|
| 1 | +import random |
| 2 | +from copy import deepcopy |
| 3 | +from typing import Iterable, List |
| 4 | + |
| 5 | +import numpy as np |
| 6 | +import torch |
| 7 | +from torch import nn |
| 8 | +from torch.optim.optimizer import Optimizer |
| 9 | + |
| 10 | + |
| 11 | +class PCGrad: |
| 12 | + """ |
| 13 | + Reference : https://github.com/WeiChengTseng/Pytorch-PCGrad |
| 14 | + Example : |
| 15 | + from pytorch_optimizer import AdamP, PCGrad |
| 16 | + ... |
| 17 | + model = YourModel() |
| 18 | + optimizer = PCGrad(AdamP(model.parameters())) |
| 19 | +
|
| 20 | + loss_1, loss_2 = nn.L1Loss(), nn.MSELoss() |
| 21 | + ... |
| 22 | + for input, output in data: |
| 23 | + optimizer.zero_grad() |
| 24 | + loss1, loss2 = loss1_fn(y_pred, output), loss2_fn(y_pred, output) |
| 25 | + optimizer.pc_backward([loss1, loss2]) |
| 26 | + optimizer.step() |
| 27 | + """ |
| 28 | + |
| 29 | + def __init__(self, optimizer: Optimizer, reduction: str = 'mean'): |
| 30 | + self.optimizer = optimizer |
| 31 | + self.reduction = reduction |
| 32 | + |
| 33 | + def check_valid_parameters(self): |
| 34 | + if self.reduction not in ('mean', 'sum'): |
| 35 | + raise ValueError(f'invalid reduction : {self.reduction}') |
| 36 | + |
| 37 | + @staticmethod |
| 38 | + def flatten_grad(grads) -> torch.Tensor: |
| 39 | + return torch.cat([g.flatten() for g in grads]) |
| 40 | + |
| 41 | + @staticmethod |
| 42 | + def un_flatten_grad(grads, shapes) -> List[torch.Tensor]: |
| 43 | + un_flatten_grad = [] |
| 44 | + idx: int = 0 |
| 45 | + for shape in shapes: |
| 46 | + length = np.prod(shape) |
| 47 | + un_flatten_grad.append(grads[idx : idx + length].view(shape).clone()) |
| 48 | + idx += length |
| 49 | + return un_flatten_grad |
| 50 | + |
| 51 | + def zero_grad(self): |
| 52 | + return self.optimizer.zero_grad(set_to_none=True) |
| 53 | + |
| 54 | + def step(self): |
| 55 | + return self.optimizer.step() |
| 56 | + |
| 57 | + def set_grad(self, grads): |
| 58 | + idx: int = 0 |
| 59 | + for group in self.optimizer.param_groups: |
| 60 | + for p in group['params']: |
| 61 | + p.grad = grads[idx] |
| 62 | + idx += 1 |
| 63 | + |
| 64 | + def pc_backward(self, objectives: Iterable[nn.Module]): |
| 65 | + """Calculate the gradient of the parameters |
| 66 | + :param objectives: Iterable[nn.Module]. a list of objectives |
| 67 | + :return: |
| 68 | + """ |
| 69 | + grads, shapes, has_grads = self.pack_grad(objectives) |
| 70 | + pc_grad = self.project_conflicting(grads, has_grads) |
| 71 | + pc_grad = self.un_flatten_grad(pc_grad, shapes[0]) |
| 72 | + self.set_grad(pc_grad) |
| 73 | + |
| 74 | + def project_conflicting(self, grads, has_grads) -> torch.Tensor: |
| 75 | + """ |
| 76 | + :param grads: a list of the gradient of the parameters |
| 77 | + :param has_grads: a list of mask represent whether the parameter has gradient |
| 78 | + :return: |
| 79 | + """ |
| 80 | + shared = torch.stack(has_grads).prod(0).bool() |
| 81 | + |
| 82 | + pc_grad = deepcopy(grads) |
| 83 | + for g_i in pc_grad: |
| 84 | + random.shuffle(grads) |
| 85 | + for g_j in grads: |
| 86 | + g_i_g_j = torch.dot(g_i, g_j) |
| 87 | + if g_i_g_j < 0: |
| 88 | + g_i -= g_i_g_j * g_j / (g_j.norm() ** 2) |
| 89 | + |
| 90 | + merged_grad = torch.zeros_like(grads[0]).to(grads[0].device) |
| 91 | + merged_grad[shared] = torch.stack([g[shared] for g in pc_grad]) |
| 92 | + |
| 93 | + if self.reduction == 'mean': |
| 94 | + merged_grad = merged_grad.mean(dim=0) |
| 95 | + else: # self.reduction == 'sum' |
| 96 | + merged_grad = merged_grad.sum(dim=0) |
| 97 | + |
| 98 | + merged_grad[~shared] = torch.stack([g[~shared] for g in pc_grad]).sum(dim=0) |
| 99 | + |
| 100 | + return merged_grad |
| 101 | + |
| 102 | + def retrieve_grad(self): |
| 103 | + """Get the gradient of the parameters of the network with specific objective |
| 104 | + :return: |
| 105 | + """ |
| 106 | + grad, shape, has_grad = [], [], [] |
| 107 | + for group in self.optimizer.param_groups: |
| 108 | + for p in group['params']: |
| 109 | + if p.grad is None: |
| 110 | + shape.append(p.shape) |
| 111 | + grad.append(torch.zeros_like(p).to(p.device)) |
| 112 | + has_grad.append(torch.zeros_like(p).to(p.device)) |
| 113 | + continue |
| 114 | + |
| 115 | + shape.append(p.grad.shape) |
| 116 | + grad.append(p.grad.clone()) |
| 117 | + has_grad.append(torch.ones_like(p).to(p.device)) |
| 118 | + |
| 119 | + return grad, shape, has_grad |
| 120 | + |
| 121 | + def pack_grad(self, objectives: Iterable[nn.Module]): |
| 122 | + """Pack the gradient of the parameters of the network for each objective |
| 123 | + :param objectives: Iterable[float]. a list of objectives |
| 124 | + :return: |
| 125 | + """ |
| 126 | + grads, shapes, has_grads = [], [], [] |
| 127 | + for objective in objectives: |
| 128 | + self.zero_grad() |
| 129 | + |
| 130 | + objective.backward(retain_graph=True) |
| 131 | + |
| 132 | + grad, shape, has_grad = self.retrieve_grad() |
| 133 | + |
| 134 | + grads.append(self.flatten_grad(grad)) |
| 135 | + has_grads.append(self.flatten_grad(has_grad)) |
| 136 | + shapes.append(shape) |
| 137 | + |
| 138 | + return grads, shapes, has_grads |
0 commit comments