File size: 7,091 Bytes
e84842d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
 Copyright (c) 2022, salesforce.com, inc.
 All rights reserved.
 SPDX-License-Identifier: BSD-3-Clause
 For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
"""

import os

import torch
import torch.nn.functional as F
from lavis.common.dist_utils import download_cached_file
from lavis.common.registry import registry
from lavis.common.utils import get_abs_path, is_url
from lavis.models.base_model import MomentumDistilationMixin
from lavis.models.blip_models.blip import BlipBase
from lavis.models.blip_models.blip_outputs import BlipIntermediateOutput, BlipOutput
from lavis.models.blip_models.nlvr_encoder import BertModel
from lavis.models.vit import VisionTransformerEncoder, interpolate_pos_embed
from torch import nn
from transformers import BertConfig


@registry.register_model("blip_nlvr")
class BlipNLVR(BlipBase, MomentumDistilationMixin):
    """
    Class for BLIP NLVR model.

    Supported model types:
        - base: model with pre-trained BLIP weights, used as initialization for fine-tuning.
        - nlvr: finetuned model on NLVR2 dataset.

    Usage:
        >>> from lavis.models import load_model
        >>> model = load_model("blip_nlvr", "nlvr")
    """

    PRETRAINED_MODEL_CONFIG_DICT = {
        "nlvr": "configs/models/blip_nlvr.yaml",
    }

    def __init__(self, image_encoder, text_encoder, num_classes):
        super().__init__()

        self.tokenizer = self.init_tokenizer()
        self.visual_encoder = image_encoder
        self.text_encoder = text_encoder

        hidden_size = text_encoder.config.hidden_size
        self.cls_head = nn.Sequential(
            nn.Linear(hidden_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, num_classes),
        )

    def forward(self, samples, is_train=True):
        """
        Forward function for training and evaluation.

        Args:
            samples (dict): a dict of input samples, which contains the following keys:
                - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.
                - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.
                - text_input (list): list of strings, each string is a natural language sentence.
                - label (torch.LongTensor): ground truth label with shape (batch_size,).
            is_train (bool): whether the model is in training mode.
                If True, the model will return the loss;
                If False, the model will return the prediction.

        Examples:
            >>> import torch
            >>> from lavis.models import load_model
            >>> model = load_model("blip_nlvr", "nlvr")
            >>> samples = {
            ...     "image0": torch.randn(2, 3, 384, 384),
            ...     "image1": torch.randn(2, 3, 384, 384),
            ...     "text_input": ["there is a ferret in tall grass", "there are lips in one of the images"],
            ...     "label": torch.tensor([0, 1]),
            ... }
            >>> output = model(samples)
            >>> output.keys()
            odict_keys(['intermediate_output', 'loss'])
        """
        text = samples["text_input"]
        text = self.tokenizer(text, padding="longest", return_tensors="pt").to(
            self.device
        )
        text.input_ids[:, 0] = self.tokenizer.enc_token_id

        targets = samples["label"]

        image0 = samples["image0"]
        image1 = samples["image1"]
        images = torch.cat([image0, image1], dim=0)

        image_embeds = self.visual_encoder.forward_features(images)
        image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(
            self.device
        )
        image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))

        encoder_output = self.text_encoder(
            text.input_ids,
            attention_mask=text.attention_mask,
            encoder_hidden_states=[image0_embeds, image1_embeds],
            encoder_attention_mask=[
                image_atts[: image0_embeds.size(0)],
                image_atts[image0_embeds.size(0) :],
            ],
            return_dict=True,
        )

        prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])

        if is_train:
            loss = F.cross_entropy(prediction, targets)
            # return {"loss": loss}
            return BlipOutput(
                loss=loss,
                intermediate_output=BlipIntermediateOutput(
                    image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),
                    encoder_output=encoder_output,
                ),
            )
        else:
            return {"predictions": prediction, "targets": targets}

    def predict(self, samples):
        output = self.forward(samples, is_train=False)
        return output

    @classmethod
    def from_config(cls, cfg=None):
        image_encoder = VisionTransformerEncoder.from_config(cfg)

        # text encoder + multimodal encoder
        bert_config = BertConfig.from_json_file(get_abs_path(cfg["med_config_path"]))
        text_encoder = BertModel(config=bert_config, add_pooling_layer=False)

        num_classes = cfg.get("num_classes", 3)

        assert num_classes > 1, "Invalid number of classes provided, found {}".format(
            num_classes
        )

        model = cls(
            image_encoder=image_encoder,
            text_encoder=text_encoder,
            num_classes=num_classes,
        )

        model.load_checkpoint_from_config(cfg)

        return model

    def load_from_pretrained(self, url_or_filename):
        if is_url(url_or_filename):
            cached_file = download_cached_file(
                url_or_filename, check_hash=False, progress=True
            )
            checkpoint = torch.load(cached_file, map_location="cpu")
        elif os.path.isfile(url_or_filename):
            checkpoint = torch.load(url_or_filename, map_location="cpu")
        else:
            raise RuntimeError("checkpoint url or path is invalid")
        state_dict = checkpoint["model"]

        state_dict["visual_encoder.pos_embed"] = interpolate_pos_embed(
            state_dict["visual_encoder.pos_embed"], self.visual_encoder
        )

        for key in list(state_dict.keys()):
            if "crossattention.self." in key:
                new_key0 = key.replace("self", "self0")
                new_key1 = key.replace("self", "self1")
                state_dict[new_key0] = state_dict[key]
                state_dict[new_key1] = state_dict[key]
            elif "crossattention.output.dense." in key:
                new_key0 = key.replace("dense", "dense0")
                new_key1 = key.replace("dense", "dense1")
                state_dict[new_key0] = state_dict[key]
                state_dict[new_key1] = state_dict[key]

        msg = self.load_state_dict(state_dict, strict=False)
        print("load checkpoint from %s" % url_or_filename)
        print(f"missing keys {msg.missing_keys}")
        return msg