영상 요약해주는 Vision Transformer 모델

1. 22분 짜리 영상을 1분으로 요약하는 ViT모델

 

2. 데이터

데이터 분야 - AI 데이터찾기 - AI-Hub (aihub.or.kr)

AI-Hub

3. 학습모델

영상요약모델 - Google Drive

영상요약모델 - Google Drive

4.  

5. run

      

      

      

    
    

      <python />
    
import torch
from training.summary.datamodule import SummaryDataset
from transformers import ViTImageProcessor
from tqdm import tqdm
import matplotlib.pyplot as plt
import cv2
import seaborn as sns
import numpy as np
from moviepy.editor import VideoFileClip, concatenate_videoclips
from v2021 import SummaryModel

preprocessor = ViTImageProcessor.from_pretrained(
    "google/vit-base-patch16-224", size=224, device='cuda'
)

SAMPLE_EVERY_SEC = 2
video_path = 'videos/news2.mp4'
cap = cv2.VideoCapture(video_path)
n_frames = cap.get(cv2.CAP_PROP_FRAME_COUNT)
fps = cap.get(cv2.CAP_PROP_FPS)
video_len = n_frames / fps
print(f'Video length {video_len:.2f} seconds!')

frames = []
last_collected = -1
while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break

    timestamp = cap.get(cv2.CAP_PROP_POS_MSEC)
    second = timestamp // 1000

    if second % SAMPLE_EVERY_SEC == 0 and second != last_collected:
        last_collected = second
        frames.append(frame)
features = preprocessor(images=frames, return_tensors="pt")["pixel_values"]
print(features.shape)


model = SummaryModel.load_from_checkpoint('summary.ckpt')
model.to('cuda')
model.eval()


features = features.to('cuda')
y_pred = []
for frame in tqdm(features):
    y_p = model(frame.unsqueeze(0))
    y_p = torch.sigmoid(y_p)

    y_pred.append(y_p.cpu().detach().numpy().squeeze())

y_pred = np.array(y_pred)
sns.displot(y_pred)

THRESHOLD = 0.67
total_secs = 0

for i, y_p in enumerate(y_pred):
    if y_p >= THRESHOLD:
        print(i * SAMPLE_EVERY_SEC)
        total_secs += SAMPLE_EVERY_SEC

total_secs


clip = VideoFileClip(video_path)
subclips = []
for i, y_p in enumerate(y_pred):
    sec = i * SAMPLE_EVERY_SEC

    if y_p >= THRESHOLD:
        subclip = clip.subclip(sec, sec + SAMPLE_EVERY_SEC)
        subclips.append(subclip)
result = concatenate_videoclips(subclips)
result.write_videofile("videos/result.mp4")
result.ipython_display(width=640, maxduration=240)

 

6. model

      

      

      

    
    

      <python />
    
import argparse
from collections import defaultdict

import numpy as np
import torch
import torch.nn as nn
from pytorch_lightning.core.lightning import LightningModule
from torch import optim
# from torchmetrics import F1
from transformers import ViTModel


class SummaryModel(LightningModule):
    def __init__(self, hidden_dim=768, individual_logs=None):
        super().__init__()
        self.vit = ViTModel.from_pretrained("google/vit-base-patch16-224-in21k")
        self.scorer = nn.Linear(hidden_dim, 1)
        # self.sigmoid = nn.Sigmoid()
        self.loss = nn.BCEWithLogitsLoss()
        # self.train_f1 = F1()
        # self.val_f1 = F1()
        # self.test_f1 = F1()
        self.individual_logs = individual_logs
        self.tta_logs = defaultdict(list)

    def forward(self, x):
        x = self.vit(x).pooler_output
        x = self.scorer(x)
        # x = self.sigmoid(x)
        return x

    def run_batch(self, batch, batch_idx, metric, training=False):
        video_name, image_features, labels = batch
        video_name = video_name[0]
        image_features = image_features.squeeze(0)
        labels = labels.squeeze(0)

        # Score - aggregated labels.
        score = torch.sum(labels, dim=0)
        score = torch.min(
            score,
            torch.ones(
                score.shape[0],
            ).to(score.device),
        )
        out = self(image_features).squeeze(1)
        try:
            loss = self.loss(out.double(), score)
            preds = (torch.sigmoid(out) > 0.7).int()
            metric.update(preds, score.int())
            f1 = metric.compute()
            tp, fp, tn, fn = metric._get_final_stats()
            self.tta_logs[video_name].append((tp.item(), fp.item(), fn.item()))
        except Exception as e:
            print(e)
            loss = 0
        return loss

    def training_step(self, batch, batch_idx):
        loss = self.run_batch(batch, batch_idx, self.train_f1, training=True)
        self.log("train_loss", loss)
        return loss

    def training_epoch_end(self, training_step_outputs):
        self.log("train_f1", self.train_f1.compute())
        self.train_f1.reset()

    def validation_step(self, batch, batch_idx):
        loss = self.run_batch(batch, batch_idx, self.val_f1)
        self.log("val_loss", loss)
        return loss

    def validation_epoch_end(self, validation_step_outputs):
        self.log("val_f1", self.val_f1.compute())
        self.val_f1.reset()

    def test_step(self, batch, batch_idx):
        loss = self.run_batch(batch, batch_idx, self.test_f1)
        self.log("test_loss", loss)
        return loss

    def test_epoch_end(self, outputs):
        f1 = self.test_f1.compute()
        self.log("test_f1", f1)
        tp, fp, tn, fn = self.test_f1._get_final_stats()
        print(f"\nTest f1: {f1}, TP: {tp}, FP: {fp}, TN: {tn}, fn: {fn}")
        self.test_f1.reset()

    def configure_optimizers(self):
        optimizer = torch.optim.AdamW(self.parameters(), lr=1e-4)
        return optimizer

 

 

출처

빵형의 개발도상국 - YouTube