-
Notifications
You must be signed in to change notification settings - Fork 1
/
dreamer.py
250 lines (227 loc) · 9.48 KB
/
dreamer.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
import numpy as np
import torch
import torch.distributions as td
import torch.nn as nn
import torch.optim as optim
import wandb
from models.agent import AgentModel
from models.rssm import get_feat, get_dist, apply_states
from utils import FreezeParameters, denormalize_images, merge_images_in_chunks
class Dreamer:
def __init__(
self,
agent: AgentModel,
model_lr=6e-4,
action_lr=8e-5,
value_lr=8e-5,
discount=0.99,
discount_lambda=0.95,
horizon=15,
free_nats=3,
kl_beta=1.0,
device: str = "cuda",
):
self.agent = agent.to(device)
self.discount = discount
self.discount_lambda = discount_lambda
self.horizon = horizon
self.free_nats = free_nats
self.kl_beta = kl_beta
self.device = device
self.model_modules = nn.ModuleList(
[
self.agent.observation_encoder,
self.agent.observation_decoder,
self.agent.rssm,
self.agent.reward_model,
]
)
self.model_optimizer = optim.Adam(
self.model_modules.parameters(),
lr=model_lr,
)
self.action_optimizer = optim.Adam(
self.agent.action_decoder.parameters(),
lr=action_lr,
)
self.value_optimizer = optim.Adam(
self.agent.value_model.parameters(),
lr=value_lr,
)
self.training_steps = 0
def update(
self, observations: torch.Tensor, actions: torch.Tensor, rewards: torch.Tensor
):
# swap batch and seq_len
observations = torch.transpose(observations, 0, 1)
actions = torch.transpose(actions, 0, 1)
rewards = torch.transpose(rewards, 0, 1).unsqueeze(-1)
self.model_optimizer.zero_grad()
self.action_optimizer.zero_grad()
self.value_optimizer.zero_grad()
model_loss, actor_loss, value_loss = self.calculate_loss(
observations, actions, rewards
)
model_loss.backward()
nn.utils.clip_grad_norm_(self.model_modules.parameters(), 100.0)
actor_loss.backward()
nn.utils.clip_grad_norm_(self.agent.action_decoder.parameters(), 100.0)
value_loss.backward()
nn.utils.clip_grad_norm_(self.agent.value_model.parameters(), 100.0)
self.model_optimizer.step()
self.action_optimizer.step()
self.value_optimizer.step()
self.training_steps += 1
def calculate_loss(self, observations, actions, rewards):
"""
Note: With observation[0], the agent took actions[0] and get rewards[0].
Assume they are float32 and on the proper device.
:param observations: (seq_len, batch_size, img_shape), normalized to [-0.5, 0.5]
:param actions: (seq_len, batch_size, act_shape)
:param rewards: (seq_len, batch_size)
:return:
"""
assert self.agent.explore is True
seq_len, batch_size = observations.shape[:2]
obs_embed = self.agent.observation_encoder(observations)
# init prev state
prev_state = self.agent.rssm.create_initial_state(
batch_size, device=self.device
)
# get prior and posterior and initialize stuff
prior, posterior = self.agent.rssm.observe(
seq_len, obs_embed, actions, prev_state
) # (seq_len, batch_size, state_dim)
prior_dist, posterior_dist = get_dist(prior), get_dist(posterior)
features = get_feat(posterior)
image_pred = self.agent.observation_decoder(features)
image_loss = -torch.mean(image_pred.log_prob(observations))
reward_pred = self.agent.reward_model(features)
reward_loss = -torch.mean(reward_pred.log_prob(rewards))
# TODO: also add pcont
kl_div = torch.maximum(
torch.mean(td.kl_divergence(posterior_dist, prior_dist)),
torch.tensor(self.free_nats, dtype=torch.float32, device=self.device),
) # to prevent penalize small KL divergence
model_loss = image_loss + reward_loss + self.kl_beta * kl_div
# produce a gradient-free posterior for action network
with torch.no_grad():
# TODO: handle pcont
flat_post = apply_states(
posterior, lambda x: x.reshape(seq_len * batch_size, -1)
)
with FreezeParameters(self.model_modules):
imagine_states, _ = self.agent.rssm.imagine(
self.horizon, self.agent.policy, flat_post
) # image_states of shape (horizon, seq_len * b, state_dim)
imagine_features = get_feat(imagine_states)
# get rewards from imagine features
with FreezeParameters(self.model_modules + [self.agent.value_model]):
imagine_reward_dists = self.agent.reward_model(imagine_features).mean
imagine_value_pred = self.agent.value_model(imagine_features).mean
# TODO: handle pcont
# compute value estimate
discount_arr = self.discount * torch.ones_like(imagine_reward_dists)
value_estimates = self.compute_value_estimate(
imagine_reward_dists[:-1],
imagine_value_pred[:-1],
discount_arr[:-1],
bootstrap=imagine_value_pred[-1],
lambda_=self.discount_lambda,
)
discount_arr = torch.cat([torch.ones_like(discount_arr[:1]), discount_arr])
discount = torch.cumprod(discount_arr[:-2], 0)
actor_loss = -torch.mean(discount * value_estimates)
with torch.no_grad():
value_features = imagine_features[:-1].detach()
value_discount = discount.detach()
value_target = value_estimates.detach()
value_pred = self.agent.value_model(value_features)
log_prob = value_pred.log_prob(value_target)
value_loss = -torch.mean(value_discount * log_prob.unsqueeze(2))
# logging
with torch.no_grad():
wandb.log(
{
"training_steps": self.training_steps,
"train/model_loss": image_loss.item(),
"train/reward_loss": reward_loss.item(),
"train/kl_div": kl_div.item(),
"train/actor_loss": actor_loss.item(),
"train/value_loss": value_loss.item(),
},
commit=False,
)
# log images
if self.training_steps % 500 == 0:
i = torch.randint(0, batch_size, (1,)).item()
# reconstruction quality
ground_truths = np.transpose(
denormalize_images(observations[:, i].detach().cpu().numpy()),
(0, 2, 3, 1),
)
pred_images = np.transpose(
denormalize_images(image_pred.mean[:, i].detach().cpu().numpy()),
(0, 2, 3, 1),
)
reconstruction_demo = merge_images_in_chunks(ground_truths, pred_images)
# prediction + reconstruction
# feed 5 obs to the model, and predict the next 45 obs
given_obs = observations[:5, i]
obs_embed = self.agent.observation_encoder(given_obs)
prev_state = self.agent.rssm.create_initial_state(1, device=self.device)
_, posterior = self.agent.rssm.observe(
5,
obs_embed.reshape(5, 1, -1),
actions[:5, i].reshape(5, 1, -1),
prev_state,
)
posterior = apply_states(posterior, lambda x: x[-1])
states = self.agent.rssm.follow(
45, actions[5:, i].reshape(45, 1, -1), posterior
)
features = get_feat(states)
pred_images = self.agent.observation_decoder(features)
pred_images = np.transpose(
np.squeeze(
denormalize_images(pred_images.mean.detach().cpu().numpy()),
axis=1,
),
(0, 2, 3, 1),
)
# first row is observed
prediction_demo = merge_images_in_chunks(
ground_truths[5:], pred_images, chunk_size=5
)
wandb.log(
{
"train/reconstruction": wandb.Image(reconstruction_demo),
"train/prediction": wandb.Image(prediction_demo),
}
)
return model_loss, actor_loss, value_loss
def compute_value_estimate(
self,
reward: torch.Tensor,
value: torch.Tensor,
discount: torch.Tensor,
bootstrap: torch.Tensor,
lambda_: float,
):
"""
Compute the discounted reward for a batch of data.
reward, value, and discount are all shape [horizon - 1, batch, 1] (last element is cut off)
Bootstrap is [batch, 1]
"""
next_values = torch.cat([value[1:], bootstrap[None]], 0)
target = reward + discount * next_values * (1 - lambda_)
timesteps = list(range(reward.shape[0] - 1, -1, -1))
outputs = []
accumulated_reward = bootstrap
for t in timesteps:
inp = target[t]
discount_factor = discount[t]
accumulated_reward = inp + discount_factor * lambda_ * accumulated_reward
outputs.append(accumulated_reward)
returns = torch.flip(torch.stack(outputs), [0])
return returns