-
Notifications
You must be signed in to change notification settings - Fork 162
/
common.py
193 lines (174 loc) · 6.76 KB
/
common.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
import sys
import time
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable
HYPERPARAMS = {
'pong': {
'env_name': "PongNoFrameskip-v4",
'stop_reward': 18.0,
'run_name': 'pong',
'replay_size': 100000,
'replay_initial': 10000,
'target_net_sync': 1000,
'epsilon_frames': 10**5,
'epsilon_start': 1.0,
'epsilon_final': 0.02,
'learning_rate': 0.0001,
'gamma': 0.99,
'batch_size': 32
},
'breakout-small': {
'env_name': "BreakoutNoFrameskip-v4",
'stop_reward': 500.0,
'run_name': 'breakout-small',
'replay_size': 3*10 ** 5,
'replay_initial': 20000,
'target_net_sync': 1000,
'epsilon_frames': 10 ** 6,
'epsilon_start': 1.0,
'epsilon_final': 0.1,
'learning_rate': 0.0001,
'gamma': 0.99,
'batch_size': 64
},
'breakout': {
'env_name': "BreakoutNoFrameskip-v4",
'stop_reward': 500.0,
'run_name': 'breakout',
'replay_size': 10 ** 6,
'replay_initial': 50000,
'target_net_sync': 10000,
'epsilon_frames': 10 ** 6,
'epsilon_start': 1.0,
'epsilon_final': 0.1,
'learning_rate': 0.00025,
'gamma': 0.99,
'batch_size': 32
},
'invaders': {
'env_name': "SpaceInvadersNoFrameskip-v4",
'stop_reward': 500.0,
'run_name': 'breakout',
'replay_size': 10 ** 6,
'replay_initial': 50000,
'target_net_sync': 10000,
'epsilon_frames': 10 ** 6,
'epsilon_start': 1.0,
'epsilon_final': 0.1,
'learning_rate': 0.00025,
'gamma': 0.99,
'batch_size': 32
},
}
def unpack_batch(batch):
states, actions, rewards, dones, last_states = [], [], [], [], []
for exp in batch:
state = np.array(exp.state, copy=False)
states.append(state)
actions.append(exp.action)
rewards.append(exp.reward)
dones.append(exp.last_state is None)
if exp.last_state is None:
last_states.append(state) # the result will be masked anyway
else:
last_states.append(np.array(exp.last_state, copy=False))
return np.array(states, copy=False), np.array(actions), np.array(rewards, dtype=np.float32), \
np.array(dones, dtype=np.uint8), np.array(last_states, copy=False)
def calc_loss_dqn(batch, net, tgt_net, gamma, cuda=False):
states, actions, rewards, dones, next_states = unpack_batch(batch)
states_v = Variable(torch.from_numpy(states))
next_states_v = Variable(torch.from_numpy(next_states), volatile=True)
actions_v = Variable(torch.from_numpy(actions))
rewards_v = Variable(torch.from_numpy(rewards))
done_mask = torch.ByteTensor(dones)
if cuda:
states_v = states_v.cuda()
next_states_v = next_states_v.cuda()
actions_v = actions_v.cuda()
rewards_v = rewards_v.cuda()
done_mask = done_mask.cuda()
state_action_values = net(states_v).gather(1, actions_v.unsqueeze(-1)).squeeze(-1)
next_state_values = tgt_net(next_states_v).max(1)[0]
next_state_values[done_mask] = 0.0
next_state_values.volatile = False
expected_state_action_values = next_state_values * gamma + rewards_v
return nn.MSELoss()(state_action_values, expected_state_action_values)
class RewardTracker:
def __init__(self, writer, stop_reward):
self.writer = writer
self.stop_reward = stop_reward
def __enter__(self):
self.ts = time.time()
self.ts_frame = 0
self.total_rewards = []
return self
def __exit__(self, *args):
self.writer.close()
def reward(self, reward, frame, epsilon=None):
self.total_rewards.append(reward)
speed = (frame - self.ts_frame) / (time.time() - self.ts)
self.ts_frame = frame
self.ts = time.time()
mean_reward = np.mean(self.total_rewards[-100:])
epsilon_str = "" if epsilon is None else ", eps %.2f" % epsilon
print("%d: done %d games, mean reward %.3f, speed %.2f f/s%s" % (
frame, len(self.total_rewards), mean_reward, speed, epsilon_str
))
sys.stdout.flush()
if epsilon is not None:
self.writer.add_scalar("epsilon", epsilon, frame)
self.writer.add_scalar("speed", speed, frame)
self.writer.add_scalar("reward_100", mean_reward, frame)
self.writer.add_scalar("reward", reward, frame)
if mean_reward > self.stop_reward:
print("Solved in %d frames!" % frame)
return True
return False
class EpsilonTracker:
def __init__(self, epsilon_greedy_selector, params):
self.epsilon_greedy_selector = epsilon_greedy_selector
self.epsilon_start = params['epsilon_start']
self.epsilon_final = params['epsilon_final']
self.epsilon_frames = params['epsilon_frames']
self.frame(0)
def frame(self, frame):
self.epsilon_greedy_selector.epsilon = \
max(self.epsilon_final, self.epsilon_start - frame / self.epsilon_frames)
def distr_projection(next_distr, rewards, dones, Vmin, Vmax, n_atoms, gamma):
"""
Perform distribution projection aka Catergorical Algorithm from the
"A Distributional Perspective on RL" paper
"""
batch_size = len(rewards)
proj_distr = np.zeros((batch_size, n_atoms), dtype=np.float32)
delta_z = (Vmax - Vmin) / (n_atoms - 1)
for atom in range(n_atoms):
tz_j = np.minimum(Vmax, np.maximum(Vmin, rewards + (Vmin + atom * delta_z) * gamma))
b_j = (tz_j - Vmin) / delta_z
l = np.floor(b_j).astype(np.int64)
u = np.ceil(b_j).astype(np.int64)
eq_mask = u == l
proj_distr[eq_mask, l[eq_mask]] += next_distr[eq_mask, atom]
ne_mask = u != l
proj_distr[ne_mask, l[ne_mask]] += next_distr[ne_mask, atom] * (u - b_j)[ne_mask]
proj_distr[ne_mask, u[ne_mask]] += next_distr[ne_mask, atom] * (b_j - l)[ne_mask]
if dones.any():
proj_distr[dones] = 0.0
tz_j = np.minimum(Vmax, np.maximum(Vmin, rewards[dones]))
b_j = (tz_j - Vmin) / delta_z
l = np.floor(b_j).astype(np.int64)
u = np.ceil(b_j).astype(np.int64)
eq_mask = u == l
eq_dones = dones.copy()
eq_dones[dones] = eq_mask
if eq_dones.any():
proj_distr[eq_dones, l] = 1.0
ne_mask = u != l
ne_dones = dones.copy()
ne_dones[dones] = ne_mask
if ne_dones.any():
proj_distr[ne_dones, l] = (u - b_j)[ne_mask]
proj_distr[ne_dones, u] = (b_j - l)[ne_mask]
return proj_distr