app.py

import os
import tempfile
import shutil
import torch
import gradio as gr
from types import SimpleNamespace
from huggingface_hub import hf_hub_download

# Import your original pipeline
import infer_videos_txt as tall_infer


# ================================
# CONFIGURE YOUR MODEL REPO HERE
# ================================
MODEL_REPO = "guard2PFE/tall4deepfake-weights"   # <-- change if needed
MODEL_FILE = "model_best.pth"                   # <-- change if needed


# ================================
# Build args object (like CLI)
# ================================
def build_args(ckpt_path):
    return SimpleNamespace(
        video_list="",
        initial_checkpoint=ckpt_path,
        dataset="ffpp",
        model="TALL_SWIN",
        device="cuda" if torch.cuda.is_available() else "cpu",
        num_workers=0,

        duration=4,
        frames_per_group=1,
        num_clips=8,
        num_crops=1,
        thumbnail_rows=2,
        input_size=224,

        threshold=0.5,

        disable_scaleup=False,
        threed_data=False,
        dense_sampling=True,
        augmentor_ver="v1",
        scale_range=[256, 320],
        modality="rgb",
        use_lmdb=False,

        hpe_to_token=False,
        rel_pos=False,
        window_size=7,
        no_token_mask=False,

        drop=0.0,
        drop_path=0.1,
        drop_block=None,
        use_checkpoint=False,

        dist_url="env://",
        world_size=1,
        local_rank=None,

        output_json="",
        output_csv="",
    )


# ================================
# Load model once (global cache)
# ================================
META = None
ARGS = None

def ensure_model_loaded():
    global META, ARGS

    if META is not None:
        return

    print("Downloading checkpoint...")
    ckpt_path = hf_hub_download(
        repo_id=MODEL_REPO,
        filename=MODEL_FILE
    )

    ARGS = build_args(ckpt_path)
    META = tall_infer.build_model_and_augmentor(ARGS)


# ================================
# Inference function
# ================================
@torch.no_grad()
def predict(video, threshold):

    ensure_model_loaded()
    ARGS.threshold = float(threshold)

    if isinstance(video, dict):
        video_path = video["name"]
    else:
        video_path = video

    if not os.path.isfile(video_path):
        return {"error": "Video not found"}

    tmp_dir = tempfile.mkdtemp(prefix="tall_space_")

    try:
        tmp_info = tall_infer.build_tmp_dataset_from_video(
            video_path,
            tmp_dir,
            image_tmpl=META["image_tmpl"]
        )

        result = tall_infer.infer_one_video_from_tmp(
            ARGS,
            META,
            tmp_dir,
            tmp_info["list_rel"],
            image_tmpl=META["image_tmpl"]
        )

        return {
            "video": os.path.basename(video_path),
            "frames": tmp_info["nframes"],
            **result,
            "device": str(META["device"])
        }

    except Exception as e:
        return {"error": str(e)}

    finally:
        shutil.rmtree(tmp_dir, ignore_errors=True)


# ================================
# Gradio UI
# ================================
demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.Video(label="Upload Video"),
        gr.Slider(0, 1, value=0.5, step=0.01, label="Threshold")
    ],
    outputs=gr.JSON(),
    title="TALL4Deepfake Detector",
    description="Video-level deepfake detection using TALL-SWIN"
)

if __name__ == "__main__":
    demo.launch()