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 | class OfflineIQLearn(Method):
__version__ = "1.0.0"
__author__ = "Garg et. al."
__method_name__ = "Inverse Q-Learn"
def __init__(
self,
environment: Env,
enjoy_criteria: int = 100,
verbose: bool = False,
config_file: str = None,
) -> None:
self.enjoy_criteria = enjoy_criteria
self.verbose = verbose
try:
self.environment_name = environment.spec.name
except AttributeError:
self.environment_name = environment.spec._env_name
self.save_path = f"./tmp/iqlearn/{self.environment_name}"
if config_file is None:
config_file = CONFIG_FILE
self.hyperparameters = import_hyperparameters(
config_file,
self.environment_name
)
super().__init__(environment, self.hyperparameters, activation=nn.ELU)
self.q_target = copy.deepcopy(self.policy)
self.q_target.to(self.device)
self.q_target.load_state_dict(self.policy.state_dict())
self.actor_optimizer = None
self.log_alpha = torch.tensor(np.log(self.hyperparameters.get("alpha", 0.5))).to(self.device)
@property
def alpha(self):
return self.log_alpha.exp()
def get_value(self, obs, actions=None):
q = self.policy(obs)
v = self.alpha * torch.logsumexp(q / self.alpha, dim=1, keepdim=True)
return v
def get_targetV(self, obs):
q = self.q_target(obs)
v = self.alpha * torch.logsumexp(q / self.alpha, dim=1, keepdim=True)
return v
def save(self, path: str = None, name: str = None) -> None:
path = self.save_path if path is None else path
if not os.path.exists(path):
os.makedirs(path, exist_ok=True)
if self.policy:
_name = "best_q_network.ckpt" if name is None else f"{name}_q_network.ckpt"
torch.save(self.policy.state_dict(), f"{path}/{_name}")
if self.q_target:
_name = "best_q_target.ckpt" if name is None else f"{name}_q_target.ckpt"
torch.save(self.q_target.state_dict(), f"{path}/{_name}")
def load(self, path: str = None, name: str = None) -> Self:
path = self.save_path if path is None else path
name = "best" if name is None else name
if not os.path.exists(path):
raise ValueError("Path does not exist.")
if self.policy:
state_dict = torch.load(
f"{path}/{name}_q_network.ckpt",
map_location=torch.device(self.device)
)
self.policy.load_state_dict(state_dict)
if self.q_target:
state_dict = torch.load(
f"{path}/{name}_q_target.ckpt",
map_location=torch.device(self.device)
)
self.q_target.load_state_dict(state_dict)
return self
@torch.no_grad()
def predict(
self,
obs: Union[torch.Tensor, np.ndarray],
transforms: Callable[[torch.Tensor], torch.Tensor] = None,
deterministic: bool = True
) -> Number:
self.policy.eval()
if isinstance(obs, np.ndarray) and transforms is None:
obs = torch.from_numpy(obs)
if transforms is not None:
obs = transforms(obs)
obs = obs.to(self.device)
if len(obs.shape) in [1, 3]:
obs = obs[None]
q_values = self.policy(obs)
if deterministic:
action = q_values.argmax(dim=1)
else:
probs = F.softmax(q_values / self.hyperparameters.get("temperature"), dim=1)
action = torch.multinomial(probs, 1)
return action.cpu().numpy()[0]
def train(
self,
n_epochs: int,
train_dataset: DataLoader,
eval_dataset: DataLoader = None,
transforms: Callable[[torch.Tensor], torch.Tensor] = None,
always_save: bool = True,
folder: str = None
) -> Self:
if folder is None:
folder = f"./benchmark_results/iqlearn/{self.environment_name}"
if not os.path.exists(folder):
os.makedirs(f"{folder}/", exist_ok=True)
board = Tensorboard(path=folder, delete=True)
hparams = {}
for key, value in self.hyperparameters.items():
if isinstance(value, list):
value = ",".join([str(v) for v in value])
hparams[key] = value
board.add_hparams(hparams)
self.policy.to(self.device)
self.q_target.to(self.device)
best_model = -np.inf
pbar = range(n_epochs)
if self.verbose:
pbar = tqdm(pbar, desc=self.__method_name__)
for epoch in pbar:
train_metrics = self.loop(train_dataset, train=True)
board.add_scalars("Train", epoch="train", **train_metrics)
if eval_dataset is not None:
eval_metrics = self.loop(eval_dataset, train=False)
board.add_scalars("Eval", epoch="eval", **eval_metrics)
board.step(["train", "eval"])
else:
board.step("train")
if epoch % self.hyperparameters.get("update_frequency", 4) == 0:
tau = self.hyperparameters.get("tau", 0.05)
for target_param, param in zip(self.q_target.parameters(), self.policy.parameters()):
target_param.data.copy_(tau * param.data + (1.0 - tau) * target_param.data)
if epoch > 0 and epoch % self.enjoy_criteria == 0:
metrics = self._enjoy(transforms=transforms)
board.add_scalars("Enjoy", epoch="enjoy", **metrics)
board.step("enjoy")
if best_model < metrics["aer"] or always_save:
if best_model < metrics["aer"]:
self.save()
best_model = metrics["aer"]
if always_save:
self.save(name=epoch)
if self.early_stop(metrics["aer"]):
return self
return self
def early_stop(self, metric: Metrics) -> bool:
return False
def loop(self, dataset: DataLoader, train: bool = True) -> Metrics:
"""
This is an offline implementation only from IQLearn.
Therefore, we are using the expert data as both policy and expert.
"""
metrics = defaultdict(list)
if train:
self.policy.train()
self.q_target.train()
else:
self.policy.eval()
self.q_target.eval()
pbar = dataset
for batch in pbar:
state, action, next_state, done = batch
state = state.to(self.device)
action = action.to(self.device)
next_state = next_state.to(self.device)
done = done.to(self.device)
is_expert = torch.ones_like(done).to(self.device)
expert_mask = is_expert.unsqueeze(-1).float()
n_expert = expert_mask.sum().clamp(min=1)
if train:
self.optimizer_fn.zero_grad()
# Compute Q(s, a) and V(s), V(s')
current_q = self.policy(state)
current_q = current_q.gather(1, action.reshape(-1, 1).long())
current_v = self.get_value(state)
# Compute target soft-value for next states (no gradients needed)
with torch.no_grad():
next_v = self.get_targetV(next_state)
# Target: y = (1 - done) * γ * V(s')
y = (1 - done.reshape(-1, 1).float()) * self.hyperparameters.get("gamma", 0.99) * next_v
# Implied reward: r = Q(s, a) - y
reward = current_q - y
# ===== Component 1: Expert soft-Q loss =====
phi_grad = reward * expert_mask
soft_q_loss = -(phi_grad.sum() / n_expert)
# ===== Component 2: Value function loss =====
value_loss = ((current_v - y) * expert_mask).sum() / n_expert
# ===== Component 3: Chi-squared regularization =====
chi2_loss = torch.tensor(0.0).to(self.device)
if self.hyperparameters.get("regularize", True):
chi2_loss = 1.0 / (4.0 * self.hyperparameters.get("alpha", 0.5)) * (reward.pow(2)).mean()
# ===== Total loss =====
if self.hyperparameters.get("regularize", True):
loss_q = soft_q_loss + value_loss + chi2_loss
else:
loss_q = soft_q_loss + value_loss
if train:
loss_q.backward()
self.optimizer_fn.step()
metrics["loss/q"].append(loss_q.item())
metrics["loss/soft_q"].append(soft_q_loss.item())
metrics["loss/value"].append(value_loss.item())
metrics["loss/chi2"].append(chi2_loss.item())
metrics["reward/mean"].append(reward.mean().item())
metrics["value/policy"].append(current_v.mean().item())
return {key: np.mean(value) for key, value in metrics.items()}
|