|
| 1 | +import torch |
| 2 | +from torch.optim.optimizer import Optimizer |
| 3 | + |
| 4 | +from pytorch_optimizer.base_optimizer import BaseOptimizer |
| 5 | +from pytorch_optimizer.types import CLOSURE, DEFAULTS, LOSS, PARAMETERS |
| 6 | +from pytorch_optimizer.utils import matrix_power |
| 7 | + |
| 8 | + |
| 9 | +class Shampoo(Optimizer, BaseOptimizer): |
| 10 | + """ |
| 11 | + Reference : https://github.com/moskomule/shampoo.pytorch/blob/master/shampoo.py |
| 12 | + Example : |
| 13 | + from pytorch_optimizer import Shampoo |
| 14 | + ... |
| 15 | + model = YourModel() |
| 16 | + optimizer = Shampoo(model.parameters()) |
| 17 | + ... |
| 18 | + for input, output in data: |
| 19 | + optimizer.zero_grad() |
| 20 | + loss = loss_function(output, model(input)) |
| 21 | + loss.backward() |
| 22 | + optimizer.step() |
| 23 | + """ |
| 24 | + |
| 25 | + def __init__( |
| 26 | + self, |
| 27 | + params: PARAMETERS, |
| 28 | + lr: float = 1e-3, |
| 29 | + momentum: float = 0.0, |
| 30 | + weight_decay: float = 0.0, |
| 31 | + update_freq: int = 1, |
| 32 | + eps: float = 1e-4, |
| 33 | + ): |
| 34 | + """Shampoo optimizer |
| 35 | + :param params: PARAMETERS. iterable of parameters to optimize or dicts defining parameter groups |
| 36 | + :param lr: float. learning rate |
| 37 | + :param momentum: float. momentum |
| 38 | + :param weight_decay: float. weight decay (L2 penalty) |
| 39 | + :param update_freq: int. update frequency to compute inverse |
| 40 | + :param eps: float. term added to the denominator to improve numerical stability |
| 41 | + """ |
| 42 | + self.lr = lr |
| 43 | + self.momentum = momentum |
| 44 | + self.weight_decay = weight_decay |
| 45 | + self.update_freq = update_freq |
| 46 | + self.eps = eps |
| 47 | + |
| 48 | + self.validate_parameters() |
| 49 | + |
| 50 | + defaults: DEFAULTS = dict( |
| 51 | + lr=lr, |
| 52 | + momentum=momentum, |
| 53 | + weight_decay=weight_decay, |
| 54 | + update_freq=update_freq, |
| 55 | + eps=eps, |
| 56 | + ) |
| 57 | + super().__init__(params, defaults) |
| 58 | + |
| 59 | + def validate_parameters(self): |
| 60 | + self.validate_learning_rate(self.lr) |
| 61 | + self.validate_momentum(self.momentum) |
| 62 | + self.validate_weight_decay(self.weight_decay) |
| 63 | + self.validate_update_frequency(self.update_freq) |
| 64 | + self.validate_epsilon(self.eps) |
| 65 | + |
| 66 | + @torch.no_grad() |
| 67 | + def reset(self): |
| 68 | + for group in self.param_groups: |
| 69 | + for p in group['params']: |
| 70 | + state = self.state[p] |
| 71 | + |
| 72 | + state['step'] = 0 |
| 73 | + |
| 74 | + @torch.no_grad() |
| 75 | + def step(self, closure: CLOSURE = None) -> LOSS: |
| 76 | + loss: LOSS = None |
| 77 | + if closure is not None: |
| 78 | + with torch.enable_grad(): |
| 79 | + loss = closure() |
| 80 | + |
| 81 | + for group in self.param_groups: |
| 82 | + for p in group['params']: |
| 83 | + if p.grad is None: |
| 84 | + continue |
| 85 | + |
| 86 | + grad = p.grad |
| 87 | + if grad.is_sparse: |
| 88 | + raise RuntimeError('Shampoo does not support sparse gradients') |
| 89 | + |
| 90 | + momentum = group['momentum'] |
| 91 | + state = self.state[p] |
| 92 | + if len(state) == 0: |
| 93 | + state['step'] = 0 |
| 94 | + |
| 95 | + if momentum > 0.0: |
| 96 | + state['momentum_buffer'] = grad.clone() |
| 97 | + |
| 98 | + # pre-condition matrices |
| 99 | + for dim_id, dim in enumerate(grad.size()): |
| 100 | + state[f'pre_cond_{dim_id}'] = group['eps'] * torch.eye(dim, out=grad.new(dim, dim)) |
| 101 | + state[f'inv_pre_cond_{dim_id}'] = grad.new(dim, dim).zero_() |
| 102 | + |
| 103 | + if momentum > 0.0: |
| 104 | + grad.mul_(1.0 - momentum).add_(state['momentum_buffer'], alpha=momentum) |
| 105 | + |
| 106 | + weight_decay = group['weight_decay'] |
| 107 | + if weight_decay > 0.0: |
| 108 | + grad.add_(p, alpha=weight_decay) |
| 109 | + |
| 110 | + order: int = grad.ndimension() |
| 111 | + original_size: int = grad.size() |
| 112 | + for dim_id, dim in enumerate(grad.size()): |
| 113 | + pre_cond = state[f'pre_cond_{dim_id}'] |
| 114 | + inv_pre_cond = state[f'inv_pre_cond_{dim_id}'] |
| 115 | + |
| 116 | + grad = grad.transpose_(0, dim_id).contiguous() |
| 117 | + transposed_size = grad.size() |
| 118 | + |
| 119 | + grad = grad.view(dim, -1) |
| 120 | + |
| 121 | + grad_t = grad.t() |
| 122 | + pre_cond.add_(grad @ grad_t) |
| 123 | + if state['step'] % group['update_freq'] == 0: |
| 124 | + inv_pre_cond.copy_(matrix_power(pre_cond, -1 / order)) |
| 125 | + |
| 126 | + if dim_id == order - 1: |
| 127 | + grad = grad_t @ inv_pre_cond |
| 128 | + grad = grad.view(original_size) |
| 129 | + else: |
| 130 | + grad = inv_pre_cond @ grad |
| 131 | + grad = grad.view(transposed_size) |
| 132 | + |
| 133 | + state['step'] += 1 |
| 134 | + state['momentum_buffer'] = grad |
| 135 | + |
| 136 | + p.add_(grad, alpha=-group['lr']) |
| 137 | + |
| 138 | + return loss |
0 commit comments