main.py

import os
import uuid
import pickle
from typing import List

try:
    from fastapi import FastAPI, File, Form, UploadFile, HTTPException
    from fastapi.middleware.cors import CORSMiddleware
except ImportError as exc:
    raise ImportError(
        "FastAPI is required to run this application. Install it with 'pip install fastapi'."
    ) from exc

import torch
import torchvision.models as models
import torchvision.transforms as transforms
from PIL import Image
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split  # Added for accuracy scoring
import io

# ── App Initialization ───────────────────────────────────────────────────────
app = FastAPI(title="Teachable Machine Backend")

# ── CORS Configuration ───────────────────────────────────────────────────────
# Enables file uploads and API calls from frontend (running on different origin/port)
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],  # Allow requests from any origin
    allow_credentials=True,
    allow_methods=["*"],  # Allow all HTTP methods
    allow_headers=["*"],  # Allow all headers
)

DATASET_DIR = os.path.join(os.path.dirname(__file__), "dataset")
MODEL_PATH  = os.path.join(os.path.dirname(__file__), "model.pkl")


# ── Shared ML Setup (runs once at startup) ───────────────────────────────────
# Loading the model once here means every request reuses the same object in
# memory instead of reloading it from disk each time — much faster.
device = torch.device("cpu")

backbone = models.mobilenet_v3_small(weights=models.MobileNet_V3_Small_Weights.DEFAULT)
# Remove the final classifier layer — we only want feature extraction.
# The 960 numbers it outputs describe the image content without predicting a category.
backbone.classifier = torch.nn.Identity()
backbone.eval()  # Disables dropout — we are inferring, not training the backbone

# These values MUST be identical during training and prediction.
# 224x224 = the size MobileNetV3 was designed for.
# mean/std = ImageNet dataset statistics the model was pre-trained on.
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]
    )
])


# ── Helper: Extract features from a PIL image ────────────────────────────────
def extract_features(pil_image: Image.Image) -> np.ndarray:
    """
    Passes an image through MobileNetV3 and returns a 960-number feature vector.
    Used by both /train and /predict to guarantee identical preprocessing.
    """
    image = pil_image.convert("RGB")   # Handles RGBA/grayscale images safely
    tensor = transform(image)
    tensor = tensor.unsqueeze(0)       # (3,224,224) → (1,3,224,224) — adds batch dim

    with torch.no_grad():              # No gradients needed — saves memory & time
        features = backbone(tensor)

    return features.squeeze().numpy() # (1,960) → (960,) numpy array for sklearn


# ── Health Check ─────────────────────────────────────────────────────────────
@app.get("/")
def health_check():
    return {"status": "Backend is running!"}


# ── Milestone 1: Upload images ───────────────────────────────────────────────
@app.post("/upload-sample")
async def upload_sample(
    class_name: str = Form(...),
    files: List[UploadFile] = File(...)
):
    """
    Accepts a class label + a batch of images.
    Saves each image into dataset/<class_name>/ with a random UUID filename.
    """
    class_name = class_name.strip().replace(" ", "_")

    if not class_name:
        raise HTTPException(status_code=400, detail="class_name cannot be empty.")

    class_folder = os.path.join(DATASET_DIR, class_name)
    os.makedirs(class_folder, exist_ok=True)

    if not files:
        raise HTTPException(status_code=400, detail="At least one image file is required.")

    saved_files = []

    for file in files:
        if not file.content_type.startswith("image/"):
            raise HTTPException(
                status_code=400,
                detail=f"File '{file.filename}' is not an image. Only image files are accepted."
            )

        extension = os.path.splitext(file.filename)[1] or ".jpg"
        random_filename = f"{uuid.uuid4()}{extension}"
        save_path = os.path.join(class_folder, random_filename)

        contents = await file.read()
        with open(save_path, "wb") as f:
            f.write(contents)

        saved_files.append(random_filename)

    return {
        "message": f"Uploaded {len(saved_files)} image(s) to class '{class_name}'",
        "class": class_name,
        "saved_files": saved_files
    }


# ── Milestone 1 Bonus: Dataset info ─────────────────────────────────────────
@app.get("/dataset-info")
def dataset_info():
    if not os.path.exists(DATASET_DIR):
        return {"classes": {}, "total_images": 0}

    summary = {}
    for class_name in os.listdir(DATASET_DIR):
        class_path = os.path.join(DATASET_DIR, class_name)
        if os.path.isdir(class_path):
            summary[class_name] = len(os.listdir(class_path))

    return {
        "classes": summary,
        "total_images": sum(summary.values())
    }


# ── Milestone 2: Train model ─────────────────────────────────────────────────
@app.post("/train")
def train_model():
    """
    Scans dataset/, extracts MobileNetV3 features from every image,
    trains a LogisticRegression classifier, and saves it to model.pkl.
    """

    # ── Step 1: Validate dataset exists ──────────────────────────────────────
    if not os.path.exists(DATASET_DIR):
        raise HTTPException(
            status_code=400,
            detail="No dataset found. Please upload images first."
        )

    classes = [
        d for d in os.listdir(DATASET_DIR)
        if os.path.isdir(os.path.join(DATASET_DIR, d))
    ]

    # Classifier needs at least 2 classes — it learns to DISTINGUISH between them.
    # With only 1 class there is nothing to distinguish.
    if len(classes) < 2:
        raise HTTPException(
            status_code=400,
            detail=f"Need at least 2 classes to train. You currently have: {classes}"
        )

    X = []  # Feature vectors — one row per image
    y = []  # Labels         — one entry per image, matched by index to X

    # ── Step 2: Extract features from every image ────────────────────────────
    for class_name in classes:
        class_folder = os.path.join(DATASET_DIR, class_name)
        image_files  = os.listdir(class_folder)

        if len(image_files) == 0:
            continue  # Skip empty class folders silently

        for filename in image_files:
            image_path = os.path.join(class_folder, filename)
            try:
                img      = Image.open(image_path)
                features = extract_features(img)
                X.append(features)
                y.append(class_name)
            except Exception as e:
                # One corrupted image should not kill the whole training run
                print(f"Skipping {filename}: {e}")
                continue

    if len(X) == 0:
        raise HTTPException(
            status_code=400,
            detail="No valid images found in dataset."
        )

    X = np.array(X)  # Shape: (num_images, 960)
    y = np.array(y)  # Shape: (num_images,)

    # ── Step 3: Train the classifier ─────────────────────────────────────────
    
    # NEW: Split the data to calculate a real accuracy metric.
    # We added a safety net: if there are fewer than 5 images total, we test 
    # on the training data so it doesn't crash during a live presentation.
    if len(X) >= 5:
        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
    else:
        X_train, X_test, y_train, y_test = X, X, y, y

    # Why LogisticRegression?
    # MobileNetV3 already converted images into meaningful 960-number vectors.
    # LogisticRegression just finds the boundary between those vectors.
    # It trains in under a second, works with very few images, and needs no GPU.
    # max_iter=1000 prevents ConvergenceWarning on small datasets.
    classifier = LogisticRegression(max_iter=1000)
    classifier.fit(X_train, y_train)
    
    # Calculate overall accuracy
    accuracy = classifier.score(X_test, y_test)

    # ── Step 4: Save classifier + class list to disk ─────────────────────────
    # We save classes explicitly so the /predict endpoint can map
    # numeric outputs back to human-readable label names.
    model_data = {
        "classifier": classifier,
        "classes": classes
    }
    with open(MODEL_PATH, "wb") as f:
        pickle.dump(model_data, f)

    return {
        "message": "Training complete!",
        "classes": classes,
        "total_images": len(X),
        "accuracy": round(accuracy * 100, 2),  # Returned safely to the frontend!
        "model_saved_at": MODEL_PATH
    }


# ── Milestone 3: Predict endpoint ────────────────────────────────────────────
@app.post("/predict")
async def predict(file: UploadFile = File(...)):
    """
    Accepts a single image, runs it through MobileNetV3 + the trained
    LogisticRegression classifier, and returns the predicted class
    with a confidence score for every class.
    """

    # ── Step 1: Check model exists ────────────────────────────────────────────
    # If the user hits /predict before ever running /train, model.pkl won't
    # exist yet. We catch this early with a clear message instead of a crash.
    if not os.path.exists(MODEL_PATH):
        raise HTTPException(
            status_code=400,
            detail="No trained model found. Please call /train first."
        )

    # ── Step 2: Validate the uploaded file is an image ────────────────────────
    if not file.content_type.startswith("image/"):
        raise HTTPException(
            status_code=400,
            detail=f"File '{file.filename}' is not an image. Only image files are accepted."
        )

    # ── Step 3: Load the saved model from disk ────────────────────────────────
    # We reload model.pkl on every prediction request.
    # Why not load it once at startup like the backbone?
    # Because model.pkl gets replaced every time /train is called.
    # If we cached it at startup, predictions would use the OLD model
    # even after the user retrains — a subtle but serious bug.
    with open(MODEL_PATH, "rb") as f:
        model_data = pickle.load(f)

    classifier = model_data["classifier"]
    classes    = model_data["classes"]

    # ── Step 4: Read and decode the uploaded image ────────────────────────────
    # file.read() gives us raw bytes. We wrap them in BytesIO so PIL
    # can treat the bytes like a file on disk — no temp file needed.
    contents = await file.read()
    image = Image.open(io.BytesIO(contents))

    # ── Step 5: Extract features using the SAME function used during training ─
    # This is the most important consistency rule in the whole project.
    # If training used 224x224 + ImageNet normalization, prediction MUST too.
    # extract_features() guarantees this since both phases call the same code.
    features = extract_features(image)

    # ── Step 6: Run prediction ────────────────────────────────────────────────
    # features is shape (960,) — we reshape to (1, 960) because sklearn
    # expects a 2D array: (number_of_samples, number_of_features)
    features_2d = features.reshape(1, -1)

    # predict() returns the winning class label e.g. ["cat"]
    predicted_class = classifier.predict(features_2d)[0]

    # predict_proba() returns confidence scores for ALL classes e.g. [0.82, 0.18]
    # Each number = how confident the model is that this image belongs to that class.
    # They always sum to 1.0 (100%).
    probabilities = classifier.predict_proba(features_2d)[0]

    # ── Step 7: Build a clean confidence scores dict ──────────────────────────
    # zip(classes, probabilities) pairs each class name with its score:
    # e.g. {"cat": 0.82, "dog": 0.18}
    # round(..., 4) keeps it readable: 0.8173 instead of 0.81734521938...
    # float() converts numpy float32 → Python float so JSON can serialize it
    confidence_scores = {
        cls: round(float(prob), 4)
        for cls, prob in zip(classifier.classes_, probabilities)
    }

    return {
        "predicted_class": predicted_class,
        "confidence": round(float(max(probabilities)), 4),
        "all_scores": confidence_scores
    }