| | import torch |
| | import torch.nn as nn |
| |
|
| | from .processor import Blip2ImageTrainProcessor |
| | from .eva_vit import create_eva_vit_g |
| |
|
| |
|
| | class EvaClipVisionTower(nn.Module): |
| |
|
| | def __init__(self, vision_tower, args, delay_load=False): |
| | super().__init__() |
| | self.is_loaded = False |
| | self.vision_tower_name = vision_tower |
| | |
| | self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch') |
| | self.args = args |
| |
|
| | if not delay_load: |
| | self.load_model() |
| |
|
| | |
| |
|
| |
|
| | def load_model(self, device_map=None): |
| | if self.is_loaded: |
| | print('{} is already loaded, `load_model` called again, skipping.'.format(self.vision_tower_name)) |
| | return |
| | |
| | if not hasattr(self.args, 'dynamic_resolution'): |
| | dynamic_resolution = None |
| | else: |
| | dynamic_resolution = self.args.dynamic_resolution |
| |
|
| |
|
| | if (not hasattr(self.args, 'freeze_vision_encoder')) or self.args.freeze_vision_encoder: |
| | use_checkpoint = False |
| | else: |
| | use_checkpoint = True |
| | assert self.args.vit_precision == 'fp32', 'if the vision encoder is training, the type needs to be fp32' |
| | |
| | |
| | self.image_processor = Blip2ImageTrainProcessor( |
| | image_size=self.args.img_size, |
| | dynamic_resolution= dynamic_resolution |
| | ) |
| | self.vision_tower = create_eva_vit_g( |
| | img_size=self.args.img_size, |
| | drop_path_rate=self.args.drop_path_rate, |
| | precision=self.args.vit_precision, |
| | vit_model_path=self.args.vit_model_path, |
| | use_checkpoint=use_checkpoint |
| | ) |
| |
|
| | |
| |
|
| | self.is_loaded = True |
| |
|
| |
|
| | def feature_select(self, image_features): |
| | if self.select_feature == 'patch': |
| | image_features = image_features[:, 1:] |
| | elif self.select_feature == 'cls_patch': |
| | image_features = image_features |
| | else: |
| | raise ValueError(f'Unexpected select feature: {self.select_feature}') |
| | return image_features |
| |
|
| | |
| | def forward(self, images): |
| | if type(images) is list: |
| | image_features = [] |
| | for image in images: |
| | image_forward_out = self.vision_tower(image.unsqueeze(0)) |
| | image_features.append(self.feature_select(image_forward_out).to(image.dtype)) |
| |
|
| | else: |
| | image_features = self.vision_tower(images.to(dtype=self.dtype)) |
| | image_features = self.feature_select(image_features).to(images.dtype) |
| |
|
| | return image_features |
| |
|
| | @property |
| | def dummy_feature(self): |
| | return torch.zeros(1, self.hidden_size, dtype=torch.float) |
| |
|
| | @property |
| | def hidden_size(self): |
| | return self.vision_tower.hidden_size |
| |
|
| | @property |
| | def num_patches(self): |
| | return (self.vision_tower.image_size // self.vision_tower.patch_size) ** 2 |
| | |
| | @property |
| | def num_patches_per_side(self): |
| | return (self.vision_tower.image_size // self.vision_tower.patch_size) |
| |
|
| | @property |
| | def dtype(self): |
| | return self.vision_tower.pos_embed.dtype |
| |
|
| |
|
