| 214 | """ |
| 215 | |
| 216 | def __init__(self, |
| 217 | lr=0.1, |
| 218 | momentum=0, |
| 219 | dampening=0, |
| 220 | weight_decay=0, |
| 221 | nesterov=False, |
| 222 | dtype=tensor.float32): |
| 223 | super(SGD, self).__init__(lr, dtype) |
| 224 | |
| 225 | # init momentum |
| 226 | if type(momentum) == float or type(momentum) == int: |
| 227 | if momentum < 0.0: |
| 228 | raise ValueError("Invalid momentum value: {}".format(momentum)) |
| 229 | self.momentum = Constant(momentum) |
| 230 | elif isinstance(momentum, DecayScheduler): |
| 231 | self.momentum = momentum |
| 232 | momentum = momentum.init_value |
| 233 | else: |
| 234 | raise TypeError("Wrong momentum type") |
| 235 | self.mom_value = self.momentum(self.step_counter).as_type(self.dtype) |
| 236 | |
| 237 | # init dampening |
| 238 | if type(dampening) == float or type(dampening) == int: |
| 239 | self.dampening = Constant(dampening) |
| 240 | elif isinstance(dampening, DecayScheduler): |
| 241 | self.dampening = dampening |
| 242 | dampening = dampening.init_value |
| 243 | else: |
| 244 | raise TypeError("Wrong dampening type") |
| 245 | self.dam_value = self.dampening(self.step_counter).as_type(self.dtype) |
| 246 | |
| 247 | # init weight_decay |
| 248 | if type(weight_decay) == float or type(weight_decay) == int: |
| 249 | if weight_decay < 0.0: |
| 250 | raise ValueError( |
| 251 | "Invalid weight_decay value: {}".format(weight_decay)) |
| 252 | self.weight_decay = Constant(weight_decay) |
| 253 | elif isinstance(weight_decay, DecayScheduler): |
| 254 | self.weight_decay = weight_decay |
| 255 | else: |
| 256 | raise TypeError("Wrong weight_decay type") |
| 257 | self.decay_value = self.weight_decay(self.step_counter).as_type( |
| 258 | self.dtype) |
| 259 | |
| 260 | # init other params |
| 261 | self.nesterov = nesterov |
| 262 | self.moments = dict() |
| 263 | |
| 264 | # check value |
| 265 | if nesterov and (momentum <= 0 or dampening != 0): |
| 266 | raise ValueError( |
| 267 | "Nesterov momentum requires a momentum and zero dampening") |
| 268 | |
| 269 | def apply(self, param_name, param_value, param_grad): |
| 270 | """Performs a single optimization step. |