Ucf I3d Model By Block Lr 0.001

Smart Surveillance System

we leveraged a pre-trained I3D model and fine-tuned it using two strategies:

Block-level tuning Adjusting and retraining groups of layers (blocks) to adapt the model to the new dataset.

Layer-level tuning Fine-tuning specific layers for more granular control over feature learning.

The final classification layer of the I3D model was removed and replaced with a custom output layer tailored to our binary classification task: predicting whether an activity represents a crime (1) or non-crime (0).

How Run

import torch
import torch.nn as nn


class UCFModel(nn.Module):
    def __init__(self, model_name="i3d_r50"):
        super().__init__()
        self.model_name = model_name

        self.model = torch.hub.load("facebookresearch/pytorchvideo", model_name, pretrained=True)

        in_features = self.model.blocks[-1].proj.in_features
        self.model.blocks[-1].proj = nn.Linear(in_features, 2)

    def forward(self, frames):
        return self.model(frames)

import torch
from PIL import Image
from huggingface_hub import hf_hub_download
from torchvision import transforms


inference_transform = transforms.Compose(
    [
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)


class UCFInferenceByFrames:
    def __init__(self, repo_id):
        self.repo_id = repo_id

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

    def load_model(self):
        model_path = hf_hub_download(repo_id=self.repo_id, filename="ucf_model.pth")
        state_dict = torch.load(model_path)

        model = UCFModel().to(device=self.device)
        model.load_state_dict(state_dict)
        model.eval()

        return model

    def inference(self, frames):
        video_tensor_list = []
        for frame in frames:
            frame_pil = Image.fromarray(frame)
            frame_tensor = inference_transform(frame_pil)
            video_tensor_list.append(frame_tensor)

        video_tensor = torch.stack(video_tensor_list)
        video_tensor = video_tensor.permute(1, 0, 2, 3).unsqueeze(0).float()

        video_tensor = video_tensor.to(self.device)

        with torch.no_grad():
            output = self.model(video_tensor)

        return output.argmax(1)

import cv2 as cv
import numpy as np

ucf = UCFInferenceByFrames("amjad-awad/ucf-i3d-model-by-block-lr-0.001")

def inference(ucf_model, video_path, max_frames=16):
    cap = cv.VideoCapture(video_path)

    if not cap.isOpened():
        print("No video")
        return
    
    frames = []

    while True:
        ret, frame = cap.read()

        if not ret:
            break

        frames.append(frame)
    
    length = len(frames)
    indices = np.linspace(0, length - 1, max_frames, dtype=int)
    frames = [frames[i] for i in indices]
    predict = ucf_model.inference(frames)

    return "Crime" if int(predict) == 1 else "No-Crime"

predict = inference(ucf_model=ucf, video_path="YOUR_VIDEO_PATH.mp4")
print(predict)