File size: 7,365 Bytes
617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 850b9a2 617fe56 850b9a2 617fe56 850b9a2 8561a1f 617fe56 8561a1f 617fe56 8561a1f 850b9a2 617fe56 850b9a2 617fe56 850b9a2 617fe56 850b9a2 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f 617fe56 8561a1f fabeb13 8561a1f 617fe56 8561a1f 617fe56 8561a1f 5c4e4bf 8561a1f 617fe56 8561a1f 617fe56 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 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 |
import math
from functools import partial
from typing import Iterator, Optional, Tuple, Union
import torch
import torch.nn.utils.parametrize as parametrize
from torch import nn
from torch.nn import Parameter
from .modeling_bert import BertModel, BertPreTrainedModel, JinaBertConfig
def initialized_weights(
shape: Tuple[int], num_adaptions: int, init: str = "kaiming"
) -> torch.Tensor:
weight_data = []
for _ in range(num_adaptions):
new_adaption = torch.zeros(shape)
if init == "kaiming":
nn.init.kaiming_uniform_(new_adaption, a=math.sqrt(5))
elif init == "normal":
nn.init.normal_(new_adaption)
else:
raise NotImplementedError
weight_data.append(new_adaption)
return torch.stack(weight_data, dim=0)
class LoRAParametrization(nn.Module):
def __init__(
self,
fan_in: int,
fan_out: int,
layer_type: str = "linear",
num_adaptions: int = 1,
rank: int = 4,
lora_dropout_p: float = 0.0,
lora_alpha: float = 1,
):
super().__init__()
# if weight is stored as (fan_out, fan_in), the memory layout of A & B follows (W + BA)x
# otherwise, it's x(W + AB). This allows us to tie the weights between linear layers and embeddings
fan_in_fan_out = layer_type == "embedding"
self.swap = (lambda x: (x[1], x[0])) if fan_in_fan_out else (lambda x: x)
if layer_type == "linear":
self.lora_A = nn.Parameter(
initialized_weights((rank, fan_in), num_adaptions, init="kaiming")
)
self.lora_B = nn.Parameter(torch.zeros((num_adaptions, fan_out, rank)))
elif layer_type == "embedding":
self.lora_A = nn.Parameter(torch.zeros((num_adaptions, fan_in, rank)))
self.lora_B = nn.Parameter(
initialized_weights(
(rank, fan_out), num_adaptions=num_adaptions, init="normal"
)
)
else:
raise NotImplementedError
self.lora_alpha, self.rank = lora_alpha, rank
self.scaling = lora_alpha / rank
self.lora_dropout = (
nn.Dropout(p=lora_dropout_p) if lora_dropout_p > 0 else lambda x: x
)
self.dropout_fn = self._dropout if lora_dropout_p > 0 else lambda x: x
self.register_buffer(
"lora_dropout_mask",
torch.ones(self.swap((1, fan_in)), dtype=self.lora_A.dtype),
persistent=False,
)
self.forward_fn = lambda x: x
self.current_task = None
def _dropout(self, A):
# to mimic the original implementation: A @ dropout(x), we do (A * dropout(ones)) @ x
return A * self.lora_dropout(self.lora_dropout_mask)
def lora_forward(self, X):
assert self.current_task is not None
return (
X
+ torch.matmul(
*self.swap(
(
self.lora_B[self.current_task],
self.dropout_fn(self.lora_A[self.current_task]),
)
)
).view(X.shape)
* self.scaling
)
def forward(self, X):
return self.forward_fn(X)
def select_task(self, task=None):
self.current_task = task
if task is None:
self.forward_fn = lambda x: x
else:
self.forward_fn = self.lora_forward
@classmethod
def from_linear(
cls,
layer: nn.Module,
num_adaptions: int = 1,
rank: int = 4,
lora_dropout_p: float = 0.0,
lora_alpha: int = 1,
):
assert isinstance(layer, nn.Linear)
fan_out, fan_in = layer.weight.shape
return cls(
fan_in,
fan_out,
num_adaptions=num_adaptions,
layer_type="linear",
rank=rank,
lora_dropout_p=lora_dropout_p,
lora_alpha=lora_alpha,
)
@classmethod
def from_embedding(
cls, layer, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1
):
assert isinstance(layer, nn.Embedding)
fan_in, fan_out = layer.weight.shape
return cls(
fan_in,
fan_out,
num_adaptions=num_adaptions,
layer_type="embedding",
rank=rank,
lora_dropout_p=lora_dropout_p,
lora_alpha=lora_alpha,
)
@classmethod
def add_to_layer(
cls, layer, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1
):
if isinstance(layer, nn.Linear):
parametrize.register_parametrization(
layer,
"weight",
cls.from_linear(
layer,
num_adaptions=num_adaptions,
rank=rank,
lora_dropout_p=lora_dropout_p,
lora_alpha=lora_alpha,
),
)
elif isinstance(layer, nn.Embedding):
parametrize.register_parametrization(
layer,
"weight",
cls.from_embedding(
layer,
num_adaptions=num_adaptions,
rank=rank,
lora_dropout_p=lora_dropout_p,
lora_alpha=lora_alpha,
),
)
@classmethod
def select_task_for_layer(cls, layer: nn.Module, task_idx: Optional[int] = None):
if isinstance(layer, LoRAParametrization):
layer.select_task(task_idx)
class BertLoRA(BertPreTrainedModel):
def __init__(self, config: JinaBertConfig, add_pooling_layer=True, num_adaptions=1):
super().__init__(config)
self.bert = BertModel(config, add_pooling_layer=add_pooling_layer)
self._register_lora(num_adaptions)
for name, param in super().named_parameters():
if "lora" not in name:
param.requires_grad_(False)
self.select_task(0)
def from_bert(self, *args, num_adaptions=1, **kwargs):
self.bert = BertModel.from_pretrained(*args, **kwargs)
self._register_lora(num_adaptions)
def _register_lora(self, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1):
self.apply(
partial(
LoRAParametrization.add_to_layer,
num_adaptions=num_adaptions,
rank=rank,
lora_dropout_p=lora_dropout_p,
lora_alpha=lora_alpha,
)
)
def select_task(self, task_idx: Union[None, int]):
self.apply(
partial(LoRAParametrization.select_task_for_layer, task_idx=task_idx)
)
def forward(self, *args, **kwargs):
return self.bert(*args, **kwargs)
def parameters(self, recurse: bool = True) -> Iterator[Parameter]:
for _, param in self.named_parameters(recurse=recurse):
yield param
def named_parameters(
self, prefix: str = "", recurse: bool = True, remove_duplicate: bool = True
) -> Iterator[Tuple[str, Parameter]]:
for name, param in super().named_parameters(
prefix=prefix, recurse=recurse, remove_duplicate=remove_duplicate
):
if "lora" in name:
yield name, param
|