Image-Text-to-Text
Transformers
PyTorch
English
doubutsu
conversational
custom_code
Inference Endpoints
File size: 4,452 Bytes
4b953b6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
from transformers import (
    PreTrainedModel,
    AutoModelForCausalLM,
    AutoModel,
    SiglipImageProcessor,
)
from .configuration_doubutsu import DoubutsuConfig


class ProjectionModule(nn.Module):
    def __init__(self, mm_hidden_size=1152, hidden_size=1536):
        super(ProjectionModule, self).__init__()

        self.model = nn.Sequential(
            nn.Linear(mm_hidden_size, hidden_size),
            nn.GELU(),
            nn.Linear(hidden_size, hidden_size),
        )

    def forward(self, x):
        return self.model(x)


class Doubutsu(PreTrainedModel):
    config_class = DoubutsuConfig

    def __init__(self, config):
        super().__init__(config)

        self.vision_model = AutoModel.from_config(self.config.vision_config)
        self.text_model = AutoModelForCausalLM.from_config(self.config.text_config)
        self.processor = SiglipImageProcessor()
        self.mm_projector = ProjectionModule(
            mm_hidden_size=config.vision_config.hidden_size,
            hidden_size=config.text_config.hidden_size,
        )

    @property
    def device(self):
        return self.text_model.device

    def encode_image(self, image):
        image = image.convert("RGB")
        image = self.processor(
            images=image,
            return_tensors="pt",
            do_resize=True,
            size={"height": 378, "width": 378},
        )["pixel_values"].to(
            device=self.vision_model.device, dtype=self.vision_model.dtype
        )
        with torch.no_grad():
            return self.vision_model(image, output_hidden_states=True).hidden_states[-2]

    def input_embeds(self, prompt, image_embeds, tokenizer):
        def _tokenize(txt):
            return tokenizer(
                txt, return_tensors="pt", add_special_tokens=False
            ).input_ids.to(self.device)

        text_emb = self.text_model.get_input_embeddings()
        embeds = []
        tokenized_prompt = _tokenize(prompt)

        # Add BOS token if it exists and isn't already at the start of the prompt
        if tokenizer.bos_token_id is not None:
            if tokenized_prompt[0][0] == tokenizer.bos_token_id:
                tokenized_prompt = tokenized_prompt[:, 1:]  # Remove existing BOS
            embeds.append(
                text_emb(torch.tensor([[tokenizer.bos_token_id]], device=self.device))
            )

        # Add image embeds
        projected_image_embeds = self.mm_projector(image_embeds.to(self.device))
        embeds.append(projected_image_embeds)

        # Add text embeds
        embeds.append(text_emb(tokenized_prompt))

        return torch.cat(embeds, dim=1)

    def get_input_embeddings(self):
        return self.text_model.get_input_embeddings()

    def generate(
        self,
        image_embeds,
        prompt,
        tokenizer,
        max_new_tokens=128,
        temperature=0.1,
        **kwargs,
    ):
        generate_config = {
            "eos_token_id": tokenizer.eos_token_id,
            "bos_token_id": tokenizer.bos_token_id,
            "pad_token_id": tokenizer.pad_token_id,
            "max_new_tokens": max_new_tokens,
            "temperature": temperature,
            **kwargs,
        }

        with torch.no_grad():
            inputs_embeds = self.input_embeds(prompt, image_embeds, tokenizer)
            output_ids = self.text_model.generate(
                inputs_embeds=inputs_embeds,
                do_sample=True,
                **generate_config,
            )
        return tokenizer.batch_decode(output_ids, skip_special_tokens=True)

    def answer_question(self, image, question, tokenizer, **kwargs):
        image_embeds = self.encode_image(image)

        chat = [
            {
                "role": "system",
                "content": "You are a helpful AI assistant that can see images and answer questions about them.",
            },
            {"role": "user", "content": question},
        ]
        prompt = tokenizer.apply_chat_template(
            chat, tokenize=False, add_generation_prompt=True
        )

        # Generate the answer
        with torch.no_grad():
            output = self.generate(
                image_embeds=image_embeds,
                prompt=prompt,
                tokenizer=tokenizer,
                **kwargs,
            )[0]

        # Clean and return the answer
        cleaned_answer = output.strip()
        return cleaned_answer