-
Notifications
You must be signed in to change notification settings - Fork 0
/
deepQAgent.py
301 lines (253 loc) · 13 KB
/
deepQAgent.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
300
301
import os
import datetime
import argparse
import random
import numpy as np
import cv2
import torch
from collections import deque
from game.doodlejump import DoodleJump
from model.networks import Deep_QNet, Deep_RQNet, DQ_Resnet18, DQ_Mobilenet, DQ_Mnasnet
from model.deepQTrainer import QTrainer
from helper import write_model_params
from torch.utils.tensorboard import SummaryWriter
class Agent:
def __init__(self, args):
self.n_games = 0
# self.epsilon = 0
self.ctr = 1
seed = args.seed
self.exploration = args.exploration
os.environ['PYTHONHASHSEED'] = str(seed)
# Torch RNG
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# Python RNG
np.random.seed(seed)
random.seed(seed)
self.store_frames = args.store_frames
self.image_h = args.height
self.image_w = args.width
self.image_c = args.channels
self.memory = deque(maxlen=args.max_memory)
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.gamma = args.gamma
self.batch_size = args.batch_size
self.lr = args.learning_rate
self.steps = 0
self.exploration_type = args.explore
self.decay_factor = args.decay_factor
self.epsilon = args.epsilon
self.eulers_constant = 2.71828
if args.explore == "epsilon_g_decay_exp":
self.epsilon = 1
if args.model=="dqn":
self.model = Deep_QNet()
elif args.model=="drqn":
self.model = Deep_RQNet()
elif args.model=='resnet':
self.model = DQ_Resnet18()
elif args.model=='mobilenet':
self.model = DQ_Mobilenet()
elif args.model=='mnasnet':
self.model = DQ_Mnasnet()
if args.model_path or args.test:
self.model.load_state_dict(torch.load(args.model_path))
self.trainer = QTrainer(model=self.model, lr=self.lr, gamma=self.gamma, device=self.device,
num_channels=self.image_c, attack_eps=args.attack_eps)
def preprocess(self, state):
# resize the image and then rotate
img = cv2.resize(state, (self.image_w, self.image_h))
M = cv2.getRotationMatrix2D((self.image_w / 2, self.image_h / 2), 270, 1.0)
img = cv2.warpAffine(img, M, (self.image_h, self.image_w))
if self.store_frames:
os.makedirs("./image_dump", exist_ok=True)
cv2.imwrite("./image_dump/"+str(self.ctr)+".jpg", img)
self.ctr+=1
imagenet_mean = [0.485, 0.456, 0.406]
imagenet_std = [0.229, 0.224, 0.225]
if self.image_c == 1:
# convert the image to grayscale
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# normalize the image with imagenet mean and std values
img = ((img/255.0) - np.mean(imagenet_mean))/np.mean(imagenet_std)
else:
# normalize the image with imagenet mean and std values
img = ((img/255.0) - imagenet_mean)/imagenet_std
# change the shape from WxHxC to CxHxW for pytorch tensor
img = img.transpose((2, 0, 1))
# Add a axis for converting image to shape: 1xCxHxW
img = np.expand_dims(img, axis=0)
return img
def get_state(self, game):
state = game.getCurrentFrame()
return self.preprocess(state)
def remember(self, state, action, reward, next_state, done):
self.memory.append((state, action, reward, next_state, done)) # popleft if MAX_MEMORY is reached
def train_long_memory(self):
if len(self.memory) > self.batch_size:
mini_sample = random.sample(self.memory, self.batch_size) # list of tuples
else:
mini_sample = self.memory
states, actions, rewards, next_states, dones = zip(*mini_sample)
self.model.train()
return self.trainer.train_step(states, actions, rewards, next_states, dones)
def train_short_memory(self, state, action, reward, next_state, done):
self.model.train()
return self.trainer.train_step(state, action, reward, next_state, done)
def should_explore(self, test_mode):
self.steps += 1
r = random.random()
if test_mode:
return False
if self.exploration_type == "epsilon_g":
pass
elif self.exploration_type == "epsilon_g_decay_exp":
self.epsilon = self.epsilon * pow((1.0 - self.decay_factor), self.steps)
elif self.exploration_type == "epsilon_g_decay_exp_cur":
self.epsilon = self.decay_factor * pow(self.eulers_constant, -self.steps)
if r > self.epsilon:
return True
return False
def get_action(self, state, test_mode=False):
final_move = [0,0,0]
if self.should_explore(test_mode):
move = random.randint(0, 2)
final_move[move] = 1
else:
state0 = torch.tensor(state, dtype=torch.float).to(self.device)
self.model.eval()
with torch.no_grad():
prediction = self.model(state0)
move = torch.argmax(prediction).item()
final_move[move] = 1
return final_move
def train(game, args, writer):
if args.macos:
os.environ['KMP_DUPLICATE_LIB_OK']='True'
sum_rewards = 0
sum_short_loss = 0
total_score = 0
record = 0
loop_ctr = 0
agent = Agent(args)
dummy_input = torch.rand(1, args.channels, args.height, args.width).to(agent.device)
writer.add_graph(agent.model, dummy_input)
print("Now playing")
while agent.n_games != args.max_games:
loop_ctr += 1
# get old state
state_old = agent.get_state(game)
# get move
final_move = agent.get_action(state_old)
# perform move and get new state
reward, done, score = game.playStep(final_move)
state_new = agent.get_state(game)
sum_rewards += reward
writer.add_scalar('Reward/curr_reward', reward, loop_ctr)
# train short memory
short_loss = agent.train_short_memory(state_old, final_move, reward, state_new, [done])
writer.add_scalar('Game/Short_Episodes', loop_ctr, loop_ctr)
sum_short_loss += short_loss
# remember
agent.remember(state_old, final_move, reward, state_new, done)
if loop_ctr%25 == 0:
writer.add_scalar('Loss/Short_train', sum_short_loss/loop_ctr, loop_ctr)
writer.add_scalar('Reward/mean_reward', sum_rewards/loop_ctr, loop_ctr)
if done:
# train long memory, plot result
game.gameReboot()
agent.n_games += 1
long_loss = agent.train_long_memory()
writer.add_scalar('Loss/Long_train', long_loss, agent.n_games)
writer.add_scalar('Game/Episodes', agent.n_games, agent.n_games)
if score > record:
record = score
# save the best model yet
agent.model.save(file_name="model_best.pth", model_folder_path="./model"+hyper_params+dstr)
if agent.n_games%100 == 0:
# save model per 100 games
agent.model.save(file_name="model_"+str(agent.n_games)+".pth", model_folder_path="./model"+hyper_params+dstr)
print('Game', agent.n_games, 'Score', score, 'Record:', record)
writer.add_scalar('Score/High_Score', record, agent.n_games)
total_score += score
mean_score = total_score / agent.n_games
writer.add_scalars('Score', {'Curr_Score':score, 'Mean_Score': mean_score}, agent.n_games)
write_model_params(agent.model, agent.n_games, writer)
writer.add_hparams(hparam_dict=vars(args),
metric_dict={'long_loss_loss': long_loss,
'mean_short_loss': sum_short_loss/loop_ctr,
'mean_reward': sum_rewards/loop_ctr,
'high_score': record,
'mean_score': mean_score
})
def test(game, args):
if args.macos:
os.environ['KMP_DUPLICATE_LIB_OK']='True'
record = 0
cum_score = 0
agent = Agent(args)
print("Now playing")
f = open("test_logs.txt", "w")
f.write("Now playing")
f.close()
while agent.n_games < args.max_games:
if args.attack:
state = agent.get_state(game) # original
adv_manip = agent.trainer.create_adv_state(state) #manipulated
final_move = agent.get_action(torch.tensor(state).to(agent.device) + adv_manip, test_mode=True)
reward, done, score = game.playStep(final_move)
else:
state_old = agent.get_state(game)
final_move = agent.get_action(state_old, test_mode=True)
reward, done, score = game.playStep(final_move)
if done:
agent.n_games += 1
cum_score += score
game.gameReboot()
if score > record:
record = score
f = open("test_logs.txt", "a")
f.write('Game: '+str(agent.n_games)+' Score: '+str(score)+' Record: '+str(record)+' Mean Score: '+str(cum_score/agent.n_games)+'\n')
f.close()
print('Game', agent.n_games, 'Score', score, 'Record:', record, 'Mean Score:', cum_score/agent.n_games)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='RL Agent for Doodle Jump')
parser.add_argument("--macos", action="store_true", help="select model to train the agent")
parser.add_argument("--human", action="store_true", help="playing the game manually without agent")
parser.add_argument("--test", action="store_true", help="playing the game with a trained agent")
parser.add_argument("-d", "--difficulty", type=str, default="EASY", choices=["EASY", "MEDIUM", "HARD"], help="select difficulty of the game")
parser.add_argument("-m", "--model", type=str, default="dqn", choices=["dqn", "drqn", "resnet", "mobilenet", "mnasnet"], help="select model to train the agent")
parser.add_argument("-p", "--model_path", type=str, help="path to weights of an earlier trained model")
parser.add_argument("-lr", "--learning_rate", type=float, default=0.001, help="set learning rate for training the model")
parser.add_argument("-g", "--gamma", type=float, default=0.9, help="set discount factor for q learning")
parser.add_argument("--max_memory", type=int, default=10000, help="Buffer memory size for long training")
parser.add_argument("--store_frames", action="store_true", help="store frames encountered during game play by agent")
parser.add_argument("--batch_size", type=int, default=1000, help="Batch size for long training")
parser.add_argument("--reward_type", type=int, default=1, choices=[1, 2, 3, 4, 5, 6], help="types of rewards formulation")
parser.add_argument("--exploration", type=int, default=40, help="number of games to explore")
parser.add_argument("--channels", type=int, default=1, help="set the image channels for preprocessing")
parser.add_argument("--height", type=int, default=80, help="set the image height post resize")
parser.add_argument("--width", type=int, default=80, help="set the image width post resize")
parser.add_argument("--server", action="store_true", help="when training on server add this flag")
parser.add_argument("--seed", type=int, default=42, help="change seed value for creating game randomness")
parser.add_argument("--max_games", type=int, default=1000, help="set the max number of games to be played by the agent")
parser.add_argument("--explore", type=str, default="epsilon_g", choices=["epsilon_g","epsilon_g_decay_exp","epsilon_g_decay_exp_cur"], help="select the exploration vs exploitation tradeoff")
parser.add_argument("--decay_factor", type=float, default=0.9, help="set the decay factor for exploration")
parser.add_argument("--epsilon", type=float, default=0.8, help="set the epsilon value for exploration")
parser.add_argument("--attack", action="store_true", help="use fast fgsm attack to manipulate the input state")
parser.add_argument("--attack_eps", type=float, default=0.3, help="epsilon value for the fgsm attack")
args = parser.parse_args()
game = DoodleJump(difficulty=args.difficulty, server=args.server, reward_type=args.reward_type)
if args.human:
game.run()
elif args.test:
test(game, args)
else:
hyper_params = "_d_"+args.difficulty+"_m_"+args.model+"_lr_"+str(args.learning_rate)+"_g_"+str(args.gamma)+"_mem_"+str(args.max_memory)+"_batch_"+str(args.batch_size)
arg_dict = vars(args)
dstr = datetime.datetime.now().strftime("_dt-%Y-%m-%d-%H-%M-%S")
writer = SummaryWriter(log_dir="model"+hyper_params+dstr)
writer.add_text('Model Parameters: ', str(arg_dict), 0)
train(game, args, writer)