data/camera_augment.py

"""
Camera/Scene Augmentation — synthetic diversity for prototype LoRA training.

Generates diverse camera perspectives from a single image by simulating
different camera types, scenes, and conditions. Used in Stage 5 (prototype
LoRA training) to create varied conditioning inputs so each prototype
specializes on a different camera/scene combination.

Augmentations:
  - Crop position: simulate PTZ pan/tilt (random crop of original)
  - Brightness/contrast: simulate day/night/indoor/outdoor lighting
  - Color temperature: simulate warm (indoor) vs cool (outdoor) cameras
  - Noise: simulate low-light camera sensor noise
  - Blur: simulate focus/motion blur from cheap cameras
  - Resolution: simulate different camera resolutions

Reference: Train on augmented data so hypernetwork learns to differentiate
camera/scene contexts and generate specialized adapters.
"""

from __future__ import annotations

import random
from typing import Optional

import torch
import torch.nn.functional as F


class CameraAugmentor:
    """
    Applies random camera-specific augmentations to simulate diverse cameras.

    Args:
        img_size: Target output image size (square)
        crop_range: Min/max fraction of image to crop (simulates zoom level)
        brightness_range: Min/max brightness multiplier
        contrast_range: Min/max contrast multiplier
        noise_range: Min/max Gaussian noise std
        blur_prob: Probability of applying Gaussian blur
        color_temp_range: Min/max color temperature shift
    """

    def __init__(
        self,
        img_size: int = 448,
        crop_range: tuple[float, float] = (0.5, 1.0),
        brightness_range: tuple[float, float] = (0.5, 1.5),
        contrast_range: tuple[float, float] = (0.7, 1.3),
        noise_range: tuple[float, float] = (0.0, 0.05),
        blur_prob: float = 0.3,
        color_temp_range: tuple[float, float] = (-0.1, 0.1),
    ):
        self.img_size = img_size
        self.crop_range = crop_range
        self.brightness_range = brightness_range
        self.contrast_range = contrast_range
        self.noise_range = noise_range
        self.blur_prob = blur_prob
        self.color_temp_range = color_temp_range

    def __call__(self, image: torch.Tensor) -> tuple[torch.Tensor, dict]:
        """
        Apply random camera augmentation.

        Args:
            image: [3, H, W] or [B, 3, H, W] image tensor (0-1 range)
        Returns:
            augmented_image: same shape, augmented
            aug_params: dict describing which augmentations were applied
        """
        squeeze = False
        if image.dim() == 3:
            image = image.unsqueeze(0)
            squeeze = True

        B, C, H, W = image.shape
        aug_params = {}

        # 1. Random crop (simulates PTZ zoom/position)
        crop_frac = random.uniform(*self.crop_range)
        crop_h = int(H * crop_frac)
        crop_w = int(W * crop_frac)
        top = random.randint(0, H - crop_h)
        left = random.randint(0, W - crop_w)
        image = image[:, :, top:top + crop_h, left:left + crop_w]
        image = F.interpolate(image, size=(self.img_size, self.img_size), mode="bilinear", align_corners=False)
        aug_params["crop"] = {"frac": crop_frac, "top": top, "left": left}

        # 2. Brightness
        brightness = random.uniform(*self.brightness_range)
        image = image * brightness
        aug_params["brightness"] = brightness

        # 3. Contrast
        contrast = random.uniform(*self.contrast_range)
        mean = image.mean(dim=[2, 3], keepdim=True)
        image = (image - mean) * contrast + mean
        aug_params["contrast"] = contrast

        # 4. Color temperature shift
        temp = random.uniform(*self.color_temp_range)
        if abs(temp) > 0.01:
            # Warm: boost red, reduce blue. Cool: opposite.
            image[:, 0] += temp       # Red channel
            image[:, 2] -= temp       # Blue channel
        aug_params["color_temp"] = temp

        # 5. Gaussian noise (sensor noise)
        noise_std = random.uniform(*self.noise_range)
        if noise_std > 0.001:
            noise = torch.randn_like(image) * noise_std
            image = image + noise
        aug_params["noise_std"] = noise_std

        # 6. Gaussian blur (focus/motion blur)
        if random.random() < self.blur_prob:
            kernel_size = random.choice([3, 5])
            sigma = random.uniform(0.5, 2.0)
            image = _gaussian_blur(image, kernel_size, sigma)
            aug_params["blur"] = {"kernel": kernel_size, "sigma": sigma}
        else:
            aug_params["blur"] = None

        # Clamp to valid range
        image = image.clamp(0, 1)

        if squeeze:
            image = image.squeeze(0)

        return image, aug_params


class ScenePresetAugmentor:
    """
    Applies pre-defined scene augmentation presets for prototype diversity.

    Each preset simulates a specific camera installation scenario.
    """

    PRESETS = {
        "lobby_daytime": {"brightness": 1.2, "contrast": 1.0, "color_temp": 0.05, "noise": 0.01, "crop_frac": 0.8},
        "lobby_nighttime": {"brightness": 0.5, "contrast": 0.8, "color_temp": -0.05, "noise": 0.04, "crop_frac": 0.8},
        "parking_day": {"brightness": 1.3, "contrast": 1.2, "color_temp": 0.0, "noise": 0.01, "crop_frac": 0.6},
        "parking_night": {"brightness": 0.3, "contrast": 0.7, "color_temp": -0.08, "noise": 0.05, "crop_frac": 0.6},
        "gate_clear": {"brightness": 1.1, "contrast": 1.1, "color_temp": 0.02, "noise": 0.01, "crop_frac": 0.9},
        "gate_rain": {"brightness": 0.7, "contrast": 0.9, "color_temp": -0.03, "noise": 0.03, "crop_frac": 0.9},
        "corridor_empty": {"brightness": 0.9, "contrast": 1.0, "color_temp": 0.08, "noise": 0.02, "crop_frac": 1.0},
        "corridor_crowded": {"brightness": 0.9, "contrast": 1.0, "color_temp": 0.08, "noise": 0.02, "crop_frac": 0.7},
        "rooftop_twilight": {"brightness": 0.4, "contrast": 1.1, "color_temp": -0.1, "noise": 0.03, "crop_frac": 0.5},
        "rooftop_noon": {"brightness": 1.4, "contrast": 1.3, "color_temp": 0.0, "noise": 0.01, "crop_frac": 0.5},
    }

    def __init__(self, img_size: int = 448):
        self.img_size = img_size

    def augment(self, image: torch.Tensor, preset_name: str) -> torch.Tensor:
        """Apply a named scene preset augmentation."""
        preset = self.PRESETS.get(preset_name, self.PRESETS["lobby_daytime"])

        if image.dim() == 3:
            image = image.unsqueeze(0)

        B, C, H, W = image.shape

        # Crop
        frac = preset["crop_frac"]
        ch, cw = int(H * frac), int(W * frac)
        top = (H - ch) // 2
        left = (W - cw) // 2
        image = image[:, :, top:top + ch, left:left + cw]
        image = F.interpolate(image, size=(self.img_size, self.img_size), mode="bilinear", align_corners=False)

        # Brightness + contrast
        mean = image.mean(dim=[2, 3], keepdim=True)
        image = ((image - mean) * preset["contrast"] + mean) * preset["brightness"]

        # Color temperature
        image[:, 0] += preset["color_temp"]
        image[:, 2] -= preset["color_temp"]

        # Noise
        if preset["noise"] > 0:
            image = image + torch.randn_like(image) * preset["noise"]

        return image.clamp(0, 1).squeeze(0)

    @classmethod
    def list_presets(cls) -> list[str]:
        return list(cls.PRESETS.keys())


def _gaussian_blur(image: torch.Tensor, kernel_size: int, sigma: float) -> torch.Tensor:
    """Apply Gaussian blur to image tensor."""
    # Create 1D Gaussian kernel
    coords = torch.arange(kernel_size, dtype=torch.float32) - kernel_size // 2
    kernel_1d = torch.exp(-0.5 * (coords / sigma) ** 2)
    kernel_1d = kernel_1d / kernel_1d.sum()

    # Create 2D kernel
    kernel_2d = kernel_1d.unsqueeze(0) * kernel_1d.unsqueeze(1)
    kernel_2d = kernel_2d.unsqueeze(0).unsqueeze(0)  # [1, 1, K, K]
    kernel_2d = kernel_2d.expand(image.shape[1], -1, -1, -1).to(image.device)  # [C, 1, K, K]

    pad = kernel_size // 2
    return F.conv2d(image, kernel_2d, padding=pad, groups=image.shape[1])