| 94 | """ |
| 95 | |
| 96 | def __init__(self, |
| 97 | lr=0.1, |
| 98 | momentum=0, |
| 99 | dampening=0, |
| 100 | weight_decay=0, |
| 101 | nesterov=False, |
| 102 | dtype=tensor.float32): |
| 103 | super(MSSGD, self).__init__(lr, dtype) |
| 104 | |
| 105 | # init momentum |
| 106 | if type(momentum) == float or type(momentum) == int: |
| 107 | if momentum < 0.0: |
| 108 | raise ValueError("Invalid momentum value: {}".format(momentum)) |
| 109 | self.momentum = Constant(momentum) |
| 110 | elif isinstance(momentum, DecayScheduler): |
| 111 | self.momentum = momentum |
| 112 | momentum = momentum.init_value |
| 113 | else: |
| 114 | raise TypeError("Wrong momentum type") |
| 115 | self.mom_value = self.momentum(self.step_counter).as_type(self.dtype) |
| 116 | |
| 117 | # init dampening |
| 118 | if type(dampening) == float or type(dampening) == int: |
| 119 | self.dampening = Constant(dampening) |
| 120 | elif isinstance(dampening, DecayScheduler): |
| 121 | self.dampening = dampening |
| 122 | dampening = dampening.init_value |
| 123 | else: |
| 124 | raise TypeError("Wrong dampening type") |
| 125 | self.dam_value = self.dampening(self.step_counter).as_type(self.dtype) |
| 126 | |
| 127 | # init weight_decay |
| 128 | if type(weight_decay) == float or type(weight_decay) == int: |
| 129 | if weight_decay < 0.0: |
| 130 | raise ValueError( |
| 131 | "Invalid weight_decay value: {}".format(weight_decay)) |
| 132 | self.weight_decay = Constant(weight_decay) |
| 133 | elif isinstance(weight_decay, DecayScheduler): |
| 134 | self.weight_decay = weight_decay |
| 135 | else: |
| 136 | raise TypeError("Wrong weight_decay type") |
| 137 | self.decay_value = self.weight_decay(self.step_counter).as_type( |
| 138 | self.dtype) |
| 139 | |
| 140 | # init other params |
| 141 | self.nesterov = nesterov |
| 142 | self.moments = dict() |
| 143 | |
| 144 | # check value |
| 145 | if nesterov and (momentum <= 0 or dampening != 0): |
| 146 | raise ValueError( |
| 147 | "Nesterov momentum requires a momentum and zero dampening") |
| 148 | |
| 149 | def apply(self, param_name, param_value, param_grad): |
| 150 | """Performs a single optimization step. |