File size: 5,739 Bytes
e4bd7f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
 Copyright (c) 2022, salesforce.com, inc.
 All rights reserved.
 SPDX-License-Identifier: BSD-3-Clause
 For full license text, see the LICENSE_Lavis file in the repo root or https://opensource.org/licenses/BSD-3-Clause
"""

import logging
import torch
from omegaconf import OmegaConf

from bubogpt.common.registry import registry
from bubogpt.models.base_model import BaseModel
from bubogpt.models.blip2 import Blip2Base
from bubogpt.processors.base_processor import BaseProcessor
from bubogpt.models.mm_gpt4 import MMGPT4


__all__ = [
    "load_model",
    "BaseModel",
    "Blip2Base",
    "MMGPT4"
]


def load_model(name, model_type, is_eval=False, device="cpu", checkpoint=None):
    """
    Load supported models.

    To list all available models and types in registry:
    >>> from bubogpt.models import model_zoo
    >>> print(model_zoo)

    Args:
        name (str): name of the model.
        model_type (str): type of the model.
        is_eval (bool): whether the model is in eval mode. Default: False.
        device (str): device to use. Default: "cpu".
        checkpoint (str): path or to checkpoint. Default: None.
            Note that expecting the checkpoint to have the same keys in state_dict as the model.

    Returns:
        model (torch.nn.Module): model.
    """

    model = registry.get_model_class(name).from_pretrained(model_type=model_type)

    if checkpoint is not None:
        model.load_checkpoint(checkpoint)

    if is_eval:
        model.eval()

    if device == "cpu":
        model = model.float()

    return model.to(device)


def load_preprocess(config):
    """
    Load preprocessor configs and construct preprocessors.

    If no preprocessor is specified, return BaseProcessor, which does not do any preprocessing.

    Args:
        config (dict): preprocessor configs.

    Returns:
        vis_processors (dict): preprocessors for visual inputs.
        txt_processors (dict): preprocessors for text inputs.

        Key is "train" or "eval" for processors used in training and evaluation respectively.
    """

    def _build_proc_from_cfg(cfg):
        return (
            registry.get_processor_class(cfg.name).from_config(cfg)
            if cfg is not None
            else BaseProcessor()
        )

    vis_processors = dict()
    txt_processors = dict()

    vis_proc_cfg = config.get("vis_processor")
    txt_proc_cfg = config.get("text_processor")

    if vis_proc_cfg is not None:
        vis_train_cfg = vis_proc_cfg.get("train")
        vis_eval_cfg = vis_proc_cfg.get("eval")
    else:
        vis_train_cfg = None
        vis_eval_cfg = None

    vis_processors["train"] = _build_proc_from_cfg(vis_train_cfg)
    vis_processors["eval"] = _build_proc_from_cfg(vis_eval_cfg)

    if txt_proc_cfg is not None:
        txt_train_cfg = txt_proc_cfg.get("train")
        txt_eval_cfg = txt_proc_cfg.get("eval")
    else:
        txt_train_cfg = None
        txt_eval_cfg = None

    txt_processors["train"] = _build_proc_from_cfg(txt_train_cfg)
    txt_processors["eval"] = _build_proc_from_cfg(txt_eval_cfg)

    return vis_processors, txt_processors


def load_model_and_preprocess(name, model_type, is_eval=False, device="cpu"):
    """
    Load model and its related preprocessors.

    List all available models and types in registry:
    >>> from bubogpt.models import model_zoo
    >>> print(model_zoo)

    Args:
        name (str): name of the model.
        model_type (str): type of the model.
        is_eval (bool): whether the model is in eval mode. Default: False.
        device (str): device to use. Default: "cpu".

    Returns:
        model (torch.nn.Module): model.
        vis_processors (dict): preprocessors for visual inputs.
        txt_processors (dict): preprocessors for text inputs.
    """
    model_cls = registry.get_model_class(name)

    # load model
    model = model_cls.from_pretrained(model_type=model_type)

    if is_eval:
        model.eval()

    # load preprocess
    cfg = OmegaConf.load(model_cls.default_config_path(model_type))
    if cfg is not None:
        preprocess_cfg = cfg.preprocess

        vis_processors, txt_processors = load_preprocess(preprocess_cfg)
    else:
        vis_processors, txt_processors = None, None
        logging.info(
            f"""No default preprocess for model {name} ({model_type}).
                This can happen if the model is not finetuned on downstream datasets,
                or it is not intended for direct use without finetuning.
            """
        )

    if device == "cpu" or device == torch.device("cpu"):
        model = model.float()

    return model.to(device), vis_processors, txt_processors


class ModelZoo:
    """
    A utility class to create string representation of available model architectures and types.

    >>> from bubogpt.models import model_zoo
    >>> # list all available models
    >>> print(model_zoo)
    >>> # show total number of models
    >>> print(len(model_zoo))
    """

    def __init__(self) -> None:
        self.model_zoo = {
            k: list(v.PRETRAINED_MODEL_CONFIG_DICT.keys())
            for k, v in registry.mapping["model_name_mapping"].items()
        }

    def __str__(self) -> str:
        return (
            "=" * 50
            + "\n"
            + f"{'Architectures':<30} {'Types'}\n"
            + "=" * 50
            + "\n"
            + "\n".join(
                [
                    f"{name:<30} {', '.join(types)}"
                    for name, types in self.model_zoo.items()
                ]
            )
        )

    def __iter__(self):
        return iter(self.model_zoo.items())

    def __len__(self):
        return sum([len(v) for v in self.model_zoo.values()])


model_zoo = ModelZoo()