torchebm.samplers.langevin_dynamics ¶
Langevin Dynamics Sampler Module.
LangevinDynamics ¶
Bases: BaseSampler
Langevin Dynamics sampler.
Update rule:
Called as sampler(model, x, n_steps=...), the model giving the score \(-\nabla_x U\) that the configured integrator drifts along; see BaseSampler.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
model | Optional[BaseModel] | Deprecated. Energy-based model sampled by | None |
step_size | Union[float, BaseScheduler] | Step size for gradient descent. Float or | 0.001 |
noise_scale | Union[float, BaseScheduler] | Scale of Gaussian noise injection. Float or | 1.0 |
decay | float | Damping coefficient (not supported). | 0.0 |
clamp | Optional[Tuple[float, float]] | Optional (min, max) bounds applied to the state after every step. Standard stabilization for image-space EBMs (e.g. [-1, 1]). | None |
dtype | dtype | Data type for computations. | float32 |
device | Optional[Union[str, device]] | Device for computations. | None |
integrator | Union[str, BaseSDERungeKuttaIntegrator, None] | SDE integrator used for the update. | None |
Example
SGLD as in JEM (Grathwohl et al., 2020): step 1 and noise std 0.01 (noise_scale times \(\sqrt{2\eta}\)), 20 steps from U(-1, 1) on CIFAR-10 images:
Source code in torchebm/samplers/langevin_dynamics.py
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 | |
forward(model, x, *, n_steps=50, generator=None, model_kwargs=None, return_trajectory=False) ¶
Run n_steps Langevin updates of model from x; see BaseSampler.forward for the arguments.
Source code in torchebm/samplers/langevin_dynamics.py
sample(x=None, dim=None, n_steps=100, n_samples=1, thin=1, return_trajectory=False, return_diagnostics=False, reset_schedulers=True, *, y=None, model_kwargs=None, generator=None) ¶
Generate samples via Langevin dynamics.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x | Optional[Tensor] | Initial state. If | None |
dim | Optional[Union[int, Tuple[int, ...]]] | State dimension (int) or shape (tuple), used when | None |
n_steps | int | Number of MCMC steps to perform. | 100 |
n_samples | int | Number of parallel chains to generate. | 1 |
thin | int | Keep every | 1 |
return_trajectory | bool | If True, return the full kept trajectory of shape | False |
return_diagnostics | bool | If True, also return a dict with keys | False |
reset_schedulers | bool | If True (default), reset registered schedulers. | True |
y | Optional[Tensor] | Optional conditioning tensor forwarded to the model; shorthand for | None |
model_kwargs | Optional[Dict[str, Any]] | Conditioning arguments (e.g. class labels) forwarded to the model at every step. Normalized to the sampler device once at entry; | None |
generator | Optional[Generator] | RNG for the initial state and the per-step Langevin noise; the global RNG when | None |
Returns:
| Type | Description |
|---|---|
Union[Tensor, Tuple[Tensor, Dict[str, Tensor]]] | Sample tensor (or trajectory if |
Union[Tensor, Tuple[Tensor, Dict[str, Tensor]]] | optionally paired with the diagnostics dict. |
Raises:
| Type | Description |
|---|---|
ValueError | If |
Source code in torchebm/samplers/langevin_dynamics.py
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | |