File size: 8,339 Bytes
122057f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import logging
import random

import torch
from torch.cuda.amp import autocast as autocast
import torch.nn as nn

from minigpt4.common.registry import registry
from minigpt4.models.base_model import disabled_train
from minigpt4.models.minigpt_base import MiniGPTBase
from minigpt4.models.Qformer import BertConfig, BertLMHeadModel


@registry.register_model("minigpt4")
class MiniGPT4(MiniGPTBase):
    """
    MiniGPT-4 model
    """

    PRETRAINED_MODEL_CONFIG_DICT = {
        "pretrain_vicuna0": "configs/models/minigpt4_vicuna0.yaml",
        "pretrain_llama2": "configs/models/minigpt4_llama2.yaml",
    }

    def __init__(
            self,
            vit_model="eva_clip_g",
            q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth",
            img_size=224,
            drop_path_rate=0,
            use_grad_checkpoint=False,
            vit_precision="fp16",
            freeze_vit=True,
            has_qformer=True,
            freeze_qformer=True,
            num_query_token=32,
            llama_model="",
            prompt_path="",
            prompt_template="",
            max_txt_len=32,
            end_sym='\n',
            low_resource=False,  # use 8 bit and put vit in cpu
            device_8bit=0,  # the device of 8bit model should be set when loading and cannot be changed anymore.
            lora_r=64,
            lora_target_modules=['query_key_value','dense'],
            lora_alpha=16,
            lora_dropout=0.05,
    ):
        super().__init__(
            vit_model=vit_model,
            img_size=img_size,
            drop_path_rate=drop_path_rate,
            use_grad_checkpoint=use_grad_checkpoint,
            vit_precision=vit_precision,
            freeze_vit=freeze_vit,
            llama_model=llama_model,
            max_txt_len=max_txt_len,
            end_sym=end_sym,
            low_resource=low_resource,
            device_8bit=device_8bit,
            lora_r=lora_r,
            lora_target_modules=lora_target_modules,
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
        )

        self.has_qformer = True
        if self.has_qformer:
            print('Loading Q-Former')
            self.Qformer, self.query_tokens = self.init_Qformer(
                num_query_token, self.visual_encoder.num_features, freeze_qformer
            )
            self.load_from_pretrained(url_or_filename=q_former_model)  # load q-former weights here

            img_f_dim = self.Qformer.config.hidden_size
            print('Loading Q-Former Done')
        else:
            img_f_dim = self.visual_encoder.num_features * 4
            print('Do not use Q-Former here.')
        print(img_f_dim,self.llama_model.config.hidden_size)
        self.llama_proj = nn.Linear(
            self.Qformer.config.hidden_size, 4096
        )
        self.llama_proj2 = nn.Linear(
            4096, self.llama_model.config.hidden_size
        )


        if prompt_path:
            with open(prompt_path, 'r') as f:
                raw_prompts = f.read().splitlines()
            filted_prompts = [raw_prompt for raw_prompt in raw_prompts if "<ImageHere>" in raw_prompt]
            self.prompt_list = [prompt_template.format(p) for p in filted_prompts]
            print('Load {} training prompts'.format(len(self.prompt_list)))
            print('Prompt Example \n{}'.format(random.choice(self.prompt_list)))
        else:
            self.prompt_list = []

    @classmethod
    def init_Qformer(cls, num_query_token, vision_width, freeze):
        encoder_config = BertConfig.from_pretrained("bert-base-uncased")
        encoder_config.encoder_width = vision_width
        # insert cross-attention layer every other block
        encoder_config.add_cross_attention = True
        encoder_config.cross_attention_freq = 2
        encoder_config.query_length = num_query_token
        Qformer = BertLMHeadModel(config=encoder_config)
        query_tokens = nn.Parameter(
            torch.zeros(1, num_query_token, encoder_config.hidden_size)
        )
        query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)

        Qformer.cls = None
        Qformer.bert.embeddings.word_embeddings = None
        Qformer.bert.embeddings.position_embeddings = None
        for layer in Qformer.bert.encoder.layer:
            layer.output = None
            layer.intermediate = None

        if freeze:
            for name, param in Qformer.named_parameters():
                param.requires_grad = False
            Qformer = Qformer.eval()
            Qformer.train = disabled_train
            query_tokens.requires_grad = False
            logging.info("freeze Qformer")

        return Qformer, query_tokens

    def encode_img(self, image):
        device = image.device

        if len(image.shape) > 4:
            image = image.reshape(-1, *image.shape[-3:])

        with self.maybe_autocast():
            image_embeds = self.ln_vision(self.visual_encoder(image)).to(device)
            if self.has_qformer:
                image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(device)

                query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
                query_output = self.Qformer.bert(
                    query_embeds=query_tokens,
                    encoder_hidden_states=image_embeds,
                    encoder_attention_mask=image_atts,
                    return_dict=True,
                )

                inputs_llama = self.llama_proj(query_output.last_hidden_state)
                inputs_llama = self.llama_proj2(inputs_llama)

            else:
                image_embeds = image_embeds[:, 1:, :]
                bs, pn, hs = image_embeds.shape
                image_embeds = image_embeds.view(bs, int(pn / 4), int(hs * 4))

                inputs_llama = self.llama_proj(image_embeds)
                inputs_llama = self.llama_proj2(inputs_llama)

            atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
        return inputs_llama, atts_llama

    @classmethod
    def from_config(cls, cfg):
        vit_model = cfg.get("vit_model", "eva_clip_g")
        q_former_model = cfg.get("q_former_model", "https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth")
        img_size = cfg.get("image_size")
        num_query_token = cfg.get("num_query_token")
        llama_model = cfg.get("llama_model")

        drop_path_rate = cfg.get("drop_path_rate", 0)
        use_grad_checkpoint = cfg.get("use_grad_checkpoint", False)
        vit_precision = cfg.get("vit_precision", "fp16")
        freeze_vit = cfg.get("freeze_vit", True)
        has_qformer = cfg.get("has_qformer", True)
        freeze_qformer = cfg.get("freeze_qformer", True)
        low_resource = cfg.get("low_resource", False)
        device_8bit = cfg.get("device_8bit", 0)

        prompt_path = cfg.get("prompt_path", "")
        prompt_template = cfg.get("prompt_template", "")
        max_txt_len = cfg.get("max_txt_len", 32)
        end_sym = cfg.get("end_sym", '\n')

        lora_r = cfg.get("lora_r", 64)
        lora_alpha = cfg.get("lora_alpha", 16)

        model = cls(
            vit_model=vit_model,
            q_former_model=q_former_model,
            img_size=img_size,
            drop_path_rate=drop_path_rate,
            use_grad_checkpoint=use_grad_checkpoint,
            vit_precision=vit_precision,
            freeze_vit=freeze_vit,
            has_qformer=has_qformer,
            freeze_qformer=freeze_qformer,
            num_query_token=num_query_token,
            llama_model=llama_model,
            prompt_path=prompt_path,
            prompt_template=prompt_template,
            max_txt_len=max_txt_len,
            end_sym=end_sym,
            low_resource=low_resource,
            device_8bit=device_8bit,
            lora_r=lora_r,
            lora_alpha=lora_alpha,
        )

        ckpt_path = cfg.get("ckpt", "")  # load weights of MiniGPT-4
        if ckpt_path:
            print("Load MiniGPT-4 Checkpoint: {}".format(ckpt_path))
            ckpt = torch.load(ckpt_path, map_location="cpu")
            msg = model.load_state_dict(ckpt['model'], strict=False)

        return model