|
import torch |
|
import torch.nn as nn |
|
from .eva_vit import create_eva_vit_g, _cfg |
|
from .processor import ImageTrainProcessor, ImageEvalProcessor |
|
|
|
|
|
class EVAVisionTower(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_layer = args.mm_vision_select_layer |
|
self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch') |
|
|
|
if not delay_load: |
|
self.load_model() |
|
else: |
|
self.cfg_only = _cfg() |
|
|
|
def load_model(self): |
|
self.image_processor = ImageTrainProcessor() |
|
self.image_eval_processor = ImageEvalProcessor() |
|
self.vision_tower = create_eva_vit_g( |
|
img_size=224, drop_path_rate=0, use_checkpoint=False, precision="fp16" |
|
) |
|
|
|
|
|
self.is_loaded = True |
|
|
|
def feature_select(self, image_forward_outs, select_feature='patch'): |
|
image_features = image_forward_outs[self.select_layer] |
|
if select_feature == 'patch': |
|
image_features = image_features[:, 1:] |
|
elif select_feature == 'cls_patch': |
|
image_features = image_features |
|
else: |
|
raise ValueError(f'Unexpected select feature: {self.select_feature}') |
|
return image_features |
|
|
|
@torch.no_grad() |
|
def forward(self, images, select_feature='patch'): |
|
if type(images) is list: |
|
image_features = [] |
|
for image in images: |
|
image_forward_out = self.vision_tower.get_intermediate_layers(image.to(device=self.device, dtype=self.dtype).unsqueeze(0),) |
|
image_feature = self.feature_select(image_forward_out, select_feature).to(image.dtype) |
|
image_features.append(image_feature) |
|
else: |
|
image_forward_outs = self.vision_tower.get_intermediate_layers(images.to(device=self.device, dtype=self.dtype)) |
|
image_features = self.feature_select(image_forward_outs, select_feature).to(images.dtype) |
|
|
|
return image_features |
|
|
|
@property |
|
def dummy_feature(self): |
|
return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype) |
|
|
|
@property |
|
def dtype(self): |
|
return self.vision_tower.cls_token.dtype |
|
|
|
@property |
|
def device(self): |
|
return self.vision_tower.cls_token.device |
|
|
|
@property |
|
def config(self): |
|
if self.is_loaded: |
|
return self.vision_tower.config |
|
else: |
|
return self.cfg_only |
|
|
|
@property |
|
def hidden_size(self): |
|
return self.vision_tower.num_features |
|
|
|
@property |
|
def num_patches(self): |
|
return (self.config.image_size // self.config.patch_size) ** 2 |
