Spaces:
Sleeping
Sleeping
| # This file is modified from https://github.com/haotian-liu/LLaVA/ | |
| import torch | |
| import torch.nn as nn | |
| from transformers import ( | |
| CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig, | |
| SiglipVisionModel, SiglipImageProcessor, SiglipVisionConfig | |
| ) | |
| class CLIPVisionTower(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() | |
| 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) | |
| 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: | |
| 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_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 | |
| def image_size(self): | |
| return self.config.image_size | |
| class SiglipVisionTower(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.input_image_size = args.input_image_size | |
| self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch') | |
| self.is_loaded = False | |
| if not delay_load: | |
| self.load_model() | |
| else: | |
| self.cfg_only = SiglipVisionConfig.from_pretrained(self.vision_tower_name) | |
| def load_model(self, device_map=None): | |
| if self. is_loaded: | |
| return | |
| self.image_processor = SiglipImageProcessor.from_pretrained(self.vision_tower_name) | |
| self.vision_tower = SiglipVisionModel.from_pretrained(self.vision_tower_name) | |
| self.image_processor.crop_size = {'height':self.input_image_size, 'width':self.input_image_size} | |
| self.is_loaded = True | |
| def feature_select(self, image_forward_outs, dtype): | |
| image_features = image_forward_outs.hidden_states | |
| if self.select_feature == 'patch': | |
| image_features = image_features[self.select_layer].to(dtype) | |
| elif self.select_feature == 'list': | |
| image_features = [feature.to(dtype) for feature in image_features[::7]] | |
| 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.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True) | |
| image_feature = self.feature_select(image_forward_out, image.dtype) | |
| image_features.append(image_feature) | |
| else: | |
| batch_size = images.shape[0] | |
| chunk_size = 256 | |
| image_features = [] | |
| for i in range(0, batch_size, chunk_size): | |
| chunk = images[i:i+chunk_size].to(device=self.device, dtype=self.dtype) | |
| chunk_forward_outs = self.vision_tower(chunk, output_hidden_states=True) | |
| chunk_features = self.feature_select(chunk_forward_outs, images.dtype) | |
| image_features.append(chunk_features) | |
| image_features = torch.cat(image_features, dim=0) | |
| 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(self): | |
| return (self.config.image_size // self.config.patch_size) ** 2 | |
| def num_patches_per_side(self): | |
| return self.config.image_size // self.config.patch_size | |
| def image_size(self): | |
| return self.config.image_size |