File size: 8,445 Bytes
80e48d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import List, Optional, Tuple, Union

from .configuration_uform_gen import VLMConfig

import torch
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
from torch import nn

from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.models.auto.modeling_auto import AutoModelForCausalLM, AutoModel
from transformers import AutoConfig
from transformers.utils import logging

from .vision_encoder import VisionEncoder


class ImageFeaturesPooler(nn.Module):
    def __init__(self, config, text_config):
        super().__init__()
        self.pooler = nn.TransformerDecoderLayer(
            config.image_encoder_hidden_size,
            config.image_pooler_num_attn_heads,
            config.image_pooler_intermediate_size,
            activation=nn.functional.silu,
            batch_first=True,
            norm_first=True,
        )
        self.image_latents = nn.Parameter(
            torch.randn(1, config.num_image_latents, config.image_encoder_hidden_size)
            * config.initializer_range**0.5
        )
        self.projection = nn.Linear(config.image_encoder_hidden_size, text_config.hidden_size)

    def forward(self, features):
        features = self.pooler(
            self.image_latents.expand(features.size(0), -1, -1), features
        )

        return self.projection(features)


class VLMPreTrainedModel(PreTrainedModel):
    config_class = VLMConfig
    base_model_prefix = "vlm"
    supports_gradient_checkpointing = True
    _no_split_modules = []
    _skip_keys_device_placement = "past_key_values"

    def _init_weights(self, module):
        pass

    def _initialize_weights(self, module):
        pass


class VLMForCausalLM(VLMPreTrainedModel):
    def __init__(self, config: VLMConfig):
        super().__init__(config)

        self.config = config
        self.text_config = AutoConfig.from_pretrained(
            config.text_decoder_name_or_path,
            trust_remote_code=True
        )
        
        self.text_decoder = AutoModelForCausalLM.from_config(
            self.text_config,
            trust_remote_code=True
        )

        self.image_encoder = VisionEncoder(
            config.image_encoder_hidden_size,
            config.image_encoder_patch_size,
            config.image_encoder_num_layers,
            config.image_encoder_num_heads,
        )

        self.image_pooler = ImageFeaturesPooler(config, self.text_config)

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

    def set_input_embeddings(self, value):
        self.text_decoder.set_input_embeddings(value)

    def get_images_embeddings(self, images):
        features = self.image_encoder(images)
        return self.image_pooler(features)

    def gather_continuous_embeddings(
        self,
        input_ids: torch.Tensor,
        word_embeddings: torch.Tensor,
        image_embeddings: torch.Tensor
    ) -> torch.Tensor:
        
        start_indices = (input_ids == self.config.image_token_id).nonzero()[:, 1]
        embeddings = []
        for sample_idx, start_idx in enumerate(start_indices.tolist()):
            embeddings.append(
                torch.cat(
                    (
                        word_embeddings[sample_idx, :start_idx],
                        image_embeddings[sample_idx],
                        word_embeddings[sample_idx, start_idx + 1 :],
                    ),
                    dim=0,
                )
            )

        return torch.stack(embeddings, dim=0)

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        images: torch.Tensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None
    ) -> Union[dict, Tuple, CausalLMOutputWithPast]:
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time"
            )
        elif input_ids is None and inputs_embeds is None:
            raise ValueError("You have to specify either input_is or inputs_embeds")

        if inputs_embeds is None and past_key_values is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

            if images is not None:
                image_embeds = self.get_images_embeddings(images)
                inputs_embeds = self.gather_continuous_embeddings(
                    input_ids,
                    inputs_embeds,
                    image_embeds
                )

        if position_ids is None:
            seq_length = (
                inputs_embeds.shape[1]
                if inputs_embeds is not None
                else input_ids.shape[1]
            )
            past_key_values_length = 0

            if past_key_values is not None:
                past_key_values_length = past_key_values[0][0].shape[2]

            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length,
                seq_length + past_key_values_length,
                dtype=torch.long,
                device=device,
            )
            position_ids = position_ids.unsqueeze(0)

        outputs = self.text_decoder(
            inputs_embeds=inputs_embeds,
            input_ids=input_ids if past_key_values is not None else None,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            use_cache=use_cache,
            return_dict=return_dict,
        )

        return outputs

    def prepare_inputs_for_generation(
        self,
        input_ids,
        images=None,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        **kwargs,
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -1].unsqueeze(-1)

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
            n_samples = inputs_embeds.shape[0]
        else:
            model_inputs = {"input_ids": input_ids}
            n_samples = input_ids.shape[0]

        if images is not None:
            model_inputs["images"] = images

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "images": images if past_key_values is None else None,
            }
        )
        return model_inputs

    @classmethod
    def from_config(cls, config, **kwargs):
        return cls._from_config(config, **kwargs)


VLMConfig.register_for_auto_class()
VLMForCausalLM.register_for_auto_class("AutoModel")