| import torch |
| import clip |
| from PIL import Image |
| from glob import glob |
| import numpy as np |
| import os |
| from editboard.utils import load_json |
| from tqdm import tqdm |
|
|
| def crop_read_image_path(image_path): |
| origin_image = Image.open(image_path) |
| w, h = origin_image.size |
| if h > w: |
| origin_image = origin_image.crop((0, h-w, w, h)) |
| return origin_image |
|
|
| def edit_success(image_path, source_prompt,target_prompt, model, preprocess, device): |
| image = preprocess(crop_read_image_path(image_path)).unsqueeze(0).to(device) |
|
|
| text = clip.tokenize([source_prompt, target_prompt]).to(device) |
| target = clip.tokenize(target_prompt).to(device) |
|
|
|
|
| with torch.no_grad(): |
| image_features = model.encode_image(image) |
| text_features = model.encode_text(text) |
| target_features = model.encode_text(target) |
|
|
| logits_per_image, logits_per_text = model(image, text) |
| probs = logits_per_image.softmax(dim=-1).cpu().numpy() |
|
|
|
|
| image_features = image_features.cpu().numpy() |
| target_features = target_features.cpu().numpy() |
| image_features_normalized = image_features / np.linalg.norm(image_features) |
| text_features_normalized = target_features / np.linalg.norm(target_features) |
| |
| |
| image_features_normalized = image_features_normalized |
| text_features_normalized = text_features_normalized |
|
|
| similarity = np.sum(image_features_normalized * text_features_normalized, -1) |
|
|
| if probs[0,1] >= probs[0,0]: |
| return 1, similarity[0] |
| |
| else: |
| return 0, similarity[0] |
|
|
| def video_score(edited_video_path, source_prompt, target_prompt, model, preprocess, device): |
| count = 0 |
| score = 0 |
| file_list = os.listdir(edited_video_path) |
| file_list = [img for img in file_list if (img.endswith('.png') or img.endswith('.jpg') or img.endswith('.jpeg'))] |
|
|
| for i in file_list: |
| image_path = os.path.join(edited_video_path, i) |
| count_sub, score_sub = edit_success(image_path, source_prompt, target_prompt, model, preprocess, device) |
| count+=count_sub |
| score+=score_sub |
|
|
| success_rate = count/len(file_list) |
| clip_similarity = score/len(file_list) |
| |
| return success_rate |
|
|
| def compute_success_rate(json_dir, device, submodules_list): |
| model, preprocess = clip.load("ViT-B/32", device=device) |
|
|
| metadata = load_json(json_dir) |
| result = {} |
| for i in tqdm(metadata): |
| score = video_score(i["edited_video_path"], i["source_prompt"], i["target_prompt"], model, preprocess, device) |
| result[i["edited_video_path"] + i["source_prompt"] + i["target_prompt"]] = score |
| return result |
