Feature Extraction
Transformers
Safetensors
custom_code
File size: 6,921 Bytes
2f947ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
from typing import Optional, Callable, Union, Tuple, Any, Dict, NamedTuple

import torch
from torch import nn

from timm.models import create_model, VisionTransformer

from .enable_cpe_support import enable_cpe
from .input_conditioner import InputConditioner
# Register extra models
from . import extra_timm_models
from .adaptor_base import AdaptorBase, RadioOutput, AdaptorInput
from . import eradio_model
from .enable_spectral_reparam import configure_spectral_reparam_from_args


class Resolution(NamedTuple):
    height: int
    width: int


class RADIOModel(nn.Module):
    def __init__(
        self,
        model: nn.Module,
        input_conditioner: InputConditioner,
        patch_size: int,
        max_resolution: int,
        preferred_resolution: Resolution,
        summary_idxs: Optional[torch.Tensor] = None,
        window_size: int = None,
        adaptors: Dict[str, AdaptorBase] = None,
    ):
        super().__init__()

        self.model = model
        self.input_conditioner = input_conditioner
        if summary_idxs is not None:
            self.register_buffer('summary_idxs', summary_idxs)
        else:
            self.summary_idxs = None

        self._preferred_resolution = preferred_resolution
        self._patch_size = patch_size
        self._max_resolution = max_resolution
        self._window_size = window_size

        adaptors = adaptors or dict()
        self.adaptors = nn.ModuleDict(adaptors)

    @property
    def num_summary_tokens(self) -> int:
        patch_gen = getattr(self.model, "patch_generator", None)
        if patch_gen is not None:
            return patch_gen.num_skip
        elif self.model.global_pool == 'avg':
            return 0
        return 1

    @property
    def patch_size(self) -> int:
        return self._patch_size

    @property
    def max_resolution(self) -> int:
        return self._max_resolution

    @property
    def preferred_resolution(self) -> Resolution:
        return self._preferred_resolution

    @property
    def window_size(self) -> int:
        return self._window_size

    @property
    def min_resolution_step(self) -> int:
        res = self.patch_size
        if self.window_size is not None:
            res *= self.window_size
        return res

    def make_preprocessor_external(self) -> Callable[[torch.Tensor], torch.Tensor]:
        ret = self.input_conditioner
        self.input_conditioner = nn.Identity()
        return ret

    def get_nearest_supported_resolution(self, height: int, width: int) -> Resolution:
        height = int(round(height / self.min_resolution_step) * self.min_resolution_step)
        width = int(round(width / self.min_resolution_step) * self.min_resolution_step)

        height = max(height, self.min_resolution_step)
        width = max(width, self.min_resolution_step)

        return Resolution(height=height, width=width)

    def switch_to_deploy(self):
        fn = getattr(self.model, 'switch_to_deploy', None)
        if fn is not None:
            fn()

    def forward(self, x: torch.Tensor) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        x = self.input_conditioner(x)
        y = self.model.forward_features(x)

        if isinstance(self.model, VisionTransformer):
            patch_gen = getattr(self.model, "patch_generator", None)
            if patch_gen is not None:
                all_summary = y[:, : patch_gen.num_cls_tokens]
                if self.summary_idxs is not None:
                    bb_summary = all_summary[:, self.summary_idxs]
                else:
                    bb_summary = all_summary
                all_feat = y[:, patch_gen.num_skip :]
            elif self.model.global_pool == "avg":
                all_summary = y[:, self.model.num_prefix_tokens :].mean(dim=1)
                bb_summary = all_summary
                all_feat = y
            else:
                all_summary = y[:, 0]
                bb_summary = all_summary
                all_feat = y[:, 1:]
        elif isinstance(self.model, eradio_model.FasterViT):
            _, f = y
            all_feat = f.flatten(2).transpose(1, 2)
            all_summary = all_feat.mean(dim=1)
            bb_summary = all_summary
        elif isinstance(y, (list, tuple)):
            all_summary, all_feat = y
            bb_summary = all_summary
        else:
            raise ValueError("Unsupported model type")

        all_feat = all_feat.float()
        ret = RadioOutput(bb_summary.flatten(1), all_feat).to(torch.float32)
        if self.adaptors:
            ret = dict(backbone=ret)
            for name, adaptor in self.adaptors.items():
                if all_summary.ndim == 3:
                    summary = all_summary[:, adaptor.head_idx]
                else:
                    summary = all_summary
                ada_input = AdaptorInput(images=x, summary=summary.float(), features=all_feat)
                v = adaptor(ada_input).to(torch.float32)
                ret[name] = v

        return ret


def create_model_from_args(args) -> nn.Module:
    in_chans = 3
    if args.in_chans is not None:
        in_chans = args.in_chans
    elif args.input_size is not None:
        in_chans = args.input_size[0]

    # Skip weight initialization unless it's explicitly requested.
    weight_init = args.model_kwargs.pop("weight_init", "skip")

    model = create_model(
        args.model,
        pretrained=args.pretrained,
        in_chans=in_chans,
        num_classes=args.num_classes,
        drop_rate=args.drop,
        drop_path_rate=args.drop_path,
        drop_block_rate=args.drop_block,
        global_pool=args.gp,
        bn_momentum=args.bn_momentum,
        bn_eps=args.bn_eps,
        scriptable=args.torchscript,
        checkpoint_path=args.initial_checkpoint,
        weight_init=weight_init,
        **args.model_kwargs,
    )

    if hasattr(model, 'norm') and not getattr(args, 'model_norm', False):
        model.norm = nn.Identity()

    model.head = nn.Identity()

    assert (
        not args.cls_token_per_teacher or args.cpe_max_size is not None
    ), "CPE must be enabled for multiple CLS tokens!"

    if args.cpe_max_size is not None:
        enable_cpe(
            model,
            args.cpe_max_size,
            num_cls_tokens=len(args.teachers) if args.cls_token_per_teacher else 1,
            register_multiple=args.register_multiple,
        )

    if args.spectral_reparam:
        configure_spectral_reparam_from_args(model, args)

    return model