inference.py

"""
Inference script for SAH-DRY-ADS-v1 (Sub-Saharan Drylands Species Classifier)

This model classifies 328 categories across eastern and southern African ecosystems,
with taxonomic fallback for uncertain species-level predictions. Trained on 2.8+ million
camera trap images from savannas, dry forests, arid shrublands, and semi-desert habitats
across 9 countries. All training data is open-source via LILA BC (https://lila.science/).

Model: Sub-Saharan Drylands v1
Input: Variable size (extracted from checkpoint, typically 480x480)
Framework: PyTorch (EfficientNet V2 Medium architecture)
Classes: 328 species and higher-level taxa with taxonomic fallback
Developer: Addax Data Science
Citation: https://joss.theoj.org/papers/10.21105/joss.05581
License: CC BY-NC-SA 4.0
Info: https://addaxdatascience.com/

Training regions: South Africa, Tanzania, Kenya, Mozambique, Botswana, Namibia,
                  Rwanda, Madagascar, Uganda

Author: Peter van Lunteren
Created: 2026-01-14
"""

from __future__ import annotations

import pathlib
import platform
from pathlib import Path

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image, ImageFile, ImageOps
from torchvision import transforms
from torchvision.models import efficientnet

# Don't freak out over truncated images
ImageFile.LOAD_TRUNCATED_IMAGES = True

# Make sure Windows-trained models work on Unix
plt = platform.system()
if plt != 'Windows':
    pathlib.WindowsPath = pathlib.PosixPath


class EfficientNetV2M(nn.Module):
    """EfficientNet V2 Medium architecture for wildlife classification."""

    def __init__(self, num_classes: int, tune: bool = True):
        super(EfficientNetV2M, self).__init__()
        self.avgpool = nn.AdaptiveAvgPool2d(1)
        self.model = efficientnet.efficientnet_v2_m(
            weights=efficientnet.EfficientNet_V2_M_Weights.DEFAULT
        )
        if tune:
            for params in self.model.parameters():
                params.requires_grad = True
        num_ftrs = self.model.classifier[1].in_features
        self.model.classifier[1] = nn.Linear(in_features=num_ftrs, out_features=num_classes)

    def forward(self, x):
        x = self.model.features(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        prediction = self.model.classifier(x)
        return prediction

class ModelInference:
    """PyTorch inference implementation for Sub-Saharan Drylands species classifier."""

    def __init__(self, model_dir: Path, model_path: Path):
        """
        Initialize with model paths.

        Args:
            model_dir: Directory containing model files
            model_path: Path to sub_saharan_drylands_v1.pt checkpoint file
        """
        self.model_dir = model_dir
        self.model_path = model_path
        self.model = None
        self.device = None
        self.image_size = None
        self.classes = []
        self.preprocess = None

    def check_gpu(self) -> bool:
        """
        Check GPU availability for PyTorch inference.

        Checks both Apple Metal Performance Shaders (MPS) and CUDA availability.

        Returns:
            True if GPU available, False otherwise
        """
        # Check Apple MPS (Apple Silicon)
        try:
            if torch.backends.mps.is_built() and torch.backends.mps.is_available():
                return True
        except Exception:
            pass

        # Check CUDA (NVIDIA)
        return torch.cuda.is_available()

    def load_model(self, device_str: str = 'cpu') -> None:
        """
        Load PyTorch model from checkpoint.

        The checkpoint contains:
        - model: State dict with trained weights
        - categories: Dict mapping class names to indices
        - image_size: Tuple with input dimensions

        Args:
            device_str: Device to load model on ('cpu', 'cuda', or 'mps')

        Raises:
            RuntimeError: If model loading fails
            FileNotFoundError: If model_path is invalid
        """
        if not self.model_path.exists():
            raise FileNotFoundError(f"Model file not found: {self.model_path}")

        try:
            # Set device
            self.device = torch.device(device_str)

            # Load checkpoint
            checkpoint = torch.load(str(self.model_path), map_location=self.device)

            # Extract metadata
            self.image_size = tuple(checkpoint['image_size'])
            categories = checkpoint['categories']
            self.classes = list(categories.keys())

            # Initialize EfficientNet V2 Medium architecture
            num_classes = len(self.classes)
            self.model = EfficientNetV2M(num_classes, tune=False)

            # Load weights
            self.model.load_state_dict(checkpoint['model'])
            self.model.to(self.device)
            self.model.eval()

            # Setup preprocessing
            self.preprocess = transforms.Compose([
                transforms.Resize(self.image_size),
                transforms.ToTensor(),
            ])

        except Exception as e:
            raise RuntimeError(f"Failed to load PyTorch model from {self.model_path}: {e}") from e

    def get_crop(
        self, image: Image.Image, bbox: tuple[float, float, float, float]
    ) -> Image.Image:
        """
        Crop image using model-specific preprocessing.

        This cropping method was developed by Dan Morris for MegaDetector and is
        designed to:
        1. Square the bounding box (max of width/height)
        2. Add padding to prevent over-enlargement of small animals
        3. Center the detection within the crop
        4. Pad with black (0) to maintain square aspect ratio

        Args:
            image: PIL Image (full resolution)
            bbox: Normalized bounding box (x, y, width, height) in range [0.0, 1.0]

        Returns:
            Cropped and padded PIL Image ready for classification

        Raises:
            ValueError: If bbox is invalid (zero size)
        """
        img_w, img_h = image.size

        # Denormalize bbox coordinates
        xmin = int(bbox[0] * img_w)
        ymin = int(bbox[1] * img_h)
        box_w = int(bbox[2] * img_w)
        box_h = int(bbox[3] * img_h)

        # Square the box (use max dimension)
        box_size = max(box_w, box_h)

        # Add padding (prevents over-enlargement of small animals)
        box_size = self._pad_crop(box_size)

        # Center the detection within the squared crop
        xmin = max(0, min(xmin - int((box_size - box_w) / 2), img_w - box_w))
        ymin = max(0, min(ymin - int((box_size - box_h) / 2), img_h - box_h))

        # Clip to image boundaries
        box_w = min(img_w, box_size)
        box_h = min(img_h, box_size)

        if box_w == 0 or box_h == 0:
            raise ValueError(f"Invalid bbox size: {box_w}x{box_h}")

        # Crop and pad to square
        crop = image.crop(box=[xmin, ymin, xmin + box_w, ymin + box_h])
        crop = ImageOps.pad(crop, size=(box_size, box_size), color=0)

        return crop

    def _pad_crop(self, box_size: int) -> int:
        """
        Calculate padded crop size to prevent over-enlargement of small animals.

        Standard network input is 224x224. This function ensures small detections
        aren't excessively upscaled while adding consistent padding to larger detections.

        Args:
            box_size: Original bounding box size (max of width/height)

        Returns:
            Padded box size
        """
        input_size_network = 224
        default_padding = 30

        if box_size >= input_size_network:
            # Large detection: add default padding
            return box_size + default_padding
        else:
            # Small detection: ensure minimum size without excessive enlargement
            diff_size = input_size_network - box_size
            if diff_size < default_padding:
                return box_size + default_padding
            else:
                return input_size_network

    def get_classification(self, crop: Image.Image) -> list[list[str, float]]:
        """
        Run PyTorch classification on cropped image.

        Args:
            crop: Cropped and preprocessed PIL Image

        Returns:
            List of [class_name, confidence] lists for ALL classes, in model order.
            Example: [["lion", 0.85], ["leopard", 0.10], ["cheetah", 0.02], ...]
            NOTE: Sorting by confidence is handled by classification_worker.py

        Raises:
            RuntimeError: If model not loaded or inference fails
        """
        if self.model is None:
            raise RuntimeError("Model not loaded - call load_model() first")

        try:
            # Preprocess image (resize and convert to tensor)
            input_tensor = self.preprocess(crop)
            input_batch = input_tensor.unsqueeze(0)  # Add batch dimension
            input_batch = input_batch.to(self.device)

            # Run inference
            with torch.no_grad():
                output = self.model(input_batch)

            # Apply softmax to get probabilities
            probabilities = F.softmax(output, dim=1)
            probabilities_np = probabilities.cpu().detach().numpy()
            confidence_scores = probabilities_np[0]

            # Build list of [class_name, confidence] pairs
            classifications = []
            for i in range(len(confidence_scores)):
                pred_class = self.classes[i]
                pred_conf = float(confidence_scores[i])
                classifications.append([pred_class, pred_conf])

            return classifications

        except Exception as e:
            raise RuntimeError(f"PyTorch classification failed: {e}") from e

    def get_class_names(self) -> dict[str, str]:
        """
        Get mapping of class IDs to species names.

        Returns:
            Dict mapping class ID (1-indexed string) to species/taxon name
            Example: {"1": "aardvark", "2": "african wild cat", ..., "328": "zebra"}

        Raises:
            RuntimeError: If model not loaded
        """
        if self.model is None:
            raise RuntimeError("Model not loaded - call load_model() first")

        try:
            # Create 1-indexed mapping of class IDs to names
            class_names = {}
            for i, class_name in enumerate(self.classes):
                class_id_str = str(i + 1)  # 1-indexed
                class_names[class_id_str] = class_name

            return class_names

        except Exception as e:
            raise RuntimeError(f"Failed to extract class names: {e}") from e

    def get_tensor(self, crop: Image.Image):
        """Preprocess a crop into a numpy array for batch inference."""
        tensor = self.preprocess(crop)
        return tensor.numpy()

    def classify_batch(self, batch):
        """Run inference on a batch of preprocessed numpy arrays."""
        tensor = torch.from_numpy(batch).to(self.device)
        with torch.no_grad():
            output = self.model(tensor)
            probs = F.softmax(output, dim=1).cpu().numpy()

        results = []
        for p in probs:
            classifications = [
                [self.classes[i], float(p[i])]
                for i in range(len(self.classes))
            ]
            results.append(classifications)
        return results