Last commit not found
import torch | |
import torch.nn as nn | |
from transformers import CLIPImageProcessor, CLIPVisionConfig, CLIPVisionModel | |
class CLIPVisionTower(nn.Module): | |
def __init__(self, vision_tower, args, freeze_vision_tower=False, 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") | |
self.freeze_vision_tower = freeze_vision_tower | |
if not delay_load: | |
self.load_model() | |
elif getattr(args, "unfreeze_mm_vision_tower", False): | |
self.load_model() | |
else: | |
self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name) | |
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 | |
self.image_processor = CLIPImageProcessor.from_pretrained( | |
self.vision_tower_name | |
) | |
self.vision_tower = CLIPVisionModel.from_pretrained( | |
self.vision_tower_name, device_map=device_map | |
) | |
if self.freeze_vision_tower: | |
self.vision_tower.requires_grad_(False) | |
self.is_loaded = True | |
def feature_select(self, image_forward_outs): | |
image_features = image_forward_outs.hidden_states[self.select_layer] | |
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: | |
if self.freeze_vision_tower: | |
with torch.no_grad(): | |
image_forward_out = self.vision_tower( | |
image.to(device=self.device, dtype=self.dtype).unsqueeze(0), | |
output_hidden_states=True, | |
) | |
image_feature = self.feature_select(image_forward_out).to( | |
image.dtype | |
) | |
image_features.append(image_feature) | |
else: | |
image_forward_out = self.vision_tower( | |
image.to(device=self.device, dtype=self.dtype).unsqueeze(0), | |
output_hidden_states=True, | |
) | |
image_feature = self.feature_select(image_forward_out).to( | |
image.dtype | |
) | |
image_features.append(image_feature) | |
else: | |
if self.freeze_vision_tower: | |
with torch.no_grad(): | |
image_forward_out = self.vision_tower( | |
images.to(device=self.device, dtype=self.dtype), | |
output_hidden_states=True, | |
) | |
image_features = self.feature_select(image_forward_out).to( | |
images.dtype | |
) | |
else: | |
image_forward_outs = self.vision_tower( | |
images.to(device=self.device, dtype=self.dtype), | |
output_hidden_states=True, | |
) | |
image_features = self.feature_select(image_forward_outs).to( | |
images.dtype | |
) | |
return image_features | |
def dummy_feature(self): | |
return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype) | |
def dtype(self): | |
return self.vision_tower.dtype | |
def device(self): | |
return self.vision_tower.device | |
def config(self): | |
if self.is_loaded: | |
return self.vision_tower.config | |
else: | |
return self.cfg_only | |
def hidden_size(self): | |
return self.config.hidden_size | |
def num_patches_per_side(self): | |
return self.config.image_size // self.config.patch_size | |
def num_patches(self): | |
return (self.config.image_size // self.config.patch_size) ** 2 | |