-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathagents.py
445 lines (337 loc) · 19.2 KB
/
agents.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributions as D
import numpy as np
from dense_models import *
from conv_models import *
from world_model import *
import utils
def get_mode(dist, n_samples = 100):
sample = dist.sample_n(n_samples)
logprob = dist.log_prob(sample)
mode_indices = torch.argmax(logprob, dim=0)
return sample[mode_indices]
def stack_states(rssm_states: list, dim=0):
return dict(
mean=torch.stack([state['mean'] for state in rssm_states], dim=dim),
std=torch.stack([state['std'] for state in rssm_states], dim=dim),
stoch=torch.stack([state['stoch'] for state in rssm_states], dim=dim),
deter=torch.stack([state['deter'] for state in rssm_states], dim=dim),
)
def flatten_state(rssm_state: dict):
return dict(
mean=torch.reshape(rssm_state['mean'], [-1, rssm_state['mean'].shape[-1]]),
std=torch.reshape(rssm_state['std'], [-1, rssm_state['std'].shape[-1]]),
stoch=torch.reshape(rssm_state['stoch'], [-1, rssm_state['stoch'].shape[-1]]),
deter=torch.reshape(rssm_state['deter'], [-1, rssm_state['deter'].shape[-1]]),
)
def detach_state(rssm_state: dict):
return dict(
mean=rssm_state['mean'].detach(),
std=rssm_state['std'].detach(),
stoch=rssm_state['stoch'].detach(),
deter=rssm_state['deter'].detach(),
)
def expand_state(rssm_state: dict, n : int):
return dict(
mean=rssm_state['mean'].expand(n, *rssm_state['mean'].shape),
std=rssm_state['std'].expand(n, *rssm_state['std'].shape),
stoch=rssm_state['stoch'].expand(n, *rssm_state['stoch'].shape),
deter=rssm_state['deter'].expand(n, *rssm_state['deter'].shape),
)
def get_dist(rssm_state: dict):
return D.independent.Independent(D.Normal(rssm_state['mean'], rssm_state['std']), 1)
class Agent(nn.Module):
def __init__(self,
config=None,
world_lr=6e-4,
policy_lr=8e-5,
value_lr=8e-5,
device='cuda' if torch.cuda.is_available() else 'cpu',
):
super().__init__()
self.config = config
self.action_dist = config['action_dist']
self.use_rewards = config['use_rewards']
self.contrastive = config['contrastive']
self._action_size = config['action_size']
self.wm = WorldModel(config)
self.world_optim = torch.optim.Adam(utils.get_parameters([self.wm]), lr=world_lr)
self.policy = ActionModel(self._action_size, 230, 200, 3, dist=self.action_dist)
self.policy_optim = torch.optim.Adam(self.policy.parameters(), lr=policy_lr)
self.reinforce = self.action_dist == 'one_hot'
self.value_model = DenseModel(self.wm._feature_size, (1,), 3, 200)
if self.reinforce or not self.use_rewards:
self.value_target = DenseModel(self.wm._feature_size, (1,), 3, 200)
else:
self.value_target = self.value_model
self.value_optim = torch.optim.Adam(self.value_model.parameters(), lr=value_lr)
self.grad_clip = 100.
self.gamma = config['discount_gamma']
self.device = device
self.add_actor_entropy = config.get('actor_entropy', False)
self.entropy_temperature = config.get('entropy_temperature', 1e-4)
# Default for the moment
self.use_rms = False
self.rew_rms = utils.RunningMeanStd()
self.ambiguity_rms = utils.RunningMeanStd()
self.ambiguity_beta = 1e-3
##
self.to(device)
def update_target_network(self, tau, network, target_network):
# Softly Update Target
target_value_params = target_network.named_parameters()
value_params = network.named_parameters()
target_value_state_dict = dict(target_value_params)
value_state_dict = dict(value_params)
for name in value_state_dict:
value_state_dict[name] = tau*value_state_dict[name].clone() + \
(1-tau)*target_value_state_dict[name].clone()
target_network.load_state_dict(value_state_dict)
def train_world(self, path_obs, path_act, path_rew, preferred_obs, path_done=None, update_target=False, get_reconstruction=False):
loss_dict = dict()
batch_t, batch_b, img_shape = path_obs.shape[0], path_obs.shape[1], path_obs.shape[2:]
batch_t -= 1
init_obs = path_obs[0]
next_obs = path_obs[1:]
obs_embed = self.wm.obs_encoder(path_obs)
init_embed = obs_embed[0]
next_embed = obs_embed[1:]
prev_state = self.wm.prior.initial_state(batch_size=batch_b, device=self.device)
prior, post = self.rollout_posterior(batch_t, next_embed, path_act, prev_state)
feat = torch.cat((post['stoch'], post['deter']), dim=-1)
if self.use_rewards:
reward_pred = self.wm.rew_model(feat)
reward_loss = -torch.mean(reward_pred.log_prob(path_rew))
else:
reward_loss = torch.zeros(1).to(self.device)
if self.contrastive:
W_c = post['stoch'].reshape(batch_t * batch_b, -1) # N
reshaped_obs = path_obs[1:].reshape(batch_b*batch_t, *img_shape)
W_c = self.wm.z_encoder(feat).reshape(batch_t * batch_b, -1)
mean_z = self.wm.w_contrastive(next_embed).reshape(batch_t * batch_b, -1)
sim_matrix = torch.mm(W_c, mean_z.T)
labels = torch.Tensor(list(range(batch_b * batch_t))).long().to(self.device)
image_loss = F.cross_entropy(sim_matrix, labels, reduction='mean')
else:
image_pred = self.wm.obs_decoder(feat)
image_loss = -torch.mean(image_pred.log_prob(next_obs))
prior_dist = get_dist(prior)
post_dist = get_dist(post)
div = torch.mean(torch.distributions.kl.kl_divergence(post_dist, prior_dist))
kl_loss = torch.max(div, div.new_full(div.size(), self.wm.free_nats))
model_loss = image_loss + reward_loss + self.wm.kl_scale * kl_loss
self.world_optim.zero_grad()
model_loss.backward()
grad_norm_world = torch.nn.utils.clip_grad_norm_(utils.get_parameters([self.wm]), self.grad_clip)
loss_dict['world_grad_norm'] = grad_norm_world
self.world_optim.step()
loss_dict = dict(reconstruction_loss=image_loss.item(), kl_loss=kl_loss, reward_loss=reward_loss.item(), **loss_dict)
if get_reconstruction and (not self.contrastive):
with torch.no_grad():
vb = 16 # video batch size
true_steps = 5
ground_truth = (path_obs[1:, :vb] + 0.5).cpu()
recon_truth = next_obs.cpu()[:, :vb] + 0.5
recon = image_pred.mean.detach()[:, :vb]
init = {k: v[true_steps-1, :vb] for k,v in post.items()}
rec_prior = self.rollout_prior(batch_t - true_steps, path_act[true_steps:, :vb], init)
rec_feat = torch.cat((rec_prior['stoch'], rec_prior['deter']), dim=-1)
future_pred = self.wm.obs_decoder(rec_feat).mean
model = torch.clamp(torch.cat([recon[:true_steps], future_pred], dim=0) + 0.5, 0, 1).cpu()
model_post = torch.clamp(recon + 0.5, 0, 1).cpu()
error = (model - recon_truth + 1) / 2
post_prior_div = (model_post - model + 1) / 2
reconstruction_dict = dict(truth=ground_truth, rencostructed_truth=recon_truth, prior_predictions=model, post_predictions=model_post, prior_truth_diff=error, post_prior_diff=post_prior_div)
else:
reconstruction_dict = dict()
return post, loss_dict, reconstruction_dict
def train_value(self, state_features, lambda_returns, discount_arr):
value_pred = self.value_model(state_features.detach())
value_pred = D.independent.Independent(D.Normal(value_pred.mean[:-1], 1), 1)
value_loss = -torch.mean(discount_arr * value_pred.log_prob(lambda_returns.detach()) )
self.value_optim.zero_grad()
value_loss.backward()
grad_norm_value = torch.nn.utils.clip_grad_norm_(utils.get_parameters([self.value_model]), self.grad_clip)
self.value_optim.step()
return value_loss.item(), grad_norm_value
def train_policy_value(self, steps, policy, states, preferred_obs, obs_batch=None):
with utils.FreezeParameters([self.wm, self.value_model]):
states = detach_state(flatten_state(states))
list_prior_states, act_logprobs, actions, act_entropies = self.rollout_policy(steps, policy, states)
prior_states = stack_states(list_prior_states, dim=0)
all_prior_feat = torch.cat((prior_states['stoch'], prior_states['deter']), dim=-1)
if self.use_rewards:
future_rew_pred = self.wm.rew_model(all_prior_feat)
free_energy, fe_dict = self.compute_free_energy(future_rew_pred.mean, preferred_obs, prior_states, actions)
else:
free_energy, fe_dict = self.compute_free_energy(None, preferred_obs, prior_states, actions, obs_batch=obs_batch)
loss_dict = dict(**fe_dict)
future_value_pred = self.value_target(all_prior_feat).mean
if not self.use_rewards:
if self.use_rms:
self.rew_rms.update(free_energy.detach().cpu().view(-1, 1).numpy())
free_energy = free_energy / np.sqrt(self.rew_rms.var.item() + 1e-8)
loss_dict['free_energy'] = free_energy.detach().mean().item()
loss_dict['action_entropy'] = act_entropies.mean().detach().cpu().item()
discount = torch.ones_like(future_value_pred) * self.gamma
expected_free_energy = utils.lambda_dreamer_values(free_energy[:,:,None], future_value_pred, gamma=discount)
loss_dict['expected_free_energy'] = expected_free_energy.detach().cpu().mean().item()
discount_arr = torch.cat( [discount[:1] / self.gamma, discount[:-1]], dim=0).detach()
discount_arr = torch.cumprod(discount_arr, dim=0).squeeze(-1)
if self.reinforce:
future_value_pred = self.value_model(all_prior_feat).mean
advantages = (expected_free_energy - future_value_pred).detach().squeeze(-1)
loss = torch.mean(discount_arr[:-1] * advantages[:-1] * act_logprobs)
else:
loss = torch.mean(discount_arr[:-1] * expected_free_energy[:-1].squeeze(-1))
if self.add_actor_entropy:
loss = loss - torch.mean(discount_arr[:-1] * act_entropies * self.entropy_temperature)
self.policy_optim.zero_grad()
loss.backward()
grad_norm_actor = torch.nn.utils.clip_grad_norm_(utils.get_parameters([self.policy]), self.grad_clip)
loss_dict['policy_grad_norm'] = grad_norm_actor
self.policy_optim.step()
with utils.FreezeParameters([self.wm, self.policy]):
value_loss_item, value_grad_norm = self.train_value(all_prior_feat, expected_free_energy[:-1].detach(), discount_arr[:-1])
loss_dict['value_logprob_loss'] = value_loss_item
loss_dict['value_grad_norm'] = value_grad_norm
return loss_dict
def compute_free_energy(self, rew, preferred_obs, prior_states=None, actions=None, obs_batch=None):
free_energy_dict = dict()
free_energy = torch.zeros(1,1)
if len(prior_states['stoch'].shape) == 2:
prior_states = expand_state(prior_states,1)
batch_t, batch_b, state_dim = prior_states['stoch'].shape[0], prior_states['stoch'].shape[1], prior_states['stoch'].shape[2]
if self.use_rewards:
free_energy = -rew.squeeze(-1)
preferences = free_energy
else:
# Contrastive AIF
if self.contrastive:
feat = torch.cat((prior_states['stoch'], prior_states['deter']), dim=-1)
pref_embed = self.wm.obs_encoder(preferred_obs).view(1, self.wm.obs_encoder.embed_size)
W_c = self.wm.z_encoder(feat).reshape(batch_t * batch_b, -1)
pref_z = self.wm.w_contrastive(pref_embed).reshape(1, 200)
pos_loss = torch.mm(W_c, pref_z.T).view(batch_t, batch_b) / W_c.shape[-1]
free_energy = -pos_loss
free_energy_dict['-pos_loss'] = pos_loss.detach().mean().item()
# Only to allow computation along the episode
if obs_batch is not None:
next_embed = self.wm.obs_encoder(obs_batch).view(-1, self.wm.obs_encoder.embed_size)
mean_z = self.wm.w_contrastive(next_embed).reshape(-1, 200)
mean_z = torch.cat([pref_z, mean_z], dim=0)
sim_matrix = torch.mm(W_c, mean_z.T) / W_c.shape[-1]
neg_loss = torch.logsumexp(sim_matrix, dim=1).view(batch_t, batch_b) - np.log(mean_z.shape[0])
free_energy = -pos_loss + neg_loss
free_energy_dict['neg_loss'] = neg_loss.detach().mean().item()
# Likelihood AIF
else:
feat = torch.cat((prior_states['stoch'], prior_states['deter']), dim=-1)
image_pred = self.wm.obs_decoder(feat)
predicted_obs = image_pred.mean
preferred_obs_dist = D.Independent(D.Laplace(preferred_obs, 1.), 3)
logprob_preferences = preferred_obs_dist.log_prob(predicted_obs) / np.prod(preferred_obs.shape)
free_energy = - logprob_preferences
free_energy_dict['-logprob_preferences'] = (-logprob_preferences).detach().mean().item()
init_states = flatten_state(prior_states)
_, posterior_states = self.wm.posterior(obs_embed=self.wm.obs_encoder(preferred_obs).expand(batch_b*batch_t, self.wm.obs_encoder.embed_size), prev_action=None, prev_state=init_states, is_init=True)
prior_dist = get_dist(init_states)
post_dist = get_dist(posterior_states)
epistemic_term = D.kl_divergence(post_dist, prior_dist).reshape(*logprob_preferences.shape) / init_states['stoch'].shape[-1]
self.ambiguity_rms.update(epistemic_term.detach().cpu().view(-1, 1).numpy())
epistemic_term = epistemic_term / np.sqrt(self.ambiguity_rms.var.item() + 1e-8)
free_energy = - logprob_preferences - self.ambiguity_beta * epistemic_term
free_energy_dict['-epistemic_term'] = (-epistemic_term).detach().mean().item()
return free_energy, free_energy_dict
def rollout_policy(self, steps, policy, prev_state):
priors = [prev_state]
act_logprobs = []
actions = []
act_entropies = []
state = prev_state
for t in range(steps):
# Act
feat = torch.cat((state['stoch'], state['deter']), dim=-1)
feat = feat.detach()
act_dist = policy(feat)
if self.reinforce:
act = act_dist.sample()
else:
act = act_dist.rsample()
actions.append(act)
act_entropies.append(act_dist.entropy())
act_logprobs.append(act_dist.log_prob(act))
# Imagine
state = self.wm.prior(actions[t], state)
priors.append(state)
actions = torch.stack(actions, dim=0) # shape (T, B, action_dim)
act_logprobs = torch.stack(act_logprobs, dim=0) # shape (T, B, action_dim)
act_entropies = torch.stack(act_entropies, dim=0) # shape (T, B, action_dim)
return priors, act_logprobs, actions, act_entropies
def rollout_policies(self, steps, policies, prev_state, take_mean_action=False):
priors = []
actions = []
state = prev_state
for t in range(steps):
# Act
feat = torch.cat((state['stoch'], state['deter']), dim=-1)
act = torch.stack([ p(f).sample() for f, p in zip(feat, policies)], dim=0)
actions.append(act)
# Imagine
state = self.wm.prior(actions[t], state)
priors.append(state)
all_prior_states = stack_states(priors, dim=0)
actions = torch.stack(actions, dim=1) # shape: (B,T,action_dim)
return all_prior_states, actions
def step(self, image_obs, rew, act, prev_state):
with torch.no_grad():
image_embed = self.wm.obs_encoder(image_obs)
if prev_state is None:
prev_state = self.wm.posterior.initial_state(batch_size=1, device=self.device)
_, post = self.rollout_posterior(1, image_embed, act, prev_state)
return flatten_state(post)
def eval_obs(self, image_obs, rew, preferred_obs, prev_state):
with torch.no_grad():
if self.use_rewards:
preferences = self.compute_free_energy(rew, preferred_obs, prior_states=prev_state)[0]
else:
preferences = self.compute_free_energy(None, preferred_obs, prior_states=prev_state)[0]
return preferences
def policy_distribution(self, steps, policies, preferred_obs, prev_state=None, eval_mode=False):
with torch.no_grad():
n_policies = len(policies)
if prev_state is None:
prev_state = self.wm.posterior.initial_state(batch_size=n_policies, device=self.device)
prev_state = flatten_state(prev_state)
all_prior_states, actions = self.rollout_policies(steps, policies, prev_state, take_mean_action=eval_mode)
all_prior_feat = torch.cat((all_prior_states['stoch'], all_prior_states['deter']), dim=-1)
if self.use_rewards:
future_rew_pred = self.wm.rew_model(all_prior_feat)
preference_loss = torch.mean(self.compute_free_energy(future_rew_pred.mean, preferred_obs, prior_states=all_prior_states)[0], dim=0)
else:
preference_loss = torch.mean(self.compute_free_energy(None, preferred_obs, prior_states=all_prior_states)[0], dim=0)
free_energy = preference_loss
policy_logits = F.softmax(-free_energy.detach(), dim=0)
policy_distr = D.Categorical(policy_logits)
expected_loss = torch.sum(free_energy.detach() * policy_logits)
return policy_distr, actions.detach(), dict(policy_expected_loss=expected_loss.detach().cpu().item())
def rollout_posterior(self, steps: int, obs_embed: torch.Tensor, action: torch.Tensor, prev_state: dict):
priors = []
posteriors = []
for t in range(steps):
prior_state, posterior_state = self.wm.posterior(obs_embed[t], action[t], prev_state, transition_model=self.wm.prior)
priors.append(prior_state)
posteriors.append(posterior_state)
prev_state = posterior_state
prior = stack_states(priors, dim=0)
post = stack_states(posteriors, dim=0)
return prior, post
def rollout_prior(self, steps: int, action: torch.Tensor, prev_state: dict):
priors = []
state = prev_state
for t in range(steps):
state = self.wm.prior(action[t], state)
priors.append(state)
return stack_states(priors, dim=0)