This repository has been archived by the owner on Oct 31, 2023. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 88
/
drqv2.py
262 lines (204 loc) · 8.83 KB
/
drqv2.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
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import hydra
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import utils
class RandomShiftsAug(nn.Module):
def __init__(self, pad):
super().__init__()
self.pad = pad
def forward(self, x):
n, c, h, w = x.size()
assert h == w
padding = tuple([self.pad] * 4)
x = F.pad(x, padding, 'replicate')
eps = 1.0 / (h + 2 * self.pad)
arange = torch.linspace(-1.0 + eps,
1.0 - eps,
h + 2 * self.pad,
device=x.device,
dtype=x.dtype)[:h]
arange = arange.unsqueeze(0).repeat(h, 1).unsqueeze(2)
base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
base_grid = base_grid.unsqueeze(0).repeat(n, 1, 1, 1)
shift = torch.randint(0,
2 * self.pad + 1,
size=(n, 1, 1, 2),
device=x.device,
dtype=x.dtype)
shift *= 2.0 / (h + 2 * self.pad)
grid = base_grid + shift
return F.grid_sample(x,
grid,
padding_mode='zeros',
align_corners=False)
class Encoder(nn.Module):
def __init__(self, obs_shape):
super().__init__()
assert len(obs_shape) == 3
self.repr_dim = 32 * 35 * 35
self.convnet = nn.Sequential(nn.Conv2d(obs_shape[0], 32, 3, stride=2),
nn.ReLU(), nn.Conv2d(32, 32, 3, stride=1),
nn.ReLU(), nn.Conv2d(32, 32, 3, stride=1),
nn.ReLU(), nn.Conv2d(32, 32, 3, stride=1),
nn.ReLU())
self.apply(utils.weight_init)
def forward(self, obs):
obs = obs / 255.0 - 0.5
h = self.convnet(obs)
h = h.view(h.shape[0], -1)
return h
class Actor(nn.Module):
def __init__(self, repr_dim, action_shape, feature_dim, hidden_dim):
super().__init__()
self.trunk = nn.Sequential(nn.Linear(repr_dim, feature_dim),
nn.LayerNorm(feature_dim), nn.Tanh())
self.policy = nn.Sequential(nn.Linear(feature_dim, hidden_dim),
nn.ReLU(inplace=True),
nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(inplace=True),
nn.Linear(hidden_dim, action_shape[0]))
self.apply(utils.weight_init)
def forward(self, obs, std):
h = self.trunk(obs)
mu = self.policy(h)
mu = torch.tanh(mu)
std = torch.ones_like(mu) * std
dist = utils.TruncatedNormal(mu, std)
return dist
class Critic(nn.Module):
def __init__(self, repr_dim, action_shape, feature_dim, hidden_dim):
super().__init__()
self.trunk = nn.Sequential(nn.Linear(repr_dim, feature_dim),
nn.LayerNorm(feature_dim), nn.Tanh())
self.Q1 = nn.Sequential(
nn.Linear(feature_dim + action_shape[0], hidden_dim),
nn.ReLU(inplace=True), nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(inplace=True), nn.Linear(hidden_dim, 1))
self.Q2 = nn.Sequential(
nn.Linear(feature_dim + action_shape[0], hidden_dim),
nn.ReLU(inplace=True), nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(inplace=True), nn.Linear(hidden_dim, 1))
self.apply(utils.weight_init)
def forward(self, obs, action):
h = self.trunk(obs)
h_action = torch.cat([h, action], dim=-1)
q1 = self.Q1(h_action)
q2 = self.Q2(h_action)
return q1, q2
class DrQV2Agent:
def __init__(self, obs_shape, action_shape, device, lr, feature_dim,
hidden_dim, critic_target_tau, num_expl_steps,
update_every_steps, stddev_schedule, stddev_clip, use_tb):
self.device = device
self.critic_target_tau = critic_target_tau
self.update_every_steps = update_every_steps
self.use_tb = use_tb
self.num_expl_steps = num_expl_steps
self.stddev_schedule = stddev_schedule
self.stddev_clip = stddev_clip
# models
self.encoder = Encoder(obs_shape).to(device)
self.actor = Actor(self.encoder.repr_dim, action_shape, feature_dim,
hidden_dim).to(device)
self.critic = Critic(self.encoder.repr_dim, action_shape, feature_dim,
hidden_dim).to(device)
self.critic_target = Critic(self.encoder.repr_dim, action_shape,
feature_dim, hidden_dim).to(device)
self.critic_target.load_state_dict(self.critic.state_dict())
# optimizers
self.encoder_opt = torch.optim.Adam(self.encoder.parameters(), lr=lr)
self.actor_opt = torch.optim.Adam(self.actor.parameters(), lr=lr)
self.critic_opt = torch.optim.Adam(self.critic.parameters(), lr=lr)
# data augmentation
self.aug = RandomShiftsAug(pad=4)
self.train()
self.critic_target.train()
def train(self, training=True):
self.training = training
self.encoder.train(training)
self.actor.train(training)
self.critic.train(training)
def act(self, obs, step, eval_mode):
obs = torch.as_tensor(obs, device=self.device)
obs = self.encoder(obs.unsqueeze(0))
stddev = utils.schedule(self.stddev_schedule, step)
dist = self.actor(obs, stddev)
if eval_mode:
action = dist.mean
else:
action = dist.sample(clip=None)
if step < self.num_expl_steps:
action.uniform_(-1.0, 1.0)
return action.cpu().numpy()[0]
def update_critic(self, obs, action, reward, discount, next_obs, step):
metrics = dict()
with torch.no_grad():
stddev = utils.schedule(self.stddev_schedule, step)
dist = self.actor(next_obs, stddev)
next_action = dist.sample(clip=self.stddev_clip)
target_Q1, target_Q2 = self.critic_target(next_obs, next_action)
target_V = torch.min(target_Q1, target_Q2)
target_Q = reward + (discount * target_V)
Q1, Q2 = self.critic(obs, action)
critic_loss = F.mse_loss(Q1, target_Q) + F.mse_loss(Q2, target_Q)
if self.use_tb:
metrics['critic_target_q'] = target_Q.mean().item()
metrics['critic_q1'] = Q1.mean().item()
metrics['critic_q2'] = Q2.mean().item()
metrics['critic_loss'] = critic_loss.item()
# optimize encoder and critic
self.encoder_opt.zero_grad(set_to_none=True)
self.critic_opt.zero_grad(set_to_none=True)
critic_loss.backward()
self.critic_opt.step()
self.encoder_opt.step()
return metrics
def update_actor(self, obs, step):
metrics = dict()
stddev = utils.schedule(self.stddev_schedule, step)
dist = self.actor(obs, stddev)
action = dist.sample(clip=self.stddev_clip)
log_prob = dist.log_prob(action).sum(-1, keepdim=True)
Q1, Q2 = self.critic(obs, action)
Q = torch.min(Q1, Q2)
actor_loss = -Q.mean()
# optimize actor
self.actor_opt.zero_grad(set_to_none=True)
actor_loss.backward()
self.actor_opt.step()
if self.use_tb:
metrics['actor_loss'] = actor_loss.item()
metrics['actor_logprob'] = log_prob.mean().item()
metrics['actor_ent'] = dist.entropy().sum(dim=-1).mean().item()
return metrics
def update(self, replay_iter, step):
metrics = dict()
if step % self.update_every_steps != 0:
return metrics
batch = next(replay_iter)
obs, action, reward, discount, next_obs = utils.to_torch(
batch, self.device)
# augment
obs = self.aug(obs.float())
next_obs = self.aug(next_obs.float())
# encode
obs = self.encoder(obs)
with torch.no_grad():
next_obs = self.encoder(next_obs)
if self.use_tb:
metrics['batch_reward'] = reward.mean().item()
# update critic
metrics.update(
self.update_critic(obs, action, reward, discount, next_obs, step))
# update actor
metrics.update(self.update_actor(obs.detach(), step))
# update critic target
utils.soft_update_params(self.critic, self.critic_target,
self.critic_target_tau)
return metrics