temporalSceneClustering.py

# Lib
import os
import glob
import json
import tqdm
import natsort
import random

from PIL import Image

import numpy as np

import torch
from torch.utils.data import Dataset, DataLoader

import clip

from torchvision import models

from config import config


class loading_img(Dataset):
    def __init__(self, img_list):
        self.img_list = img_list

    def __len__(self):
        return len(self.img_list)

    def __getitem__(self, idx):
        return preprocess(Image.open(self.img_list[idx]))


# select frames
def select_frames(folder_list, preprocess, resnet18_pretrained):
    for folder in folder_list: 
        img_list = natsort.natsorted(glob.glob(f"{folder}/*.jpg"))
        img_feats = []

        img_set = loading_img(img_list)
        img_loader = DataLoader(img_set, batch_size=64, shuffle=False, num_workers=16)

        for imgtensor in img_loader: img_feats.append(imgtensor)
        img_feats = torch.concat(img_feats, dim=0).to(device)

        with torch.no_grad():
            featuremap = resnet18_pretrained(img_feats)
            frame_num = featuremap.shape[0]

            dist_list = []
            for img_feat in featuremap: dist_list.append(torch.mean(torch.sqrt((featuremap-img_feat)**2), dim=-1))
            dist_list = torch.concat(dist_list).reshape(frame_num, frame_num)

            idx_list = [_ for _ in range(frame_num)]
            loop_idx = 0
            out_frames = []

            output_results = []
            while len(idx_list) > 5:
                dist_idx = idx_list.pop(0)

                data = dist_list[dist_idx, idx_list].softmax(dim=-1)
                mu, std = torch.mean(data), torch.std(data)
                pop_idx_list = torch.where(data < mu-std*(np.exp(1-loop_idx/config.divlam)))[0].detach().cpu().numpy()
                result = list(np.array(idx_list)[pop_idx_list])
                result.append(dist_idx)
                output_results.append(result)

                num_picks = 18
                if len(result) > num_picks:
                    idx_result_list = sorted(random.sample(result, num_picks)) 
                    img_list = np.array(img_list)
                    idx_result_list = np.array(idx_result_list)
                    out_frames.extend(img_list[idx_result_list])
                else:
                    idx_result_list = sorted(result)
                    img_list = np.array(img_list)
                    idx_result_list = np.array(idx_result_list)
                    out_frames.extend(img_list[idx_result_list])

                loop_idx += 1
                
                for pop_idx in reversed(pop_idx_list): idx_list.pop(pop_idx)

    return out_frames, output_results


# Init
random.seed(10)

device = "cuda" if torch.cuda.is_available() else "cpu"

resnet18_pretrained = models.resnet18(pretrained=True).to(device)
resnet18_pretrained.fc = torch.nn.Identity()
resnet18_pretrained.avgpool = torch.nn.Identity()
resnet18_pretrained.eval()

model, preprocess = clip.load("ViT-B/32", device=device)

objs_acts = config.f_path
questions = config.q_path

questions = [json.loads(q) for q in open(os.path.expanduser(questions), "r")]
objs_acts = [json.loads(q) for q in open(os.path.expanduser(objs_acts), "r")]

answer_path = os.path.expanduser(config.a_path)
os.makedirs(os.path.dirname(answer_path), exist_ok=True)
ans_file = open(answer_path, "w")

output_results = []
for question in tqdm.tqdm(questions):
    test_token = True

    for objs_act in objs_acts:
        if objs_act['q_uid'] == question['q_uid']:
            question['Object'] = objs_act["Activity"]
            question['Activity'] = objs_act["Activity"]

            folder_list = glob.glob(f"{config.img_folder}/{question['q_uid']}/")
            out_frames, output_result = select_frames(folder_list, preprocess, resnet18_pretrained)
            output_results.append(output_result)
            question['filepath'] = out_frames

            ans_file.write(json.dumps(question) + "\n")
            test_token = False
            break