-
Notifications
You must be signed in to change notification settings - Fork 2
/
rainbow.py
540 lines (462 loc) · 20.2 KB
/
rainbow.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
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
import gym
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from typing import Deque, Dict, List, Tuple
from replay_buffer import *
from utils import *
from IPython.display import clear_output
from torch.nn.utils import clip_grad_norm_
from preprocess_frame import *
import copy
from frame_stack import *
import os
class DQNAgent:
"""DQN Agent interacting with environment.
Attribute:
env (gym.Env): openAI Gym environment
memory (PrioritizedReplayBuffer): replay memory to store transitions
batch_size (int): batch size for sampling
target_update (int): period for target model's hard update
gamma (float): discount factor
dqn (Network): model to train and select actions
dqn_target (Network): target model to update
optimizer (torch.optim): optimizer for training dqn
transition (list): transition information including
state, action, reward, next_state, done
v_min (float): min value of support
v_max (float): max value of support
atom_size (int): the unit number of support
support (torch.Tensor): support for categorical dqn
n_step (int): step number to calculate n-step td error
hidden_size (int): size of the layers of the top dense network
lr (float): learning rate
plot (bool): if True, plot training charts
frame_interval (int): steps per plotting refresh
alpha (float): determines how much prioritization is used
beta (float): determines how much importance sampling is used
prior_eps (float): guarantees every transition can be sampled
max_epsilon (float): starting epsilon value
min_epsilon (float): finish epsilon value
epsilon_decay (float): epsilon decay for each step
"""
def __init__(
self,
env: gym.Env,
memory_size: int = 1024,
batch_size: int = 32,
target_update: int = 100,
gamma: float = 0.99,
lr: float =0.001,
hidden_size=128,
# PER parameters
alpha: float = 0.2,
beta: float = 0.6,
prior_eps: float = 1e-6,
# Categorical DQN parameters
v_min: float = 0.0,
v_max: float = 200.0,
atom_size: int = 51,
# N-step Learning
n_step: int = 3,
# Plotting
plot: bool = True,
frame_interval: int = 100,
# Options
no_dueling: bool = False,
no_double: bool = False,
no_noise: bool = False,
no_categorical: bool = False,
no_priority: bool = False,
no_n_step: bool = False,
# Only if no_noise is True
max_epsilon: float = 1.,
min_epsilon: float = 0.1,
epsilon_decay: float = 0.0005,
# Reward clipping
max_reward: float = None,
min_reward: float = None,
# Input preprocessing functions
frame_preprocess: np.array = None, # this is a function
# Early stopping
early_stopping: bool = True,
# Frames_stacking
n_frames_stack: int = 1,
# training delay
training_delay: int = 0, # how many frames to skip before start training (used to fill the memory buffer)
# used for saving and loading
model_path: str = "models",
model_name: str = "rainbow"
):
obs_shape = env.observation_space.shape # get shape of an observation
if frame_preprocess is not None:
obs_shape = frame_preprocess(np.zeros(obs_shape)).shape # get shape of preprocessed observation
if n_frames_stack > 1: # if the input consists of more than 1 frame, compute its total dimension
obs_shape = list(obs_shape)
obs_shape[0] *= n_frames_stack
assert len(obs_shape) == 3 or len(obs_shape) == 1
if len(obs_shape) == 1: # observation is an array
print("Using DenseNet")
self.obs_dim = [obs_shape[0]]
self.mode = "dense"
self.frame_stack = FrameStack(n_frames_stack, mode="array")
else:
print("Using ConvNet") # observation is a frame
# remember: gym has dimension (w, h, c) but pytorch has (c, h, w)
self.obs_dim = [obs_shape[0], obs_shape[1], obs_shape[2]]
self.mode = "conv"
self.frame_stack = FrameStack(n_frames_stack, mode="pixels")
self.action_dim = env.action_space.n # get number of possible actions
self.n_frames_stack = n_frames_stack # number of stacked input observations
self.env = env # the the gym environment
self.batch_size = batch_size
self.target_update = target_update
self.gamma = gamma
self.hidden_size = hidden_size # this parameters is used only with DenseNet
# device: cpu / gpu
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu"
)
print("Device", self.device)
# PER and memory for N-step Learning (PER = Prioritized Experience Replay)
self.beta = beta
self.prior_eps = prior_eps
self.memory = PrioritizedReplayBuffer(
self.obs_dim, memory_size, batch_size, alpha=alpha, n_step=n_step, gamma=gamma
)
self.n_step = n_step
# Categorical DQN parameters
self.v_min = v_min
self.v_max = v_max
self.atom_size = atom_size
self.support = torch.linspace(
self.v_min, self.v_max, self.atom_size
).to(self.device)
# networks: dqn, dqn_target
if self.mode == "dense":
# if input is 1d array use dense layers
self.dqn = DenseNet(
self.obs_dim[0], self.action_dim, self.atom_size, self.support, self.hidden_size, no_dueling, no_noise
).to(self.device)
self.dqn_target = DenseNet(
self.obs_dim[0], self.action_dim, self.atom_size, self.support, self.hidden_size, no_dueling, no_noise
).to(self.device)
else:
# if input is 3d frame use convolutional layers
self.dqn = ConvNet(
self.obs_dim, self.action_dim, self.atom_size, self.support, no_dueling, no_noise
).to(self.device)
self.dqn_target = ConvNet(
self.obs_dim, self.action_dim, self.atom_size, self.support, no_dueling, no_noise
).to(self.device)
self.dqn_target.load_state_dict(self.dqn.state_dict()) # DQN <- targetDQN
self.dqn_target.eval()
# early stropping
self.early_stopping = early_stopping
# optimizer
self.lr = lr
self.optimizer = optim.Adam(self.dqn.parameters(), lr=self.lr)
# current transition to store in memory
self.transition = list()
# training delay
self.training_delay = training_delay
# mode: train / test
self.is_test = False
# reward clipping
self.max_reward = max_reward
self.min_reward = min_reward
# save / load
self.model_dir = model_path
self.model_path = os.path.join(model_path, model_name + ".tar")
# observation preprocess function (convert to grayscale, crop, resize...)
self.frame_preprocess = frame_preprocess
# plot
self.plot = plot
self.frame_interval = frame_interval
# epsilon (used only if noisy net is disabled)
self.max_epsilon = max_epsilon
self.min_epsilon = min_epsilon
self.epsilon = self.max_epsilon
self.epsilon_decay = epsilon_decay
# options to disable some features
self.no_dueling = no_dueling # no dueling
self.no_double = no_double # no double
if no_double:
self.dqn_target = self.dqn
self.no_noise = no_noise # no noise
if no_noise:
self.epsilon, self.max_epsilon, self.min_epsilon = 0, 0, 0
self.no_categorical = no_categorical # no categorical
self.no_n_step = no_n_step # no n_step
if no_n_step:
self.n_step = 1
self.memory = PrioritizedReplayBuffer(
self.obs_dim, memory_size, batch_size, alpha=alpha, n_step=self.n_step, gamma=gamma
)
self.no_priority = no_priority # no priority memory
if no_priority:
self.alpha = 0
self.memory = PrioritizedReplayBuffer(
self.obs_dim, memory_size, batch_size, alpha=alpha, n_step=self.n_step, gamma=gamma
)
def select_action(self, state: np.ndarray) -> np.ndarray:
"""Select an action from the input state and return state and action."""
# epsilon greedy policy
if self.no_noise and self.epsilon > np.random.random():
# Select a random action
selected_action = self.env.action_space.sample()
else:
# Select best action: no epsilon greedy action selection but NoisyNet
selected_action = self.dqn(
torch.FloatTensor(state).unsqueeze(0).to(self.device)
).argmax()
selected_action = selected_action.detach().cpu().numpy()
if not self.is_test:
self.transition = [state, selected_action]
return selected_action
def step(self, action: np.ndarray) -> Tuple[np.ndarray, np.float64, bool]:
"""Take an action and return the response of the env (next state, rewards, done)."""
next_state, reward, done, _ = self.env.step(action)
if self.frame_preprocess is not None:
next_state = self.frame_preprocess(next_state)
if self.n_frames_stack > 1:
next_state = self.get_n_frames(next_state)
if self.max_reward is not None:
if reward > self.max_reward:
reward = self.max_reward
if self.min_reward is not None:
if reward < self.min_reward:
reward = self.min_reward
if not self.is_test:
self.transition += [reward, next_state, done]
self.memory.store(*self.transition) # store a full transition
return next_state, reward, done
def update_model(self) -> torch.Tensor:
"""Update the model by gradient descent."""
# sample transitions
samples = self.memory.sample_batch(self.beta)
# PER needs beta to calculate weights
weights = torch.FloatTensor(samples["weights"].reshape(-1, 1)).to(self.device)
indices = samples["indices"]
# N-step Learning loss
gamma = self.gamma ** self.n_step
elementwise_loss = self._compute_dqn_loss(samples, gamma)
# PER: importance sampling before average
loss = torch.mean(elementwise_loss * weights)
self.optimizer.zero_grad()
loss.backward()
clip_grad_norm_(self.dqn.parameters(), 10.0)
self.optimizer.step()
# PER: update priorities
loss_for_prior = elementwise_loss.detach().cpu().numpy()
new_priorities = loss_for_prior + self.prior_eps
self.memory.update_priorities(indices, new_priorities)
# NoisyNet: reset noise
if not self.no_noise:
self.dqn.reset_noise()
self.dqn_target.reset_noise()
return loss.item()
def train(self, num_frames: int) -> (List[int], List[int]):
"""Train the agent."""
self.is_test = False
state = self.env.reset()
# get the first state
state = self.init_first_frame(state)
update_cnt = 0 # counts the number of steps between each update
losses = [] # loss for each training step
scores = [] # score for each episode
frame_scores = [] # average score each frame_interval frames
score = 0 # current score
if self.early_stopping:
best_model = copy.deepcopy(self.dqn.state_dict())
best_average_score = -np.inf
for frame_idx in range(1, num_frames + 1):
action = self.select_action(state)
next_state, reward, done = self.step(action)
state = next_state
score += reward
self.update_beta()
# if episode ends
if done:
scores.append(score)
state = self.env.reset()
state = self.init_first_frame(state)
score = 0
# linearly decrease epsilon
if self.no_noise:
self.set_epsilon()
# if training is ready
if len(self.memory) >= self.batch_size and self.training_delay <= 0:
loss = self.update_model()
losses.append(loss)
update_cnt += 1
# if hard update is needed
if update_cnt % self.target_update == 0 and not self.no_double:
self._target_hard_update()
if frame_idx % self.frame_interval == 0:
if len(scores) == 0:
if len(frame_scores) > 0:
# if no episodes have been completed in the current interval
# then take the last score
frame_scores.append(float(frame_scores[-1]))
else:
# if no episodes have been completed since the beginning of the game
frame_scores.append(0.)
else:
frame_scores.append(float(np.mean(scores)))
if self.plot:
self._plot(frame_idx, frame_scores, losses)
scores = []
# early stopping
if self.early_stopping and frame_scores[-1] > best_average_score:
best_average_score = frame_scores[-1]
best_model = copy.deepcopy(self.dqn.state_dict())
# save temporary model
self.save()
if self.training_delay > 0:
self.training_delay -= 1
if self.early_stopping:
self.dqn.load_state_dict(best_model)
self.env.close()
return frame_scores, losses
def test(self, get_frames=False, get_actions=False) -> (int, List[int]) or (int, List[np.ndarray]):
"""Test the agent on one episode."""
self.is_test = True
state = self.env.reset()
state = self.init_first_frame(state)
done = False
score = 0
actions = []
frames = []
while not done:
self.env.render()
action = self.select_action(state)
if get_actions:
actions.append(action)
next_state, reward, done = self.step(action)
state = next_state
score += reward
if get_frames:
frames.append(self.env.render(mode='rgb_array'))
self.env.close()
if get_frames and not get_actions:
return score, frames
if not get_frames and get_actions:
return score, actions
if get_frames and get_actions:
return score, frames, actions
return score
def _compute_dqn_loss(self, samples: Dict[str, np.ndarray], gamma: float) -> torch.Tensor:
"""Return the loss."""
device = self.device # for shortening the following lines
state = torch.FloatTensor(samples["obs"]).to(device)
next_state = torch.FloatTensor(samples["next_obs"]).to(device)
action = torch.LongTensor(samples["acts"]).to(device)
reward = torch.FloatTensor(samples["rews"].reshape(-1, 1)).to(device)
done = torch.FloatTensor(samples["done"].reshape(-1, 1)).to(device)
if not self.no_categorical:
# # Compute categorical distribution loss
delta_z = float(self.v_max - self.v_min) / (self.atom_size - 1)
with torch.no_grad():
next_action = self.dqn(next_state).argmax(1)
next_dist = self.dqn_target.dist(next_state)
next_dist = next_dist[range(self.batch_size), next_action]
t_z = reward + (1 - done) * gamma * self.support
t_z = t_z.clamp(min=self.v_min, max=self.v_max)
b = (t_z - self.v_min) / delta_z
l = b.floor().long()
u = b.ceil().long()
offset = (
torch.linspace(
0, (self.batch_size - 1) * self.atom_size, self.batch_size
).long().unsqueeze(1).expand(self.batch_size, self.atom_size).to(self.device)
)
proj_dist = torch.zeros(next_dist.size(), device=self.device)
proj_dist.view(-1).index_add_(
0, (l + offset).view(-1), (next_dist * (u.float() - b)).view(-1)
)
proj_dist.view(-1).index_add_(
0, (u + offset).view(-1), (next_dist * (b - l.float())).view(-1)
)
dist = self.dqn.dist(state)
log_p = torch.log(dist[range(self.batch_size), action])
loss = -(proj_dist * log_p).sum(1)
else:
# Compute normal value estimation loss
# G_t = r + gamma * v(s_{t+1}) if state != Terminal
# = r otherwise
action = torch.LongTensor(samples["acts"].reshape(-1, 1)).to(device)
curr_q_value = self.dqn(state).gather(1, action)
next_q_value = self.dqn_target(next_state).gather(
1, self.dqn(next_state).argmax(dim=1, keepdim=True)
)[0].detach()
mask = 1 - done
target = (reward + self.gamma * next_q_value * mask).to(self.device)
# calculate dqn loss
loss = F.smooth_l1_loss(curr_q_value, target, reduction="none")
return loss
def _target_hard_update(self):
"""Hard update: target <- local."""
self.dqn_target.load_state_dict(self.dqn.state_dict())
def init_first_frame(self, state):
"""Preprocess the first frame and fill the empty frame stack repeating it"""
if self.frame_preprocess is not None:
state = self.frame_preprocess(state)
if self.n_frames_stack > 1:
self.frame_stack.clear()
state = self.get_n_frames(state)
return state
def set_epsilon(self):
"""Set the value of epsilon"""
self.epsilon = max(
self.min_epsilon, self.epsilon - (
self.max_epsilon - self.min_epsilon
) * self.epsilon_decay
)
def update_beta(self):
"""Update beta parameter for PER"""
self.epsilon = max(
self.min_epsilon, self.epsilon - (
self.max_epsilon - self.min_epsilon
) * self.epsilon_decay
)
def get_n_frames(self, frame: np.ndarray) -> np.ndarray:
"""Return the last n frames"""
if self.frame_stack.full():
self.frame_stack.stack(frame, 1)
return self.frame_stack.frames
else:
self.frame_stack.stack(frame, self.n_frames_stack)
return self.frame_stack.frames
def save(self):
if not os.path.exists(self.model_dir):
os.makedirs(self.model_dir)
print("Saving model...")
torch.save({
'model': self.dqn.state_dict(),
}, os.path.join(self.model_path))
print("Model saved in: " + str(self.model_path))
def load(self):
print("Restoring saved model...")
checkpoint = torch.load(self.model_path)
self.dqn.load_state_dict(checkpoint['model'])
print("Model restored from: " + str(self.model_path))
def _plot(self, frame_idx: int, scores: List[float], losses: List[float]):
"""Plot the training progresses."""
clear_output(True)
plt.figure(figsize=(20, 5))
plt.subplot(131)
plt.title('Frame %s. Mean Score: %.4s' % (frame_idx,
np.mean(scores[-10:])))
plt.plot(scores)
plt.xlabel("Frames x " + str(self.frame_interval))
plt.ylabel("Score")
plt.subplot(132)
plt.title('Loss')
plt.plot(losses)
plt.xlabel("Frames")
plt.ylabel("Loss")
plt.show()