app.py

# import gradio as gr
# import cv2
# import numpy as np
# from detectron2.config import get_cfg
# from detectron2.engine import DefaultPredictor
# from detectron2.utils.visualizer import Visualizer, ColorMode
# from detectron2.data import MetadataCatalog
# from huggingface_hub import hf_hub_download
# import os

# REPO_ID = os.getenv("MODEL_REPO_ID", "PUSHPENDAR/hrsid-ship-detection")

# os.makedirs("/app/hf_cache", exist_ok=True)

# print("Downloading model files...")
# MODEL_PATH  = hf_hub_download(repo_id=REPO_ID, filename="model_final.pth", cache_dir="/app/hf_cache")
# CONFIG_PATH = hf_hub_download(repo_id=REPO_ID, filename="config.yaml",     cache_dir="/app/hf_cache")
# print(f"Model: {MODEL_PATH} ✅")
# print(f"Config: {CONFIG_PATH} ✅")

# print("Loading Faster R-CNN model...")
# cfg = get_cfg()
# cfg.merge_from_file(CONFIG_PATH)
# cfg.MODEL.WEIGHTS    = MODEL_PATH
# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
# cfg.MODEL.DEVICE     = "cpu"

# MetadataCatalog.get("__unused").set(thing_classes=["ship"])
# predictor = DefaultPredictor(cfg)
# print("Model loaded ✅")


# def detect_ships(image, confidence_threshold):
#     if image is None:
#         return None, "Please upload an image."

#     cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = confidence_threshold
#     img_bgr  = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
#     outputs  = predictor(img_bgr)
#     instances = outputs["instances"].to("cpu")
#     keep     = instances.scores >= confidence_threshold
#     instances = instances[keep]

#     metadata = MetadataCatalog.get("__unused")
#     v   = Visualizer(img_bgr[:, :, ::-1], metadata=metadata, scale=1.0, instance_mode=ColorMode.IMAGE)
#     out = v.draw_instance_predictions(instances)

#     result_img = out.get_image()
#     num_ships  = len(instances)
#     scores     = instances.scores.tolist()

#     info = f"✅ Detected {num_ships} ship(s)\n"
#     if scores:
#         info += "Confidence scores: " + ", ".join([f"{s:.2f}" for s in scores])
#         if hasattr(instances, "pred_boxes"):
#             boxes = instances.pred_boxes.tensor.tolist()
#             info += "\n\nBounding boxes (x1,y1,x2,y2):\n"
#             for i, (box, score) in enumerate(zip(boxes, scores)):
#                 x1, y1, x2, y2 = [int(v) for v in box]
#                 info += f"  Ship {i+1}: [{x1},{y1},{x2},{y2}] conf={score:.2f}\n"
#     else:
#         info += "No ships detected above threshold."

#     return result_img, info


# with gr.Blocks(title="🚢 HRSID Ship Detection") as demo:
#     gr.Markdown("# 🚢 HRSID Ship Detection")
#     gr.Markdown("Upload a SAR image to detect ships using Faster R-CNN with ResNet-101, trained on HRSID dataset.")
#     with gr.Row():
#         with gr.Column():
#             image_input = gr.Image(type="pil", label="Upload SAR Image")
#             threshold   = gr.Slider(0.1, 0.9, value=0.5, step=0.05, label="Confidence Threshold")
#             btn         = gr.Button("Detect Ships", variant="primary")
#         with gr.Column():
#             image_output = gr.Image(type="numpy", label="Detection Result")
#             info_output  = gr.Textbox(label="Detection Info", lines=10)

#     btn.click(fn=detect_ships, inputs=[image_input, threshold], outputs=[image_output, info_output])

# if __name__ == "__main__":
#     demo.launch(server_name="0.0.0.0", server_port=7860)
import os
import tempfile
from copy import deepcopy

import cv2
import gradio as gr
import numpy as np
from detectron2.config import get_cfg
from detectron2.data import MetadataCatalog
from detectron2.engine import DefaultPredictor
from detectron2.utils.visualizer import ColorMode, Visualizer
from huggingface_hub import hf_hub_download

# ── Model loading ────────────────────────────────────────────────────────────

REPO_ID = os.getenv("MODEL_REPO_ID", "PUSHPENDAR/hrsid-ship-detection")

os.makedirs("/app/hf_cache", exist_ok=True)

print("Downloading model files...")
MODEL_PATH = hf_hub_download(
    repo_id=REPO_ID,
    filename="model_final.pth",
    cache_dir="/app/hf_cache",
    token=os.getenv("HF_TOKEN"),  # uses secret if set, else None (public repos)
)
CONFIG_PATH = hf_hub_download(
    repo_id=REPO_ID,
    filename="config.yaml",
    cache_dir="/app/hf_cache",
    token=os.getenv("HF_TOKEN"),
)
print(f"Model:  {MODEL_PATH} ✅")
print(f"Config: {CONFIG_PATH} ✅")

print("Loading Faster R-CNN model...")
_base_cfg = get_cfg()
_base_cfg.merge_from_file(CONFIG_PATH)
_base_cfg.MODEL.WEIGHTS = MODEL_PATH
_base_cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
_base_cfg.MODEL.DEVICE = "cpu"
_base_cfg.freeze()  # make it immutable so we always deepcopy before mutating

MetadataCatalog.get("__unused").set(thing_classes=["ship"])
print("Model loaded ✅")


# ── Helpers ──────────────────────────────────────────────────────────────────

def get_predictor(confidence_threshold: float) -> DefaultPredictor:
    """Return a fresh predictor with the requested threshold.
    deepcopy avoids mutating the global frozen cfg across concurrent requests.
    """
    cfg = deepcopy(_base_cfg)
    cfg.defrost()
    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = confidence_threshold
    return DefaultPredictor(cfg)


def run_inference(img_bgr: np.ndarray, confidence_threshold: float):
    """Run detection on a single BGR frame. Returns (result_bgr, instances)."""
    predictor = get_predictor(confidence_threshold)
    outputs = predictor(img_bgr)
    instances = outputs["instances"].to("cpu")
    instances = instances[instances.scores >= confidence_threshold]

    metadata = MetadataCatalog.get("__unused")
    v = Visualizer(
        img_bgr[:, :, ::-1],
        metadata=metadata,
        scale=1.0,
        instance_mode=ColorMode.IMAGE,
    )
    out = v.draw_instance_predictions(instances)
    result_rgb = out.get_image()  # H×W×3 RGB
    result_bgr = cv2.cvtColor(result_rgb, cv2.COLOR_RGB2BGR)
    return result_bgr, instances


def build_info(instances) -> str:
    num = len(instances)
    scores = instances.scores.tolist()
    info = f"✅ Detected {num} ship(s)\n"
    if scores:
        info += "Confidence scores: " + ", ".join([f"{s:.2f}" for s in scores])
        if hasattr(instances, "pred_boxes"):
            boxes = instances.pred_boxes.tensor.tolist()
            info += "\n\nBounding boxes (x1,y1,x2,y2):\n"
            for i, (box, score) in enumerate(zip(boxes, scores)):
                x1, y1, x2, y2 = [int(c) for c in box]
                info += f"  Ship {i+1}: [{x1},{y1},{x2},{y2}]  conf={score:.2f}\n"
    else:
        info += "No ships detected above threshold."
    return info


# ── Image tab ────────────────────────────────────────────────────────────────

def detect_ships_image(image, confidence_threshold):
    if image is None:
        return None, "Please upload an image."
    img_bgr = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
    result_bgr, inst = run_inference(img_bgr, confidence_threshold)
    result_rgb = cv2.cvtColor(result_bgr, cv2.COLOR_BGR2RGB)
    return result_rgb, build_info(inst)


# ── Video tab ────────────────────────────────────────────────────────────────

def detect_ships_video(video_path, confidence_threshold, progress=gr.Progress()):
    if video_path is None:
        return None, "Please upload a video."

    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        return None, "Could not open video file."

    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = cap.get(cv2.CAP_PROP_FPS) or 25
    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

    out_file = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
    out_path = out_file.name
    out_file.close()

    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
    writer = cv2.VideoWriter(out_path, fourcc, fps, (w, h))

    frame_idx = 0
    total_ships = 0
    max_per_frame = 0

    while True:
        ret, frame = cap.read()
        if not ret:
            break

        result_bgr, inst = run_inference(frame, confidence_threshold)
        writer.write(result_bgr)

        n = len(inst)
        total_ships += n
        max_per_frame = max(max_per_frame, n)
        frame_idx += 1

        if total_frames > 0:
            progress(
                frame_idx / total_frames,
                desc=f"Processing frame {frame_idx}/{total_frames}",
            )

    cap.release()
    writer.release()

    info = (
        f"✅ Video processed: {frame_idx} frames\n"
        f"Total ship detections across all frames: {total_ships}\n"
        f"Peak ships in a single frame: {max_per_frame}\n"
        f"FPS: {fps:.1f} | Resolution: {w}×{h}"
    )
    return out_path, info


# ── UI ───────────────────────────────────────────────────────────────────────

with gr.Blocks(title="🚢 HRSID Ship Detection") as demo:
    gr.Markdown("# 🚢 HRSID Ship Detection")
    gr.Markdown(
        "Detect ships in SAR images **or videos** using "
        "Faster R-CNN with ResNet-101, trained on the HRSID dataset."
    )

    with gr.Tabs():

        with gr.Tab("🖼️ Image Detection"):
            with gr.Row():
                with gr.Column():
                    img_input = gr.Image(type="pil", label="Upload SAR Image")
                    img_thresh = gr.Slider(
                        0.1, 0.9, value=0.5, step=0.05, label="Confidence Threshold"
                    )
                    img_btn = gr.Button("Detect Ships", variant="primary")
                with gr.Column():
                    img_output = gr.Image(type="numpy", label="Detection Result")
                    img_info = gr.Textbox(label="Detection Info", lines=10)

            img_btn.click(
                fn=detect_ships_image,
                inputs=[img_input, img_thresh],
                outputs=[img_output, img_info],
            )

        with gr.Tab("🎥 Video Detection"):
            gr.Markdown(
                "> ⚠️ CPU inference is slow. Short clips (< 30 s) are recommended."
            )
            with gr.Row():
                with gr.Column():
                    vid_input = gr.Video(label="Upload SAR Video")
                    vid_thresh = gr.Slider(
                        0.1, 0.9, value=0.5, step=0.05, label="Confidence Threshold"
                    )
                    vid_btn = gr.Button("Detect Ships in Video", variant="primary")
                with gr.Column():
                    vid_output = gr.Video(label="Detection Result Video")
                    vid_info = gr.Textbox(label="Detection Summary", lines=8)

            vid_btn.click(
                fn=detect_ships_video,
                inputs=[vid_input, vid_thresh],
                outputs=[vid_output, vid_info],
            )

if __name__ == "__main__":
    demo.queue()
    demo.launch(server_name="0.0.0.0", server_port=7860)  # NO share=True