Inference time schedules
ContinuousInferenceSchedule
Bases: InferenceSchedule
A base class for continuous time inference schedules.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
__init__(nsteps, inclusive_end=False, min_t=0, padding=0, dilation=0, direction=TimeDirection.UNIFIED, device='cpu')
Initialize the ContinuousInferenceSchedule.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
int
|
Number of time steps. |
required |
inclusive_end
|
bool
|
If True, include the end value (1.0) in the schedule otherwise ends at 1.0-1/nsteps (default is False). |
False
|
min_t
|
Float
|
minimum time value defaults to 0. |
0
|
padding
|
Float
|
padding time value defaults to 0. |
0
|
dilation
|
Float
|
dilation time value defaults to 0 ie the number of replicates. |
0
|
direction
|
Optional[str]
|
TimeDirection to synchronize the schedule with. If the schedule is defined with a different direction, this parameter allows to flip the direction to match the specified one (default is None). |
UNIFIED
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
'cpu'
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 |
|
discretize(nsteps=None, schedule=None, device=None)
Discretize the time schedule into a list of time deltas.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optioanl[int]
|
Number of time steps. If None, uses the value from initialization. |
None
|
schedule
|
Optional[Tensor]
|
Time scheudle if None will generate it with generate_schedule. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Returns:
Name | Type | Description |
---|---|---|
Tensor |
Tensor
|
A tensor of time deltas. |
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
DiscreteInferenceSchedule
Bases: InferenceSchedule
A base class for discrete time inference schedules.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
discretize(nsteps=None, device=None)
Discretize the time schedule into a list of time deltas.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optioanl[int]
|
Number of time steps. If None, uses the value from initialization. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Returns:
Name | Type | Description |
---|---|---|
Tensor |
Tensor
|
A tensor of time deltas. |
Source code in bionemo/moco/schedules/inference_time_schedules.py
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 |
|
DiscreteLinearInferenceSchedule
Bases: DiscreteInferenceSchedule
A linear time schedule for discrete time inference.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
__init__(nsteps, min_t=0, padding=0, dilation=0, direction=TimeDirection.UNIFIED, device='cpu')
Initialize the DiscreteLinearInferenceSchedule.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
int
|
Number of time steps. |
required |
min_t
|
Float
|
minimum time value defaults to 0. |
0
|
padding
|
Float
|
padding time value defaults to 0. |
0
|
dilation
|
Float
|
dilation time value defaults to 0 ie the number of replicates. |
0
|
direction
|
Optional[str]
|
TimeDirection to synchronize the schedule with. If the schedule is defined with a different direction, this parameter allows to flip the direction to match the specified one (default is None). |
UNIFIED
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
'cpu'
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 |
|
generate_schedule(nsteps=None, device=None)
Generate the linear time schedule as a tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optional[int]
|
Number of time steps. If None uses the value from initialization. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Returns:
Name | Type | Description |
---|---|---|
Tensor |
Tensor
|
A tensor of time steps. |
Tensor |
Tensor
|
A tensor of time steps. |
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
EntropicInferenceSchedule
Bases: ContinuousInferenceSchedule
Generates an entropic time schedule.
It remapping time based on the remaps cumulative information gain provided by a predictor function.
Source code in bionemo/moco/schedules/inference_time_schedules.py
463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 |
|
__init__(predictor, x_0_sampler, x_1_sampler, nsteps, n_approx_entropy_points=100, batch_size=128, inclusive_end=False, min_t=0, direction=TimeDirection.UNIFIED, device='cpu', generator=None)
Inspired by the work from Dejan Stancevic, Florian Handke, & Luca Ambrogioni. (2025).
Entropic Time Schedulers for Generative Diffusion Models. https://arxiv.org/abs/2504.13612
This entropy rate is used to create an optimized, data-dependent time-stepping schedule for the generative process.
Approach benefits: - Can improve inference performance - Prevents oversampling of less-informative time steps - Prevents undersampling of critical time windows - Easily adapted into any model architecture leveraging flow-matching - Sample-eficient way to generate data
Parameters:
Name | Type | Description | Default |
---|---|---|---|
predictor
|
Callable[[Tensor, Tensor], Tensor]
|
A callable (e.g., a function
or functools.partial) that takes a time tensor |
required |
x_0_sampler
|
Callable[[int], Tensor]
|
A function that takes a batch size and returns a tensor of samples from the initial distribution p0. |
required |
x_1_sampler
|
Callable[[int], Tensor]
|
A function that takes a batch size and returns a tensor of samples from the target distribution p1. |
required |
nsteps
|
int
|
The final number of time steps for the inference schedule. |
required |
n_approx_entropy_points
|
int
|
The number of points used to approximate the cumulative entropy curve. Higher is more accurate but slower. |
100
|
batch_size
|
int
|
Batch size for calculating divergence at each time step. |
128
|
inclusive_end
|
bool
|
If True, include 1.0, otherwise end just before. |
False
|
min_t
|
Float
|
The minimum time value for the schedule. |
0
|
direction
|
Union[TimeDirection, str]
|
'UNIFIED' for forward (0->1) or 'DIFFUSION' for reverse (1->0). |
UNIFIED
|
device
|
Union[str, device]
|
The device for computation. |
'cpu'
|
generator
|
Optional[Generator]
|
A PyTorch generator for reproducible random number generation. |
None
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 |
|
generate_schedule(nsteps=None, device=None)
Generates the entropic time schedule.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
InferenceSchedule
Bases: ABC
A base class for inference time schedules.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
__init__(nsteps, min_t=0, padding=0, dilation=0, direction=TimeDirection.UNIFIED, device='cpu')
Initialize the InferenceSchedule.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
int
|
Number of time steps. |
required |
min_t
|
Float
|
minimum time value defaults to 0. |
0
|
padding
|
Float
|
padding time value defaults to 0. |
0
|
dilation
|
Float
|
dilation time value defaults to 0 ie the number of replicates. |
0
|
direction
|
Optional[str]
|
TimeDirection to synchronize the schedule with. If the schedule is defined with a different direction, this parameter allows to flip the direction to match the specified one (default is None). |
UNIFIED
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
'cpu'
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
generate_schedule(nsteps=None, device=None)
abstractmethod
Generate the time schedule as a tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optioanl[int]
|
Number of time steps. If None, uses the value from initialization. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
58 59 60 61 62 63 64 65 66 67 68 |
|
pad_time(n_samples, scalar_time, device=None)
Creates a tensor of shape (n_samples,) filled with a scalar time value.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
n_samples
|
int
|
The desired dimension of the output tensor. |
required |
scalar_time
|
Float
|
The scalar time value to fill the tensor with. |
required |
device
|
Optional[Union[str, device]]
|
The device to place the tensor on. Defaults to None, which uses the default device. |
None
|
Returns:
Name | Type | Description |
---|---|---|
Tensor |
Tensor
|
A tensor of shape (n_samples,) filled with the scalar time value. |
Source code in bionemo/moco/schedules/inference_time_schedules.py
70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 |
|
LinearInferenceSchedule
Bases: ContinuousInferenceSchedule
A linear time schedule for continuous time inference.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
__init__(nsteps, inclusive_end=False, min_t=0, padding=0, dilation=0, direction=TimeDirection.UNIFIED, device='cpu')
Initialize the LinearInferenceSchedule.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
int
|
Number of time steps. |
required |
inclusive_end
|
bool
|
If True, include the end value (1.0) in the schedule otherwise ends at 1.0-1/nsteps (default is False). |
False
|
min_t
|
Float
|
minimum time value defaults to 0. |
0
|
padding
|
Float
|
padding time value defaults to 0. |
0
|
dilation
|
Float
|
dilation time value defaults to 0 ie the number of replicates. |
0
|
direction
|
Optional[str]
|
TimeDirection to synchronize the schedule with. If the schedule is defined with a different direction, this parameter allows to flip the direction to match the specified one (default is None). |
UNIFIED
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
'cpu'
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 |
|
generate_schedule(nsteps=None, device=None)
Generate the linear time schedule as a tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optional[int]
|
Number of time steps. If None uses the value from initialization. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Returns:
Name | Type | Description |
---|---|---|
Tensor |
Tensor
|
A tensor of time steps. |
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
LogInferenceSchedule
Bases: ContinuousInferenceSchedule
A log time schedule for inference, where time steps are generated by taking the logarithm of a uniform schedule.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
__init__(nsteps, inclusive_end=False, min_t=0, padding=0, dilation=0, exponent=-2.0, direction=TimeDirection.UNIFIED, device='cpu')
Initialize the LogInferenceSchedule.
Returns a log space time schedule.
Which for 100 steps with default parameters is
tensor([0.0000, 0.0455, 0.0889, 0.1303, 0.1699, 0.2077, 0.2439, 0.2783, 0.3113, 0.3427, 0.3728, 0.4015, 0.4288, 0.4550, 0.4800, 0.5039, 0.5266, 0.5484, 0.5692, 0.5890, 0.6080, 0.6261, 0.6434, 0.6599, 0.6756, 0.6907, 0.7051, 0.7188, 0.7319, 0.7444, 0.7564, 0.7678, 0.7787, 0.7891, 0.7991, 0.8086, 0.8176, 0.8263, 0.8346, 0.8425, 0.8500, 0.8572, 0.8641, 0.8707, 0.8769, 0.8829, 0.8887, 0.8941, 0.8993, 0.9043, 0.9091, 0.9136, 0.9180, 0.9221, 0.9261, 0.9299, 0.9335, 0.9369, 0.9402, 0.9434, 0.9464, 0.9492, 0.9520, 0.9546, 0.9571, 0.9595, 0.9618, 0.9639, 0.9660, 0.9680, 0.9699, 0.9717, 0.9734, 0.9751, 0.9767, 0.9782, 0.9796, 0.9810, 0.9823, 0.9835, 0.9847, 0.9859, 0.9870, 0.9880, 0.9890, 0.9899, 0.9909, 0.9917, 0.9925, 0.9933, 0.9941, 0.9948, 0.9955, 0.9962, 0.9968, 0.9974, 0.9980, 0.9985, 0.9990, 0.9995])
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
int
|
Number of time steps. |
required |
inclusive_end
|
bool
|
If True, include the end value (1.0) in the schedule otherwise ends at <1.0 (default is False). |
False
|
min_t
|
Float
|
minimum time value defaults to 0. |
0
|
padding
|
Float
|
padding time value defaults to 0. |
0
|
dilation
|
Float
|
dilation time value defaults to 0 ie the number of replicates. |
0
|
exponent
|
Float
|
log space exponent parameter defaults to -2.0. The lower number the more aggressive the acceleration of 0 to 0.9 will be thus having more steps from 0.9 to 1.0. |
-2.0
|
direction
|
Optional[str]
|
TimeDirection to synchronize the schedule with. If the schedule is defined with a different direction, this parameter allows to flip the direction to match the specified one (default is None). |
UNIFIED
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
'cpu'
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
generate_schedule(nsteps=None, device=None)
Generate the log time schedule as a tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optional[int]
|
Number of time steps. If None uses the value from initialization. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
PowerInferenceSchedule
Bases: ContinuousInferenceSchedule
A power time schedule for inference, where time steps are generated by raising a uniform schedule to a specified power.
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
__init__(nsteps, inclusive_end=False, min_t=0, padding=0, dilation=0, exponent=1.0, direction=TimeDirection.UNIFIED, device='cpu')
Initialize the PowerInferenceSchedule.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
int
|
Number of time steps. |
required |
inclusive_end
|
bool
|
If True, include the end value (1.0) in the schedule otherwise ends at <1.0 (default is False). |
False
|
min_t
|
Float
|
minimum time value defaults to 0. |
0
|
padding
|
Float
|
padding time value defaults to 0. |
0
|
dilation
|
Float
|
dilation time value defaults to 0 ie the number of replicates. |
0
|
exponent
|
Float
|
Power parameter defaults to 1.0. |
1.0
|
direction
|
Optional[str]
|
TimeDirection to synchronize the schedule with. If the schedule is defined with a different direction, this parameter allows to flip the direction to match the specified one (default is None). |
UNIFIED
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
'cpu'
|
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|
generate_schedule(nsteps=None, device=None)
Generate the power time schedule as a tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
nsteps
|
Optional[int]
|
Number of time steps. If None uses the value from initialization. |
None
|
device
|
Optional[str]
|
Device to place the schedule on (default is "cpu"). |
None
|
Returns:
Name | Type | Description |
---|---|---|
Tensor |
Tensor
|
A tensor of time steps. |
Tensor |
Tensor
|
A tensor of time steps. |
Source code in bionemo/moco/schedules/inference_time_schedules.py
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 |
|