File size: 10,361 Bytes
617fe56
5549314
8561a1f
617fe56
8561a1f
 
 
617fe56
8561a1f
5549314
8561a1f
 
 
 
617fe56
 
 
850b9a2
 
 
617fe56
850b9a2
617fe56
850b9a2
 
 
 
 
 
 
8561a1f
faa9951
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
617fe56
 
 
 
 
 
 
 
 
 
8561a1f
 
 
617fe56
8561a1f
850b9a2
617fe56
 
 
 
850b9a2
617fe56
850b9a2
617fe56
 
 
 
 
850b9a2
 
 
8561a1f
 
617fe56
 
 
8561a1f
617fe56
 
 
 
 
8561a1f
 
 
 
 
 
 
 
 
617fe56
 
 
 
 
 
 
 
 
 
 
 
8561a1f
 
 
 
0ff7c3d
 
 
 
 
 
 
8561a1f
 
 
 
 
 
617fe56
 
 
 
 
 
 
 
 
8561a1f
 
617fe56
 
 
 
 
 
 
8561a1f
 
 
617fe56
 
 
 
8561a1f
 
617fe56
 
 
 
 
 
 
8561a1f
 
 
617fe56
 
 
8561a1f
617fe56
 
 
 
 
 
 
 
 
 
 
8561a1f
617fe56
 
 
 
 
 
 
 
 
 
 
8561a1f
 
617fe56
8561a1f
0ff7c3d
8561a1f
 
 
a416a9d
8561a1f
851184a
 
 
 
a416a9d
 
cdf5490
20706dd
cdf5490
 
 
 
 
 
 
 
 
8561a1f
617fe56
cdf5490
8561a1f
851184a
151f328
851184a
 
151f328
851184a
5549314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8561a1f
617fe56
 
 
 
 
 
 
 
 
8561a1f
6aad619
 
 
 
 
 
cdf5490
20706dd
462e28d
 
 
 
8561a1f
e93b0fd
9410275
 
8561a1f
 
 
5c4e4bf
8561a1f
 
 
617fe56
8561a1f
617fe56
 
 
cdf5490
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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
import math
import os
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 transformers import PretrainedConfig

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):
    """
    This LoRA implementation was inspired by  https://github.com/cccntu/minLoRA

    The MIT License (MIT) Copyright (c) 2020 Andrej Karpathy

    Permission is hereby granted, free of charge, to any person obtaining a copy of this software
    and associated documentation files (the "Software"), to deal in the Software without restriction,
    including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense,
    and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so,
    subject to the following conditions:

    The above copyright notice and this permission notice shall be included in all copies or substantial
    portions of the Software.

    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT
    LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
    IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
    WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
    SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
    """
    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)

    @property
    def current_task(self):
        return self._current_task

    @current_task.setter
    def current_task(self, task: Union[None, int]):
        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.current_task = task_idx


class BertLoRA(BertPreTrainedModel):
    def __init__(self, config: JinaBertConfig, bert: Optional[BertModel] = None, add_pooling_layer=True):
        super().__init__(config)
        if bert is None:
            self.bert = BertModel(config, add_pooling_layer=add_pooling_layer)
        else:
            self.bert = bert
        self._num_adaptions = config.num_loras
        self._register_lora(self._num_adaptions)
        self.main_params_trainable = False
        self._task_idx = None
        self.current_task = 0

    @property
    def main_params_trainable(self):
        return self._main_params_trainable

    @main_params_trainable.setter
    def main_params_trainable(self, val):
        self._main_params_trainable = val
        for name, param in super().named_parameters():
            if "lora" not in name:
                param.requires_grad_(val)

    @classmethod
    def from_bert(cls, *args, **kwargs):
        bert = BertModel.from_pretrained(*args, **kwargs)
        config = JinaBertConfig.from_pretrained(*args, **kwargs)
        return cls(config, bert=bert)

    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
        *model_args,
        config: Optional[Union[PretrainedConfig, str, os.PathLike]] = None,
        cache_dir: Optional[Union[str, os.PathLike]] = None,
        ignore_mismatched_sizes: bool = False,
        force_download: bool = False,
        local_files_only: bool = False,
        token: Optional[Union[str, bool]] = None,
        revision: str = "main",
        use_safetensors: bool = None,
        **kwargs,
    ):
        # TODO: choose between from_bert and super().from_pretrained
        return cls.from_bert(pretrained_model_name_or_path)

    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,
            )
        )

    @property
    def current_task(self):
        return self._task_idx

    @current_task.setter
    def current_task(self, task_idx: Union[None, int]):
        assert task_idx is None or 0 <= task_idx < self._num_adaptions
        if self._task_idx != task_idx:
            self._task_idx = task_idx
            self.apply(
                partial(LoRAParametrization.select_task_for_layer, task_idx=task_idx)
            )

    def forward(self, *args, current_task: Union[None, int] = -1, **kwargs):
        if current_task is None or current_task >= 0:
            self.current_task = current_task
        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 or self.main_params_trainable:
                yield name, param