app.py

import gradio as gr
import numpy as np
from PIL import Image

import torch

# Hugging Face Hub helpers
from huggingface_hub import hf_hub_download

# FastAI types (needed if your exported Learner references custom transforms)
from fastai.vision.all import ItemTransform, PILImage, PILMask

try:
    import albumentations as A  # noqa: F401
except Exception:
    # If your learner was exported with Albumentations objects inside, you need this installed.
    A = None

# -----------------------------
# Config
# -----------------------------
REPO_ID = "magomerob/resnet-segmentation"
EXAMPLE_IMAGE = "bicho.jpg"

_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


class SegmentationAlbumentationsTransform(ItemTransform):
    """Defined to satisfy pickles exported from notebooks (__main__).

    If your exported Learner includes this transform in its pipeline, it must exist
    (same name) when loading `model.pkl`.
    """

    split_idx = 0

    def __init__(self, aug):
        self.aug = aug

    def encodes(self, x):
        img, mask = x
        aug = self.aug(image=np.array(img), mask=np.array(mask))
        return PILImage.create(aug["image"]), PILMask.create(aug["mask"])


_LEARN = None


def _load_learner_from_hub():
    """Load fastai Learner from Hub with a clearer error if unpickling fails."""
    global _LEARN

    if _LEARN is not None:
        return _LEARN

    pkl_path = hf_hub_download(repo_id=REPO_ID, filename="model.pkl")

    try:
        # torch.load will surface the *real* unpickling error (missing class/import, etc.)
        learn = torch.load(pkl_path, map_location="cpu")
    except Exception as e:
        raise RuntimeError(
            "Failed to load 'model.pkl'. This is usually caused by missing custom "
            "classes/functions used when exporting the Learner (e.g., custom ItemTransform) "
            "or missing dependencies (e.g., albumentations).\n\n"
            f"Original error: {type(e).__name__}: {e}"
        ) from e

    # Move to device and eval
    learn.model.to(_device)
    learn.model.eval()

    _LEARN = learn
    return _LEARN


def _ensure_pil_rgb(x) -> Image.Image:
    if isinstance(x, np.ndarray):
        return Image.fromarray(x).convert("RGB")
    return x.convert("RGB")


def _build_palette(n_classes: int) -> np.ndarray:
    """Deterministic bright-ish palette, class 0 = black."""
    rng = np.random.default_rng(0)
    pal = rng.integers(0, 256, size=(max(n_classes, 1), 3), dtype=np.uint8)
    if len(pal) > 0:
        pal[0] = np.array([0, 0, 0], dtype=np.uint8)
    return pal


def _colorize_mask(mask: np.ndarray, palette: np.ndarray) -> np.ndarray:
    # mask: (H, W) int
    h, w = mask.shape
    out = np.zeros((h, w, 3), dtype=np.uint8)
    m = np.clip(mask, 0, len(palette) - 1)
    out[:] = palette[m]
    return out


def segment_image(input_image, overlay_alpha=0.5, output_mode="overlay"):
    """Run semantic segmentation and return either overlay or raw mask (colored)."""
    try:
        learn = _load_learner_from_hub()
    except Exception as e:
        raise gr.Error(str(e))

    img = _ensure_pil_rgb(input_image)

    # FastAI predict returns (pred_class, pred, probs)
    # For segmentation, pred is usually a TensorMask (H, W)
    with torch.no_grad():
        _, pred, _ = learn.predict(img)

    # pred can be a TensorMask / Tensor
    mask = pred.detach().cpu().numpy().astype(np.int64)

    # Infer number of classes from vocab/config if possible
    n_classes = None
    try:
        n_classes = len(learn.dls.vocab)
    except Exception:
        pass
    if not n_classes:
        n_classes = int(mask.max()) + 1 if mask.size else 1

    palette = _build_palette(n_classes)
    mask_rgb = _colorize_mask(mask, palette)

    if output_mode == "mask":
        return Image.fromarray(mask_rgb)

    # Overlay
    img_np = np.array(img, dtype=np.uint8)
    a = float(overlay_alpha)
    a = 0.0 if a < 0 else (1.0 if a > 1 else a)
    blended = (img_np * (1 - a) + mask_rgb * a).astype(np.uint8)
    return Image.fromarray(blended)


with gr.Blocks() as demo:
    gr.Markdown(
        f"# Semantic Segmentation\nModel: `{REPO_ID}`\n\nUpload an image and get a segmentation mask (or an overlay)."
    )

    with gr.Row():
        inp = gr.Image(type="numpy", label="Input image")
        out = gr.Image(type="pil", label="Output")

    with gr.Row():
        overlay_alpha = gr.Slider(0, 1, step=0.05, value=0.5, label="Overlay alpha")
        output_mode = gr.Radio(["overlay", "mask"], value="overlay", label="Output mode")

    btn = gr.Button("Run segmentation")

    btn.click(
        fn=segment_image,
        inputs=[inp, overlay_alpha, output_mode],
        outputs=[out],
    )

    gr.Examples(
        examples=[[EXAMPLE_IMAGE, 0.5, "overlay"], [EXAMPLE_IMAGE, 0.5, "mask"]],
        inputs=[inp, overlay_alpha, output_mode],
    )


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