import torch from torchvision import transforms as T import numpy as np from CLIP import clip_explainability as clip device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # https://github.com/hila-chefer/Transformer-MM-Explainability/blob/main/CLIP_explainability.ipynb class ClipRelevancy(torch.nn.Module): def __init__(self, cfg): super().__init__() self.cfg = cfg # TODO it would make more sense not to load ths model again (already done in the extractor) self.model = clip.load("ViT-B/32", device=device, jit=False)[0] clip_input_size = 224 self.preprocess = T.Compose( [ T.Resize((clip_input_size, clip_input_size)), T.Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711]), ] ) input_prompts = cfg["bootstrap_text"] if type(input_prompts) == str: input_prompts = [input_prompts] self.text = clip.tokenize(input_prompts).to(cfg["device"]) if self.cfg["use_negative_bootstrap"]: input_negative_prompts = cfg["bootstrap_negative_text"] if type(input_negative_prompts) == str: input_negative_prompts = [input_negative_prompts] self.bootstrap_negative_text = clip.tokenize(input_negative_prompts).to(cfg["device"]) def image_relevance(self, image_relevance): patch_size = 32 # hardcoded for ViT-B/32 which we use h = w = 224 image_relevance = image_relevance.reshape(1, 1, h // patch_size, w // patch_size) image_relevance = torch.nn.functional.interpolate(image_relevance, size=(h, w), mode="bilinear") image_relevance = image_relevance.reshape(h, w).to(device) image_relevance = (image_relevance - image_relevance.min()) / (image_relevance.max() - image_relevance.min()) return image_relevance def interpret(self, image, negative=False): text = self.text if not negative else self.bootstrap_negative_text batch_size = text.shape[0] images = image.repeat(batch_size, 1, 1, 1) # TODO this is pretty inefficient, we can calculate the text embeddings instead of recomputing at each call logits_per_image, logits_per_text = self.model(images, text) probs = logits_per_image.softmax(dim=-1).detach().cpu().numpy() index = [i for i in range(batch_size)] one_hot = np.zeros((logits_per_image.shape[0], logits_per_image.shape[1]), dtype=np.float32) one_hot[torch.arange(logits_per_image.shape[0]), index] = 1 one_hot = torch.from_numpy(one_hot).requires_grad_(True) one_hot = torch.sum(one_hot.to(device) * logits_per_image) self.model.zero_grad() image_attn_blocks = list(dict(self.model.visual.transformer.resblocks.named_children()).values()) num_tokens = image_attn_blocks[0].attn_probs.shape[-1] R = torch.eye(num_tokens, num_tokens, dtype=image_attn_blocks[0].attn_probs.dtype).to(device) R = R.unsqueeze(0).expand(batch_size, num_tokens, num_tokens) for i, blk in enumerate(image_attn_blocks): if i <= self.cfg["relevancy_num_layers"]: continue grad = torch.autograd.grad(one_hot, [blk.attn_probs], retain_graph=True)[0].detach() cam = blk.attn_probs.detach() cam = cam.reshape(-1, cam.shape[-1], cam.shape[-1]) grad = grad.reshape(-1, grad.shape[-1], grad.shape[-1]) cam = grad * cam cam = cam.reshape(batch_size, -1, cam.shape[-1], cam.shape[-1]) cam = cam.clamp(min=0).mean(dim=1) R = R + torch.bmm(cam, R) image_relevance = R[:, 0, 1:] return image_relevance def forward(self, img, preprocess=True, negative=False): if preprocess: img = self.preprocess(img) R_image = self.interpret(img, negative=negative) res = [] for el in R_image: res.append(self.image_relevance(el).float()) res = torch.stack(res, dim=0) return res