Berzerk-v0':
if reward > 0:
return 1.0
elif reward < 0:
return -5.0
else:
return -0.001
elif game_id == 'CartPole-v1':
return reward if reward > 0 else -1.0
elif game_id == 'Skiing-v0':
if reward > 0:
return 1.0
elif reward < 0:
return -0.1
else:
return 0.0
elif game_id == 'Breakout-v0':
if reward > 0:
return 1.0
elif reward < 0:
return -1.0
else:
return 0.0
elif game_id in ('Breakout-ram-v0', 'Breakout-v0'):
if reward > 0:
return 1.0
elif reward < 0:
return -5.0
else:
return 0.0
else:
return reward
def explore(game_id='Berzerk-v0'):
logging.info('Exploring {0}'.format(game_id))
env = gym.make(game_id)
print('Envrionment action space {0}, state space {1}'.format(
env.action_space, env.observation_space))
env.reset()
rewards = []
for _ in range(5):
done = False
while not done:
env.render()
_, reward, done, _ = env.step(env.action_space.sample())
rewards.append(reward)
rewards_np = np.array(rewards)
print('Reward history size {0}, mean {1}, range {2} to {3}'.format(
len(rewards), rewards_np.mean(), rewards_np.min(), rewards_np.max()))
env.close()
def train(game_id='Berzerk-v0',
max_eps=10000,
max_buffer_size=5000,
batch_size=512,
gamma=0.95,
max_eval_eps=3,
update_freq=3,
eval_freq=10,
epsilon=0.8,
epsilon_decay=0.99,
max_steps=2000,
render=True,
checkpoint_location='most_recent',
model_location='most_recent',
save_checkpoint=True,
learning_rate=1e-3,
train_per_episode=1,
rounds_per_episode=1,
state_buffer_size=4,
mode='train'):
logging.info('Training {0}'.format(game_id))
env = gym.make(game_id)
eval_env = gym.make(game_id)
use_epsilon = epsilon
if render:
env.render()
if mode == 'train':
eval_env.render()
best_avg_reward = 0.0
action_space_size = 0
action_space_size = env.action_space.n
state_buffer = deque(maxlen=state_buffer_size)
sanitized_initial_state = sanitize_state(env.reset(), game_id)
logging.info('Sanitized state shape is {0}'.format(sanitized_initial_state.shape))
for _ in range(state_buffer_size):
state_buffer.append(sanitized_initial_state)
state_shape = np.swapaxes(np.array(list(state_buffer)), 0, 3).shape[1:]
q_net = build_q_net(state_shape, action_space_size, learning_rate, game_id)
target_q_net = build_q_net(
state_shape, action_space_size, learning_rate, game_id)
checkpoint = tf.train.Checkpoint(step=tf.Variable(1), net=q_net)
checkpoint_manager = tf.train.CheckpointManager(
checkpoint, os.path.join('./checkpoint', checkpoint_location), max_to_keep=10)
try:
checkpoint.restore(checkpoint_manager.latest_checkpoint)
target_q_net.set_weights(q_net.get_weights())
except:
q_net = build_q_net(state_shape, action_space_size,
learning_rate, game_id)
target_q_net = build_q_net(
state_shape, action_space_size, learning_rate, game_id)
q_net.summary()
replay_buffer = deque(maxlen=max_buffer_size)
for eps in range(max_eps):
# Sampling
if mode in ('train', 'eval'):
for _ in range(rounds_per_episode):
done = False
state = env.reset()
for _ in range(state_buffer_size):
state_buffer.append(sanitized_initial_state)
step = 0
while not done and step < max_steps:
if render:
env.render()
frames = np.swapaxes(list(state_buffer), 0, 3)
q_vals = q_net(tf.convert_to_tensor(
frames, dtype=tf.float32))
action = sample_action(
q_vals, use_epsilon, action_space_size, collect=(mode == 'train'))
next_state, raw_reward, done, _ = env.step(action)
reward = normalize_reward(raw_reward, game_id)
state_buffer.append(sanitize_state(next_state, game_id))
updated_frames = np.swapaxes(list(state_buffer), 0, 3)
next_q_vals = target_q_net(tf.convert_to_tensor(updated_frames, dtype=tf.float32)).numpy()[0]
update_q_val = q_vals.numpy()[0]
update_q_val[action] = reward
if not done:
update_q_val[action] += gamma * next_q_vals.max()
replay_buffer.append(
(frames[0], updated_frames[0], reward, done, action, update_q_val))
state = next_state
step += 1
# Training
if mode == 'train':
for _ in range(train_per_episode):
# Prepare batch for training
sampled_batch = []
if len(replay_buffer) < batch_size:
sampled_batch = list(replay_buffer)
else:
sampled_batch = random.sample(replay_buffer, batch_size)
# actual_batch_size = len(sampled_batch)
logging.info('Preparing batch...')
states, next_states, rewards, actions, terminals, update_q_vals = [], [], [], [], [], []
for state, next_state, reward, done, action, update_q_val in tqdm(sampled_batch):
states.append(state)
next_states.append(next_state)
rewards.append([reward])
terminals.append([0.0 if done else 1.0])
actions.append(action)
update_q_vals.append(update_q_val)
"""
logging.info('Calculating loss...')
target_q_vals = q_net(tf.convert_to_tensor(
states, dtype=tf.float32))
next_q_vals = target_q_net(tf.convert_to_tensor(
next_states, dtype=tf.float32))
max_next_q_vals = tf.expand_dims(
tf.reduce_max(next_q_vals, axis=1), axis=1)
action_onehot = tf.one_hot(actions, action_space_size)
action_onehot_reverse = tf.ones_like(
action_onehot) - action_onehot
exclude_update_q_vals = target_q_vals * action_onehot_reverse
update_q_vals_reward = action_onehot * \
tf.convert_to_tensor(rewards)
update_q_vals_discounted_max_q = action_onehot * \
max_next_q_vals * gamma * tf.convert_to_tensor(terminals)
target_q_vals = exclude_update_q_vals + \
update_q_vals_discounted_max_q + update_q_vals_reward
logging.info('Start training...')
q_net.fit(x=tf.convert_to_tensor(states, dtype=tf.float32),
y=tf.convert_to_tensor(
target_q_vals, dtype=tf.float32),
batch_size=16, verbose=1)
"""
q_net.fit(x=tf.convert_to_tensor(states, dtype=tf.float32),
y=tf.convert_to_tensor(
update_q_vals, dtype=tf.float32),
batch_size=16, verbose=1)
if save_checkpoint:
checkpoint_manager.save()
if eps % update_freq == 0 and eps != 0 and mode == 'train':
target_q_net.set_weights(q_net.get_weights())
if eps % eval_freq == 0 and eps != 0 and mode == 'train':
avg_reward, avg_normalized_reward = eval(q_net, max_eval_eps, eval_env,
action_space_size, max_steps, render, game_id, state_buffer_size)
if avg_reward > best_avg_reward:
best_avg_reward = avg_reward
tf.saved_model.save(q_net, os.path.join(
'./best_model', model_location))
print('Current eval average reward is {0} and normalized reward is {1}'.format(
avg_reward, avg_normalized_reward))
use_epsilon *= epsilon_decay
print('Finished episode {0}/{1} and the next epsilon is {2}'.format(eps, max_eps, use_epsilon))
env.close()
eval_env.close()