Skip to content

velora.models.nf.modules

Documentation

Customization: Modules

Class-based modules used inside Velora's agents, simplifying underlying PyTorch functionality.

ActorModule

Bases: BaseModule

An Actor module for NeuroFlow. Uses a Liquid NCP SAC Actor with a Gaussian policy.

Usable with continuous action spaces.

Source code in velora/models/nf/modules.py
Python
 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
class ActorModule(BaseModule):
    """
    An Actor module for NeuroFlow. Uses a Liquid NCP SAC Actor with a
    Gaussian policy.

    Usable with continuous action spaces.
    """

    def __init__(
        self,
        state_dim: int,
        n_neurons: int,
        action_dim: int,
        action_scale: torch.Tensor,
        action_bias: torch.Tensor,
        *,
        log_std_min: float = -5,
        log_std_max: float = 2,
        optim: Type[optim.Optimizer] = optim.Adam,
        lr: float = 3e-4,
        device: torch.device | None = None,
    ):
        """
        Parameters:
            state_dim (int): dimension of the state space
            n_neurons (int): number of decision/hidden neurons
            action_dim (int): dimension of the action space
            action_scale (torch.Tensor): scale factor to map normalized actions to
                environment's action range
            action_bias (torch.Tensor): bias/offset to center normalized actions to
                environment's action range
            log_std_min (float, optional): minimum log standard deviation
            log_std_max (float, optional): maximum log standard deviation
            optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
            lr (float, optional): optimizer learning rate
            device (torch.device, optional): the device to perform computations on
        """
        self.state_dim = state_dim
        self.n_neurons = n_neurons
        self.action_dim = action_dim
        self.action_scale = action_scale
        self.action_bias = action_bias
        self.log_std = (log_std_min, log_std_max)
        self.lr = lr
        self.device = device

        self.network = SACActor(
            state_dim,
            n_neurons,
            action_dim,
            action_scale,
            action_bias,
            log_std_min=log_std_min,
            log_std_max=log_std_max,
            device=device,
        ).to(device)

        self.hidden_size = self.network.ncp.hidden_size

        self.optim = optim(self.network.parameters(), lr=lr)

        self.config = self.network.config()

        self.active_params = self.config.active_params
        self.total_params = self.config.total_params

        self.network: SACActor = torch.jit.script(self.network)

    def gradient_step(self, loss: torch.Tensor) -> None:
        """
        Performs a gradient update step.

        Parameters:
            loss (torch.Tensor): network loss
        """
        self.optim.zero_grad()
        loss.backward()
        self.optim.step()

    def predict(
        self,
        obs: torch.Tensor,
        hidden: torch.Tensor | None = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Makes a deterministic prediction using the network.

        Parameters:
            obs (torch.Tensor): the batch of state observations
            hidden (torch.Tensor, optional): the current hidden state

        Returns:
            actions (torch.Tensor): the action predictions.
            hidden (torch.Tensor): the new hidden state.
        """
        action, hidden = self.network.predict(obs, hidden)
        return action, hidden

    def forward(
        self, obs: torch.Tensor, hidden: torch.Tensor | None = None
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Performs a forward pass through the network.

        Parameters:
            obs (torch.Tensor): the batch of state observations
            hidden (torch.Tensor, optional): the current hidden state

        Returns:
            actions (torch.Tensor): the action predictions.
            log_prob (torch.Tensor): log probabilities of actions.
            hidden (torch.Tensor): the new hidden state.
        """
        action, log_prob, hidden = self.network(obs, hidden)
        return action, log_prob, hidden

    def eval_mode(self) -> None:
        """Sets the network to evaluation mode."""
        self.network.eval()

    def train_mode(self) -> None:
        """Sets the network to training mode."""
        self.network.train()

    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        return {
            "actor": self.network.state_dict(),
            "actor_optim": self.optim.state_dict(),
        }

    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        self.network.load_state_dict(state_dict["actor"])
        self.optim.load_state_dict(state_dict["actor_optim"])

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}("
            f"state_dim={self.state_dim}, "
            f"n_neurons={self.n_neurons}, "
            f"action_dim={self.action_dim}, "
            f"action_scale={self.action_scale}, "
            f"action_bias={self.action_bias}, "
            f"optim={type(self.optim).__name__}, "
            f"log_std={self.log_std}, "
            f"lr={self.lr}, "
            f"device={self.device})"
        )

__init__(state_dim, n_neurons, action_dim, action_scale, action_bias, *, log_std_min=-5, log_std_max=2, optim=optim.Adam, lr=0.0003, device=None)

Parameters:

Name Type Description Default
state_dim int

dimension of the state space

 required
n_neurons int

number of decision/hidden neurons

 required
action_dim int

dimension of the action space

 required
action_scale torch.Tensor

scale factor to map normalized actions to environment's action range

 required
action_bias torch.Tensor

bias/offset to center normalized actions to environment's action range

 required
log_std_min float

minimum log standard deviation

 -5
log_std_max float

maximum log standard deviation

 2
optim Type[optim.Optimizer]

a PyTorch optimizer class

 optim.Adam
lr float

optimizer learning rate

 0.0003
device torch.device

the device to perform computations on

 None
Source code in velora/models/nf/modules.py
Python
 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
def __init__(
    self,
    state_dim: int,
    n_neurons: int,
    action_dim: int,
    action_scale: torch.Tensor,
    action_bias: torch.Tensor,
    *,
    log_std_min: float = -5,
    log_std_max: float = 2,
    optim: Type[optim.Optimizer] = optim.Adam,
    lr: float = 3e-4,
    device: torch.device | None = None,
):
    """
    Parameters:
        state_dim (int): dimension of the state space
        n_neurons (int): number of decision/hidden neurons
        action_dim (int): dimension of the action space
        action_scale (torch.Tensor): scale factor to map normalized actions to
            environment's action range
        action_bias (torch.Tensor): bias/offset to center normalized actions to
            environment's action range
        log_std_min (float, optional): minimum log standard deviation
        log_std_max (float, optional): maximum log standard deviation
        optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
        lr (float, optional): optimizer learning rate
        device (torch.device, optional): the device to perform computations on
    """
    self.state_dim = state_dim
    self.n_neurons = n_neurons
    self.action_dim = action_dim
    self.action_scale = action_scale
    self.action_bias = action_bias
    self.log_std = (log_std_min, log_std_max)
    self.lr = lr
    self.device = device

    self.network = SACActor(
        state_dim,
        n_neurons,
        action_dim,
        action_scale,
        action_bias,
        log_std_min=log_std_min,
        log_std_max=log_std_max,
        device=device,
    ).to(device)

    self.hidden_size = self.network.ncp.hidden_size

    self.optim = optim(self.network.parameters(), lr=lr)

    self.config = self.network.config()

    self.active_params = self.config.active_params
    self.total_params = self.config.total_params

    self.network: SACActor = torch.jit.script(self.network)

eval_mode()

Sets the network to evaluation mode.

Source code in velora/models/nf/modules.py
Python
157
158
159
def eval_mode(self) -> None:
    """Sets the network to evaluation mode."""
    self.network.eval()

forward(obs, hidden=None)

Performs a forward pass through the network.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required
hidden torch.Tensor

the current hidden state

 None

Returns:

Name Type Description
actions torch.Tensor

the action predictions.
log_prob torch.Tensor

log probabilities of actions.
hidden torch.Tensor

the new hidden state.
Source code in velora/models/nf/modules.py
Python
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
def forward(
    self, obs: torch.Tensor, hidden: torch.Tensor | None = None
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Performs a forward pass through the network.

    Parameters:
        obs (torch.Tensor): the batch of state observations
        hidden (torch.Tensor, optional): the current hidden state

    Returns:
        actions (torch.Tensor): the action predictions.
        log_prob (torch.Tensor): log probabilities of actions.
        hidden (torch.Tensor): the new hidden state.
    """
    action, log_prob, hidden = self.network(obs, hidden)
    return action, log_prob, hidden

gradient_step(loss)

Performs a gradient update step.

Parameters:

Name Type Description Default
loss torch.Tensor

network loss

 required
Source code in velora/models/nf/modules.py
Python
109
110
111
112
113
114
115
116
117
118
def gradient_step(self, loss: torch.Tensor) -> None:
    """
    Performs a gradient update step.

    Parameters:
        loss (torch.Tensor): network loss
    """
    self.optim.zero_grad()
    loss.backward()
    self.optim.step()

predict(obs, hidden=None)

Makes a deterministic prediction using the network.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required
hidden torch.Tensor

the current hidden state

 None

Returns:

Name Type Description
actions torch.Tensor

the action predictions.
hidden torch.Tensor

the new hidden state.
Source code in velora/models/nf/modules.py
Python
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
def predict(
    self,
    obs: torch.Tensor,
    hidden: torch.Tensor | None = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Makes a deterministic prediction using the network.

    Parameters:
        obs (torch.Tensor): the batch of state observations
        hidden (torch.Tensor, optional): the current hidden state

    Returns:
        actions (torch.Tensor): the action predictions.
        hidden (torch.Tensor): the new hidden state.
    """
    action, hidden = self.network.predict(obs, hidden)
    return action, hidden

train_mode()

Sets the network to training mode.

Source code in velora/models/nf/modules.py
Python
161
162
163
def train_mode(self) -> None:
    """Sets the network to training mode."""
    self.network.train()

ActorModuleDiscrete

Bases: BaseModule

An Actor module for NeuroFlow. Uses a Liquid NCP SAC Actor with a Categorical policy.

Usable with discrete action spaces.

Source code in velora/models/nf/modules.py
Python
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
class ActorModuleDiscrete(BaseModule):
    """
    An Actor module for NeuroFlow. Uses a Liquid NCP SAC Actor with a
    Categorical policy.

    Usable with discrete action spaces.
    """

    def __init__(
        self,
        state_dim: int,
        n_neurons: int,
        action_dim: int,
        *,
        optim: Type[optim.Optimizer] = optim.Adam,
        lr: float = 3e-4,
        device: torch.device | None = None,
    ):
        """
        Parameters:
            state_dim (int): dimension of the state space
            n_neurons (int): number of decision/hidden neurons
            action_dim (int): dimension of the action space
            optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
            lr (float, optional): optimizer learning rate
            device (torch.device, optional): the device to perform computations on
        """
        self.state_dim = state_dim
        self.n_neurons = n_neurons
        self.action_dim = action_dim
        self.lr = lr
        self.device = device

        self.network = SACActorDiscrete(
            state_dim,
            n_neurons,
            action_dim,
            device=device,
        ).to(device)

        self.hidden_size = self.network.ncp.hidden_size

        self.optim = optim(self.network.parameters(), lr=lr)

        self.config = self.network.config()

        self.active_params = self.config.active_params
        self.total_params = self.config.total_params

        self.network: SACActorDiscrete = torch.jit.script(self.network)

    def gradient_step(self, loss: torch.Tensor) -> None:
        """
        Performs a gradient update step.

        Parameters:
            loss (torch.Tensor): network loss
        """
        self.optim.zero_grad()
        loss.backward()
        self.optim.step()

    def predict(
        self,
        obs: torch.Tensor,
        hidden: torch.Tensor | None = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Makes a deterministic prediction using the network.

        Parameters:
            obs (torch.Tensor): the batch of state observations
            hidden (torch.Tensor, optional): the current hidden state

        Returns:
            actions (torch.Tensor): the action predictions.
            hidden (torch.Tensor): the new hidden state.
        """
        action, hidden = self.network.predict(obs, hidden)
        return action, hidden

    def forward(
        self, obs: torch.Tensor, hidden: torch.Tensor | None = None
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Performs a forward pass through the network.

        Parameters:
            obs (torch.Tensor): the batch of state observations
            hidden (torch.Tensor, optional): the current hidden state

        Returns:
            actions (torch.Tensor): the action predictions.
            probs (torch.Tensor): softmax probabilities for each action.
            log_prob (torch.Tensor): log probabilities of actions.
            hidden (torch.Tensor): the new hidden state.
        """
        actions, probs, log_prob, hidden = self.network(obs, hidden)
        return actions, probs, log_prob, hidden

    def eval_mode(self) -> None:
        """Sets the network to evaluation mode."""
        self.network.eval()

    def train_mode(self) -> None:
        """Sets the network to training mode."""
        self.network.train()

    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        return {
            "actor": self.network.state_dict(),
            "actor_optim": self.optim.state_dict(),
        }

    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        self.network.load_state_dict(state_dict["actor"])
        self.optim.load_state_dict(state_dict["actor_optim"])

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}("
            f"state_dim={self.state_dim}, "
            f"n_neurons={self.n_neurons}, "
            f"action_dim={self.action_dim}, "
            f"optim={type(self.optim).__name__}, "
            f"lr={self.lr}, "
            f"device={self.device})"
        )

__init__(state_dim, n_neurons, action_dim, *, optim=optim.Adam, lr=0.0003, device=None)

Parameters:

Name Type Description Default
state_dim int

dimension of the state space

 required
n_neurons int

number of decision/hidden neurons

 required
action_dim int

dimension of the action space

 required
optim Type[optim.Optimizer]

a PyTorch optimizer class

 optim.Adam
lr float

optimizer learning rate

 0.0003
device torch.device

the device to perform computations on

 None
Source code in velora/models/nf/modules.py
Python
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
def __init__(
    self,
    state_dim: int,
    n_neurons: int,
    action_dim: int,
    *,
    optim: Type[optim.Optimizer] = optim.Adam,
    lr: float = 3e-4,
    device: torch.device | None = None,
):
    """
    Parameters:
        state_dim (int): dimension of the state space
        n_neurons (int): number of decision/hidden neurons
        action_dim (int): dimension of the action space
        optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
        lr (float, optional): optimizer learning rate
        device (torch.device, optional): the device to perform computations on
    """
    self.state_dim = state_dim
    self.n_neurons = n_neurons
    self.action_dim = action_dim
    self.lr = lr
    self.device = device

    self.network = SACActorDiscrete(
        state_dim,
        n_neurons,
        action_dim,
        device=device,
    ).to(device)

    self.hidden_size = self.network.ncp.hidden_size

    self.optim = optim(self.network.parameters(), lr=lr)

    self.config = self.network.config()

    self.active_params = self.config.active_params
    self.total_params = self.config.total_params

    self.network: SACActorDiscrete = torch.jit.script(self.network)

eval_mode()

Sets the network to evaluation mode.

Source code in velora/models/nf/modules.py
Python
290
291
292
def eval_mode(self) -> None:
    """Sets the network to evaluation mode."""
    self.network.eval()

forward(obs, hidden=None)

Performs a forward pass through the network.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required
hidden torch.Tensor

the current hidden state

 None

Returns:

Name Type Description
actions torch.Tensor

the action predictions.
probs torch.Tensor

softmax probabilities for each action.
log_prob torch.Tensor

log probabilities of actions.
hidden torch.Tensor

the new hidden state.
Source code in velora/models/nf/modules.py
Python
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
def forward(
    self, obs: torch.Tensor, hidden: torch.Tensor | None = None
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Performs a forward pass through the network.

    Parameters:
        obs (torch.Tensor): the batch of state observations
        hidden (torch.Tensor, optional): the current hidden state

    Returns:
        actions (torch.Tensor): the action predictions.
        probs (torch.Tensor): softmax probabilities for each action.
        log_prob (torch.Tensor): log probabilities of actions.
        hidden (torch.Tensor): the new hidden state.
    """
    actions, probs, log_prob, hidden = self.network(obs, hidden)
    return actions, probs, log_prob, hidden

gradient_step(loss)

Performs a gradient update step.

Parameters:

Name Type Description Default
loss torch.Tensor

network loss

 required
Source code in velora/models/nf/modules.py
Python
241
242
243
244
245
246
247
248
249
250
def gradient_step(self, loss: torch.Tensor) -> None:
    """
    Performs a gradient update step.

    Parameters:
        loss (torch.Tensor): network loss
    """
    self.optim.zero_grad()
    loss.backward()
    self.optim.step()

predict(obs, hidden=None)

Makes a deterministic prediction using the network.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required
hidden torch.Tensor

the current hidden state

 None

Returns:

Name Type Description
actions torch.Tensor

the action predictions.
hidden torch.Tensor

the new hidden state.
Source code in velora/models/nf/modules.py
Python
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
def predict(
    self,
    obs: torch.Tensor,
    hidden: torch.Tensor | None = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Makes a deterministic prediction using the network.

    Parameters:
        obs (torch.Tensor): the batch of state observations
        hidden (torch.Tensor, optional): the current hidden state

    Returns:
        actions (torch.Tensor): the action predictions.
        hidden (torch.Tensor): the new hidden state.
    """
    action, hidden = self.network.predict(obs, hidden)
    return action, hidden

train_mode()

Sets the network to training mode.

Source code in velora/models/nf/modules.py
Python
294
295
296
def train_mode(self) -> None:
    """Sets the network to training mode."""
    self.network.train()

BaseModule

Bases: ABC

A base module for all agent modules.

Source code in velora/models/nf/modules.py
Python
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
class BaseModule(ABC):
    """
    A base module for all agent modules.
    """

    @abstractmethod
    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        """
        Creates a state dictionary for the module.

        Returns:
            state_dict (Dict[str, Dict[str, Any]]): the state dicts for the module, including networks and optimizers.
        """
        pass  # pragma: no cover

    @abstractmethod
    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        """
        Load the modules state dict from a previously saved state.

        Parameters:
            state_dict (Dict[str, Dict[str, Any]]): a previously saved state dict
        """
        pass  # pragma: no cover

load_state_dict(state_dict) abstractmethod

Load the modules state dict from a previously saved state.

Parameters:

Name Type Description Default
state_dict Dict[str, Dict[str, Any]]

a previously saved state dict

 required
Source code in velora/models/nf/modules.py
Python
30
31
32
33
34
35
36
37
38
@abstractmethod
def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
    """
    Load the modules state dict from a previously saved state.

    Parameters:
        state_dict (Dict[str, Dict[str, Any]]): a previously saved state dict
    """
    pass  # pragma: no cover

state_dict() abstractmethod

Creates a state dictionary for the module.

Returns:

Name Type Description
state_dict Dict[str, Dict[str, Any]]

the state dicts for the module, including networks and optimizers.
Source code in velora/models/nf/modules.py
Python
20
21
22
23
24
25
26
27
28
@abstractmethod
def state_dict(self) -> Dict[str, Dict[str, Any]]:
    """
    Creates a state dictionary for the module.

    Returns:
        state_dict (Dict[str, Dict[str, Any]]): the state dicts for the module, including networks and optimizers.
    """
    pass  # pragma: no cover

CriticModule

Bases: BaseModule

A Critic module for NeuroFlow. Uses a pair of NCP SAC Critic's with separate target networks to estimate Q-values.

Usable with continuous action spaces.

Source code in velora/models/nf/modules.py
Python
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
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
class CriticModule(BaseModule):
    """
    A Critic module for NeuroFlow. Uses a pair of NCP SAC Critic's with separate
    target networks to estimate Q-values.

    Usable with continuous action spaces.
    """

    def __init__(
        self,
        state_dim: int,
        n_neurons: int,
        action_dim: int,
        *,
        optim: Type[optim.Optimizer] = optim.Adam,
        lr: float = 3e-4,
        tau: float = 0.005,
        device: torch.device | None = None,
    ):
        """
        Parameters:
            state_dim (int): dimension of the state space
            n_neurons (int): number of decision/hidden neurons
            action_dim (int): dimension of the action space
            optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
            lr (float, optional): optimizer learning rates
            tau (float, optional): soft target network update factor
            device (torch.device, optional): the device to perform computations on
        """
        self.state_dim = state_dim
        self.n_neurons = n_neurons
        self.action_dim = action_dim
        self.lr = lr
        self.tau = tau
        self.device = device

        self.network1 = SACCriticNCP(
            state_dim,
            n_neurons,
            action_dim,
            device=device,
        ).to(device)

        self.network2 = SACCriticNCP(
            state_dim,
            n_neurons,
            action_dim,
            device=device,
        ).to(device)

        self.target1 = deepcopy(self.network1)
        self.target2 = deepcopy(self.network2)

        self.optim1 = optim(self.network1.parameters(), lr=lr)
        self.optim2 = optim(self.network2.parameters(), lr=lr)

        self.config = CriticConfig(
            critic1=self.network1.config(),
            critic2=self.network2.config(),
        )

        self.active_params = (
            self.config.critic1.active_params + self.config.critic2.active_params
        )
        self.total_params = (
            self.config.critic1.total_params + self.config.critic2.total_params
        )

        self.network1: SACCriticNCP = torch.jit.script(self.network1)
        self.network2: SACCriticNCP = torch.jit.script(self.network2)

        self.target1: SACCriticNCP = torch.jit.script(self.target1)
        self.target2: SACCriticNCP = torch.jit.script(self.target2)

    def update_targets(self) -> None:
        """
        Performs a soft update on the target networks.
        """
        soft_update(self.network1, self.target1, tau=self.tau)
        soft_update(self.network2, self.target2, tau=self.tau)

    def gradient_step(self, c1_loss: torch.Tensor, c2_loss: torch.Tensor) -> None:
        """
        Performs a gradient update step.

        Parameters:
            c1_loss (torch.Tensor): critic loss for first network
            c2_loss (torch.Tensor): critic loss for second network
        """
        self.optim1.zero_grad()
        c1_loss.backward()
        self.optim1.step()

        self.optim2.zero_grad()
        c2_loss.backward()
        self.optim2.step()

    def predict(
        self, obs: torch.Tensor, actions: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Makes a prediction using the critic networks.

        Parameters:
            obs (torch.Tensor): the batch of state observations
            actions (torch.Tensor): the batch of actions

        Returns:
            q_values1 (torch.Tensor): the Q-Value predictions from the first network.
            q_values2 (torch.Tensor): the Q-Value predictions from the second network.
        """
        q_values1 = self.network1(obs, actions)
        q_values2 = self.network2(obs, actions)

        return q_values1, q_values2

    def target_predict(self, obs: torch.Tensor, actions: torch.Tensor) -> torch.Tensor:
        """
        Makes a prediction using the target networks.

        Parameters:
            obs (torch.Tensor): the batch of state observations
            actions (torch.Tensor): the batch of actions

        Returns:
            next_q (torch.Tensor): the smallest Q-Value predictions between the target networks.
        """
        q_values1 = self.target1(obs, actions)
        q_values2 = self.target2(obs, actions)

        return torch.min(q_values1, q_values2)

    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        return {
            "critic": self.network1.state_dict(),
            "critic2": self.network2.state_dict(),
            "critic_target": self.target1.state_dict(),
            "critic2_target": self.target2.state_dict(),
            "critic_optim": self.optim1.state_dict(),
            "critic2_optim": self.optim2.state_dict(),
        }

    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        self.network1.load_state_dict(state_dict["critic"])
        self.network2.load_state_dict(state_dict["critic2"])

        self.target1.load_state_dict(state_dict["critic_target"])
        self.target2.load_state_dict(state_dict["critic2_target"])

        self.optim1.load_state_dict(state_dict["critic_optim"])
        self.optim2.load_state_dict(state_dict["critic2_optim"])

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}("
            f"state_dim={self.state_dim}, "
            f"n_neurons={self.n_neurons}, "
            f"action_dim={self.action_dim}, "
            f"optim={type(self.optim1).__name__}, "
            f"lr={self.lr}, "
            f"tau={self.tau}, "
            f"device={self.device})"
        )

__init__(state_dim, n_neurons, action_dim, *, optim=optim.Adam, lr=0.0003, tau=0.005, device=None)

Parameters:

Name Type Description Default
state_dim int

dimension of the state space

 required
n_neurons int

number of decision/hidden neurons

 required
action_dim int

dimension of the action space

 required
optim Type[optim.Optimizer]

a PyTorch optimizer class

 optim.Adam
lr float

optimizer learning rates

 0.0003
tau float

soft target network update factor

 0.005
device torch.device

the device to perform computations on

 None
Source code in velora/models/nf/modules.py
Python
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
def __init__(
    self,
    state_dim: int,
    n_neurons: int,
    action_dim: int,
    *,
    optim: Type[optim.Optimizer] = optim.Adam,
    lr: float = 3e-4,
    tau: float = 0.005,
    device: torch.device | None = None,
):
    """
    Parameters:
        state_dim (int): dimension of the state space
        n_neurons (int): number of decision/hidden neurons
        action_dim (int): dimension of the action space
        optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
        lr (float, optional): optimizer learning rates
        tau (float, optional): soft target network update factor
        device (torch.device, optional): the device to perform computations on
    """
    self.state_dim = state_dim
    self.n_neurons = n_neurons
    self.action_dim = action_dim
    self.lr = lr
    self.tau = tau
    self.device = device

    self.network1 = SACCriticNCP(
        state_dim,
        n_neurons,
        action_dim,
        device=device,
    ).to(device)

    self.network2 = SACCriticNCP(
        state_dim,
        n_neurons,
        action_dim,
        device=device,
    ).to(device)

    self.target1 = deepcopy(self.network1)
    self.target2 = deepcopy(self.network2)

    self.optim1 = optim(self.network1.parameters(), lr=lr)
    self.optim2 = optim(self.network2.parameters(), lr=lr)

    self.config = CriticConfig(
        critic1=self.network1.config(),
        critic2=self.network2.config(),
    )

    self.active_params = (
        self.config.critic1.active_params + self.config.critic2.active_params
    )
    self.total_params = (
        self.config.critic1.total_params + self.config.critic2.total_params
    )

    self.network1: SACCriticNCP = torch.jit.script(self.network1)
    self.network2: SACCriticNCP = torch.jit.script(self.network2)

    self.target1: SACCriticNCP = torch.jit.script(self.target1)
    self.target2: SACCriticNCP = torch.jit.script(self.target2)

gradient_step(c1_loss, c2_loss)

Performs a gradient update step.

Parameters:

Name Type Description Default
c1_loss torch.Tensor

critic loss for first network

 required
c2_loss torch.Tensor

critic loss for second network

 required
Source code in velora/models/nf/modules.py
Python
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
def gradient_step(self, c1_loss: torch.Tensor, c2_loss: torch.Tensor) -> None:
    """
    Performs a gradient update step.

    Parameters:
        c1_loss (torch.Tensor): critic loss for first network
        c2_loss (torch.Tensor): critic loss for second network
    """
    self.optim1.zero_grad()
    c1_loss.backward()
    self.optim1.step()

    self.optim2.zero_grad()
    c2_loss.backward()
    self.optim2.step()

predict(obs, actions)

Makes a prediction using the critic networks.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required
actions torch.Tensor

the batch of actions

 required

Returns:

Name Type Description
q_values1 torch.Tensor

the Q-Value predictions from the first network.
q_values2 torch.Tensor

the Q-Value predictions from the second network.
Source code in velora/models/nf/modules.py
Python
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
def predict(
    self, obs: torch.Tensor, actions: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Makes a prediction using the critic networks.

    Parameters:
        obs (torch.Tensor): the batch of state observations
        actions (torch.Tensor): the batch of actions

    Returns:
        q_values1 (torch.Tensor): the Q-Value predictions from the first network.
        q_values2 (torch.Tensor): the Q-Value predictions from the second network.
    """
    q_values1 = self.network1(obs, actions)
    q_values2 = self.network2(obs, actions)

    return q_values1, q_values2

target_predict(obs, actions)

Makes a prediction using the target networks.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required
actions torch.Tensor

the batch of actions

 required

Returns:

Name Type Description
next_q torch.Tensor

the smallest Q-Value predictions between the target networks.
Source code in velora/models/nf/modules.py
Python
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
def target_predict(self, obs: torch.Tensor, actions: torch.Tensor) -> torch.Tensor:
    """
    Makes a prediction using the target networks.

    Parameters:
        obs (torch.Tensor): the batch of state observations
        actions (torch.Tensor): the batch of actions

    Returns:
        next_q (torch.Tensor): the smallest Q-Value predictions between the target networks.
    """
    q_values1 = self.target1(obs, actions)
    q_values2 = self.target2(obs, actions)

    return torch.min(q_values1, q_values2)

update_targets()

Performs a soft update on the target networks.

Source code in velora/models/nf/modules.py
Python
394
395
396
397
398
399
def update_targets(self) -> None:
    """
    Performs a soft update on the target networks.
    """
    soft_update(self.network1, self.target1, tau=self.tau)
    soft_update(self.network2, self.target2, tau=self.tau)

CriticModuleDiscrete

Bases: BaseModule

A Critic module for NeuroFlow. Uses a pair of NCP SAC Critic's with separate target networks to estimate Q-values.

Usable with discrete action spaces.

Source code in velora/models/nf/modules.py
Python
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
class CriticModuleDiscrete(BaseModule):
    """
    A Critic module for NeuroFlow. Uses a pair of NCP SAC Critic's with separate
    target networks to estimate Q-values.

    Usable with discrete action spaces.
    """

    def __init__(
        self,
        state_dim: int,
        n_neurons: int,
        action_dim: int,
        *,
        optim: Type[optim.Optimizer] = optim.Adam,
        lr: float = 3e-4,
        tau: float = 0.005,
        device: torch.device | None = None,
    ):
        """
        Parameters:
            state_dim (int): dimension of the state space
            n_neurons (int): number of decision/hidden neurons
            action_dim (int): dimension of the action space
            optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
            lr (float, optional): optimizer learning rates
            tau (float, optional): soft target network update factor
            device (torch.device, optional): the device to perform computations on
        """
        self.state_dim = state_dim
        self.n_neurons = n_neurons
        self.action_dim = action_dim
        self.lr = lr
        self.tau = tau
        self.device = device

        self.network1 = SACCriticNCPDiscrete(
            state_dim,
            n_neurons,
            action_dim,
            device=device,
        ).to(device)

        self.network2 = SACCriticNCPDiscrete(
            state_dim,
            n_neurons,
            action_dim,
            device=device,
        ).to(device)

        self.target1 = deepcopy(self.network1)
        self.target2 = deepcopy(self.network2)

        self.optim1 = optim(self.network1.parameters(), lr=lr)
        self.optim2 = optim(self.network2.parameters(), lr=lr)

        self.config = CriticConfig(
            critic1=self.network1.config(),
            critic2=self.network2.config(),
        )

        self.active_params = (
            self.config.critic1.active_params + self.config.critic2.active_params
        )
        self.total_params = (
            self.config.critic1.total_params + self.config.critic2.total_params
        )

        self.network1: SACCriticNCPDiscrete = torch.jit.script(self.network1)
        self.network2: SACCriticNCPDiscrete = torch.jit.script(self.network2)

        self.target1: SACCriticNCPDiscrete = torch.jit.script(self.target1)
        self.target2: SACCriticNCPDiscrete = torch.jit.script(self.target2)

    def update_targets(self) -> None:
        """
        Performs a soft update on the target networks.
        """
        soft_update(self.network1, self.target1, tau=self.tau)
        soft_update(self.network2, self.target2, tau=self.tau)

    def gradient_step(self, c1_loss: torch.Tensor, c2_loss: torch.Tensor) -> None:
        """
        Performs a gradient update step.

        Parameters:
            c1_loss (torch.Tensor): critic loss for first network
            c2_loss (torch.Tensor): critic loss for second network
        """
        self.optim1.zero_grad()
        c1_loss.backward()
        self.optim1.step()

        self.optim2.zero_grad()
        c2_loss.backward()
        self.optim2.step()

    def predict(self, obs: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Makes a prediction using the critic networks.

        Parameters:
            obs (torch.Tensor): the batch of state observations

        Returns:
            q_values1 (torch.Tensor): the Q-Value predictions from the first network.
            q_values2 (torch.Tensor): the Q-Value predictions from the second network.
        """
        q_values1 = self.network1(obs)
        q_values2 = self.network2(obs)

        return q_values1, q_values2

    def target_predict(self, obs: torch.Tensor) -> torch.Tensor:
        """
        Makes a prediction using the target networks.

        Parameters:
            obs (torch.Tensor): the batch of state observations

        Returns:
            next_q (torch.Tensor): the smallest Q-Value predictions between the target networks.
        """
        q_values1 = self.target1(obs)
        q_values2 = self.target2(obs)

        return torch.min(q_values1, q_values2)

    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        return {
            "critic": self.network1.state_dict(),
            "critic2": self.network2.state_dict(),
            "critic_target": self.target1.state_dict(),
            "critic2_target": self.target2.state_dict(),
            "critic_optim": self.optim1.state_dict(),
            "critic2_optim": self.optim2.state_dict(),
        }

    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        self.network1.load_state_dict(state_dict["critic"])
        self.network2.load_state_dict(state_dict["critic2"])

        self.target1.load_state_dict(state_dict["critic_target"])
        self.target2.load_state_dict(state_dict["critic2_target"])

        self.optim1.load_state_dict(state_dict["critic_optim"])
        self.optim2.load_state_dict(state_dict["critic2_optim"])

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}("
            f"state_dim={self.state_dim}, "
            f"n_neurons={self.n_neurons}, "
            f"action_dim={self.action_dim}, "
            f"optim={type(self.optim1).__name__}, "
            f"lr={self.lr}, "
            f"tau={self.tau}, "
            f"device={self.device})"
        )

__init__(state_dim, n_neurons, action_dim, *, optim=optim.Adam, lr=0.0003, tau=0.005, device=None)

Parameters:

Name Type Description Default
state_dim int

dimension of the state space

 required
n_neurons int

number of decision/hidden neurons

 required
action_dim int

dimension of the action space

 required
optim Type[optim.Optimizer]

a PyTorch optimizer class

 optim.Adam
lr float

optimizer learning rates

 0.0003
tau float

soft target network update factor

 0.005
device torch.device

the device to perform computations on

 None
Source code in velora/models/nf/modules.py
Python
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
def __init__(
    self,
    state_dim: int,
    n_neurons: int,
    action_dim: int,
    *,
    optim: Type[optim.Optimizer] = optim.Adam,
    lr: float = 3e-4,
    tau: float = 0.005,
    device: torch.device | None = None,
):
    """
    Parameters:
        state_dim (int): dimension of the state space
        n_neurons (int): number of decision/hidden neurons
        action_dim (int): dimension of the action space
        optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
        lr (float, optional): optimizer learning rates
        tau (float, optional): soft target network update factor
        device (torch.device, optional): the device to perform computations on
    """
    self.state_dim = state_dim
    self.n_neurons = n_neurons
    self.action_dim = action_dim
    self.lr = lr
    self.tau = tau
    self.device = device

    self.network1 = SACCriticNCPDiscrete(
        state_dim,
        n_neurons,
        action_dim,
        device=device,
    ).to(device)

    self.network2 = SACCriticNCPDiscrete(
        state_dim,
        n_neurons,
        action_dim,
        device=device,
    ).to(device)

    self.target1 = deepcopy(self.network1)
    self.target2 = deepcopy(self.network2)

    self.optim1 = optim(self.network1.parameters(), lr=lr)
    self.optim2 = optim(self.network2.parameters(), lr=lr)

    self.config = CriticConfig(
        critic1=self.network1.config(),
        critic2=self.network2.config(),
    )

    self.active_params = (
        self.config.critic1.active_params + self.config.critic2.active_params
    )
    self.total_params = (
        self.config.critic1.total_params + self.config.critic2.total_params
    )

    self.network1: SACCriticNCPDiscrete = torch.jit.script(self.network1)
    self.network2: SACCriticNCPDiscrete = torch.jit.script(self.network2)

    self.target1: SACCriticNCPDiscrete = torch.jit.script(self.target1)
    self.target2: SACCriticNCPDiscrete = torch.jit.script(self.target2)

gradient_step(c1_loss, c2_loss)

Performs a gradient update step.

Parameters:

Name Type Description Default
c1_loss torch.Tensor

critic loss for first network

 required
c2_loss torch.Tensor

critic loss for second network

 required
Source code in velora/models/nf/modules.py
Python
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
def gradient_step(self, c1_loss: torch.Tensor, c2_loss: torch.Tensor) -> None:
    """
    Performs a gradient update step.

    Parameters:
        c1_loss (torch.Tensor): critic loss for first network
        c2_loss (torch.Tensor): critic loss for second network
    """
    self.optim1.zero_grad()
    c1_loss.backward()
    self.optim1.step()

    self.optim2.zero_grad()
    c2_loss.backward()
    self.optim2.step()

predict(obs)

Makes a prediction using the critic networks.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required

Returns:

Name Type Description
q_values1 torch.Tensor

the Q-Value predictions from the first network.
q_values2 torch.Tensor

the Q-Value predictions from the second network.
Source code in velora/models/nf/modules.py
Python
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
def predict(self, obs: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Makes a prediction using the critic networks.

    Parameters:
        obs (torch.Tensor): the batch of state observations

    Returns:
        q_values1 (torch.Tensor): the Q-Value predictions from the first network.
        q_values2 (torch.Tensor): the Q-Value predictions from the second network.
    """
    q_values1 = self.network1(obs)
    q_values2 = self.network2(obs)

    return q_values1, q_values2

target_predict(obs)

Makes a prediction using the target networks.

Parameters:

Name Type Description Default
obs torch.Tensor

the batch of state observations

 required

Returns:

Name Type Description
next_q torch.Tensor

the smallest Q-Value predictions between the target networks.
Source code in velora/models/nf/modules.py
Python
598
599
600
601
602
603
604
605
606
607
608
609
610
611
def target_predict(self, obs: torch.Tensor) -> torch.Tensor:
    """
    Makes a prediction using the target networks.

    Parameters:
        obs (torch.Tensor): the batch of state observations

    Returns:
        next_q (torch.Tensor): the smallest Q-Value predictions between the target networks.
    """
    q_values1 = self.target1(obs)
    q_values2 = self.target2(obs)

    return torch.min(q_values1, q_values2)

update_targets()

Performs a soft update on the target networks.

Source code in velora/models/nf/modules.py
Python
559
560
561
562
563
564
def update_targets(self) -> None:
    """
    Performs a soft update on the target networks.
    """
    soft_update(self.network1, self.target1, tau=self.tau)
    soft_update(self.network2, self.target2, tau=self.tau)

EntropyModule

Bases: BaseModule

An Entropy module for NeuroFlow. Uses automatic entropy tuning from SAC based on the paper: Soft Actor-Critic Algorithms and Applications.

Usable with continuous action spaces.

Source code in velora/models/nf/modules.py
Python
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
class EntropyModule(BaseModule):
    """
    An Entropy module for NeuroFlow. Uses automatic entropy tuning from SAC
    based on the paper: [Soft Actor-Critic Algorithms and Applications](https://arxiv.org/abs/1812.05905).

    Usable with continuous action spaces.
    """

    def __init__(
        self,
        action_dim: int,
        *,
        initial_alpha: float = 1.0,
        optim: Type[optim.Optimizer] = optim.Adam,
        lr: float = 3e-4,
        device: torch.device | None = None,
    ):
        """
        Parameters:
            action_dim (int): dimension of the action space
            initial_alpha (float, optional): the starting entropy coefficient value
            optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
            lr (float, optional): optimizer learning rates
            device (torch.device, optional): the device to perform computations on
        """
        self.action_dim = action_dim
        self.initial_alpha = initial_alpha
        self.lr = lr
        self.device = device

        self.target = -action_dim
        self.log_alpha = nn.Parameter(torch.tensor(initial_alpha, device=device).log())

        self.optim = optim([self.log_alpha], lr=lr)

    @property
    def alpha(self) -> torch.Tensor:
        """
        Get the current entropy coefficient (alpha).

        Returns:
            alpha (torch.Tensor): the entropy coefficient.
        """
        return self.log_alpha.exp()

    def compute_loss(self, log_probs: torch.Tensor) -> torch.Tensor:
        """
        Computes the entropy coefficient loss.

        Parameters:
            log_probs (torch.Tensor): log probabilities for actions

        Returns:
            loss (torch.Tensor): the entropy loss value.
        """
        loss = torch.tensor(0.0, device=log_probs.device)

        entropy = (log_probs + self.target).detach()
        loss = -(self.log_alpha * entropy).mean()

        return loss

    def gradient_step(self, loss: torch.Tensor) -> None:
        """
        Performs a gradient update step.

        Parameters:
            loss (torch.Tensor): loss to backpropagate
        """
        self.optim.zero_grad()
        loss.backward()
        self.optim.step()

    def config(self) -> EntropyParameters:
        """
        Creates a module config.

        Returns:
            config (EntropyParameters): a parameter config model.
        """
        return EntropyParameters(
            lr=self.lr,
            initial_alpha=self.initial_alpha,
            target=self.target,
        )

    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        return {
            "entropy_optim": self.optim.state_dict(),
        }

    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        self.optim.load_state_dict(state_dict["entropy_optim"])

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}("
            f"action_dim={self.action_dim}, "
            f"initial_alpha={self.initial_alpha}, "
            f"optim={type(self.optim).__name__}, "
            f"lr={self.lr}, "
            f"device={self.device})"
        )

alpha property

Get the current entropy coefficient (alpha).

Returns:

Name Type Description
alpha torch.Tensor

the entropy coefficient.

__init__(action_dim, *, initial_alpha=1.0, optim=optim.Adam, lr=0.0003, device=None)

Parameters:

Name Type Description Default
action_dim int

dimension of the action space

 required
initial_alpha float

the starting entropy coefficient value

 1.0
optim Type[optim.Optimizer]

a PyTorch optimizer class

 optim.Adam
lr float

optimizer learning rates

 0.0003
device torch.device

the device to perform computations on

 None
Source code in velora/models/nf/modules.py
Python
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
def __init__(
    self,
    action_dim: int,
    *,
    initial_alpha: float = 1.0,
    optim: Type[optim.Optimizer] = optim.Adam,
    lr: float = 3e-4,
    device: torch.device | None = None,
):
    """
    Parameters:
        action_dim (int): dimension of the action space
        initial_alpha (float, optional): the starting entropy coefficient value
        optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
        lr (float, optional): optimizer learning rates
        device (torch.device, optional): the device to perform computations on
    """
    self.action_dim = action_dim
    self.initial_alpha = initial_alpha
    self.lr = lr
    self.device = device

    self.target = -action_dim
    self.log_alpha = nn.Parameter(torch.tensor(initial_alpha, device=device).log())

    self.optim = optim([self.log_alpha], lr=lr)

compute_loss(log_probs)

Computes the entropy coefficient loss.

Parameters:

Name Type Description Default
log_probs torch.Tensor

log probabilities for actions

 required

Returns:

Name Type Description
loss torch.Tensor

the entropy loss value.
Source code in velora/models/nf/modules.py
Python
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
def compute_loss(self, log_probs: torch.Tensor) -> torch.Tensor:
    """
    Computes the entropy coefficient loss.

    Parameters:
        log_probs (torch.Tensor): log probabilities for actions

    Returns:
        loss (torch.Tensor): the entropy loss value.
    """
    loss = torch.tensor(0.0, device=log_probs.device)

    entropy = (log_probs + self.target).detach()
    loss = -(self.log_alpha * entropy).mean()

    return loss

config()

Creates a module config.

Returns:

Name Type Description
config EntropyParameters

a parameter config model.
Source code in velora/models/nf/modules.py
Python
719
720
721
722
723
724
725
726
727
728
729
730
def config(self) -> EntropyParameters:
    """
    Creates a module config.

    Returns:
        config (EntropyParameters): a parameter config model.
    """
    return EntropyParameters(
        lr=self.lr,
        initial_alpha=self.initial_alpha,
        target=self.target,
    )

gradient_step(loss)

Performs a gradient update step.

Parameters:

Name Type Description Default
loss torch.Tensor

loss to backpropagate

 required
Source code in velora/models/nf/modules.py
Python
708
709
710
711
712
713
714
715
716
717
def gradient_step(self, loss: torch.Tensor) -> None:
    """
    Performs a gradient update step.

    Parameters:
        loss (torch.Tensor): loss to backpropagate
    """
    self.optim.zero_grad()
    loss.backward()
    self.optim.step()

EntropyModuleDiscrete

Bases: BaseModule

An Entropy module for NeuroFlow. Uses automatic entropy tuning from SAC based on the paper: Soft Actor-Critic for Discrete Action Settings.

Usable with discrete action spaces.

Source code in velora/models/nf/modules.py
Python
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
class EntropyModuleDiscrete(BaseModule):
    """
    An Entropy module for NeuroFlow. Uses automatic entropy tuning from SAC
    based on the paper: [Soft Actor-Critic for Discrete Action Settings](https://arxiv.org/abs/1910.07207).

    Usable with discrete action spaces.
    """

    def __init__(
        self,
        action_dim: int,
        *,
        initial_alpha: float = 1.0,
        optim: Type[optim.Optimizer] = optim.Adam,
        lr: float = 3e-4,
        device: torch.device | None = None,
    ):
        """
        Parameters:
            action_dim (int): dimension of the action space
            initial_alpha (float, optional): the starting entropy coefficient value
            optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
            lr (float, optional): optimizer learning rates
            device (torch.device, optional): the device to perform computations on
        """
        self.action_dim = action_dim
        self.initial_alpha = initial_alpha
        self.lr = lr
        self.device = device

        self.target = 0.98 * torch.tensor(1 / action_dim, device=device).log()
        self.log_alpha = nn.Parameter(torch.tensor(initial_alpha, device=device).log())

        self.optim = optim([self.log_alpha], lr=lr)

    @property
    def alpha(self) -> torch.Tensor:
        """
        Get the current entropy coefficient (alpha).

        Returns:
            alpha (torch.Tensor): the entropy coefficient.
        """
        return self.log_alpha.exp()

    def compute_loss(
        self, probs: torch.Tensor, log_probs: torch.Tensor
    ) -> torch.Tensor:
        """
        Computes the entropy coefficient loss.

        Parameters:
            probs (torch.Tensor): probabilities for actions
            log_probs (torch.Tensor): log probabilities for actions

        Returns:
            loss (torch.Tensor): the entropy loss value.
        """
        with torch.no_grad():
            entropy_mean = -torch.sum(probs * log_probs, dim=1).mean()

        loss = self.log_alpha * (entropy_mean - self.target)
        return loss

    def gradient_step(self, loss: torch.Tensor) -> None:
        """
        Performs a gradient update step.

        Parameters:
            loss (torch.Tensor): loss to backpropagate
        """
        self.optim.zero_grad()
        loss.backward()
        self.optim.step()

    def config(self) -> EntropyParameters:
        """
        Creates a module config.

        Returns:
            config (EntropyParameters): a parameter config model.
        """
        return EntropyParameters(
            lr=self.lr,
            initial_alpha=self.initial_alpha,
            target=self.target,
        )

    def state_dict(self) -> Dict[str, Dict[str, Any]]:
        return {
            "entropy_optim": self.optim.state_dict(),
        }

    def load_state_dict(self, state_dict: Dict[str, Dict[str, Any]]) -> None:
        self.optim.load_state_dict(state_dict["entropy_optim"])

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}("
            f"action_dim={self.action_dim}, "
            f"initial_alpha={self.initial_alpha}, "
            f"optim={type(self.optim).__name__}, "
            f"lr={self.lr}, "
            f"device={self.device})"
        )

alpha property

Get the current entropy coefficient (alpha).

Returns:

Name Type Description
alpha torch.Tensor

the entropy coefficient.

__init__(action_dim, *, initial_alpha=1.0, optim=optim.Adam, lr=0.0003, device=None)

Parameters:

Name Type Description Default
action_dim int

dimension of the action space

 required
initial_alpha float

the starting entropy coefficient value

 1.0
optim Type[optim.Optimizer]

a PyTorch optimizer class

 optim.Adam
lr float

optimizer learning rates

 0.0003
device torch.device

the device to perform computations on

 None
Source code in velora/models/nf/modules.py
Python
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
def __init__(
    self,
    action_dim: int,
    *,
    initial_alpha: float = 1.0,
    optim: Type[optim.Optimizer] = optim.Adam,
    lr: float = 3e-4,
    device: torch.device | None = None,
):
    """
    Parameters:
        action_dim (int): dimension of the action space
        initial_alpha (float, optional): the starting entropy coefficient value
        optim (Type[optim.Optimizer], optional): a `PyTorch` optimizer class
        lr (float, optional): optimizer learning rates
        device (torch.device, optional): the device to perform computations on
    """
    self.action_dim = action_dim
    self.initial_alpha = initial_alpha
    self.lr = lr
    self.device = device

    self.target = 0.98 * torch.tensor(1 / action_dim, device=device).log()
    self.log_alpha = nn.Parameter(torch.tensor(initial_alpha, device=device).log())

    self.optim = optim([self.log_alpha], lr=lr)

compute_loss(probs, log_probs)

Computes the entropy coefficient loss.

Parameters:

Name Type Description Default
probs torch.Tensor

probabilities for actions

 required
log_probs torch.Tensor

log probabilities for actions

 required

Returns:

Name Type Description
loss torch.Tensor

the entropy loss value.
Source code in velora/models/nf/modules.py
Python
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
def compute_loss(
    self, probs: torch.Tensor, log_probs: torch.Tensor
) -> torch.Tensor:
    """
    Computes the entropy coefficient loss.

    Parameters:
        probs (torch.Tensor): probabilities for actions
        log_probs (torch.Tensor): log probabilities for actions

    Returns:
        loss (torch.Tensor): the entropy loss value.
    """
    with torch.no_grad():
        entropy_mean = -torch.sum(probs * log_probs, dim=1).mean()

    loss = self.log_alpha * (entropy_mean - self.target)
    return loss

config()

Creates a module config.

Returns:

Name Type Description
config EntropyParameters

a parameter config model.
Source code in velora/models/nf/modules.py
Python
826
827
828
829
830
831
832
833
834
835
836
837
def config(self) -> EntropyParameters:
    """
    Creates a module config.

    Returns:
        config (EntropyParameters): a parameter config model.
    """
    return EntropyParameters(
        lr=self.lr,
        initial_alpha=self.initial_alpha,
        target=self.target,
    )

gradient_step(loss)

Performs a gradient update step.

Parameters:

Name Type Description Default
loss torch.Tensor

loss to backpropagate

 required
Source code in velora/models/nf/modules.py
Python
815
816
817
818
819
820
821
822
823
824
def gradient_step(self, loss: torch.Tensor) -> None:
    """
    Performs a gradient update step.

    Parameters:
        loss (torch.Tensor): loss to backpropagate
    """
    self.optim.zero_grad()
    loss.backward()
    self.optim.step()