app.py

import gradio as gr
import os
import torch
import tensorflow as tf
tf.config.set_visible_devices([], 'GPU')
import numpy as np
from PIL import Image
from PIL import Image as PILImage
from pathlib import Path
import matplotlib.pyplot as plt
import io
from skimage.io import imread
from skimage.color import rgb2gray
from csbdeep.utils import normalize
from stardist.models import StarDist2D
from stardist.plot import render_label
from MEDIARFormer import MEDIARFormer
from Predictor import Predictor
from cellpose import models as cellpose_models, io as cellpose_io, plot as cellpose_plot
from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation

# === Setup for GPU or CPU ===
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Load SegFormer
processor_segformer = SegformerImageProcessor(do_reduce_labels=False)
model_segformer = SegformerForSemanticSegmentation.from_pretrained(
    "nvidia/segformer-b0-finetuned-ade-512-512",
    num_labels=8,
    ignore_mismatched_sizes=True
)
model_segformer.load_state_dict(torch.load("trained_model_200.pt", map_location=device))
model_segformer.to(device)
model_segformer.eval()

# Load StarDist model (CPU-only, no GPU support)
model_stardist = StarDist2D.from_pretrained('2D_versatile_fluo')

# Load Cellpose model with GPU if available
model_cellpose = cellpose_models.CellposeModel(gpu=torch.cuda.is_available())

# SegFormer Inference
def infer_segformer(image):
    image = image.convert("RGB")
    inputs = processor_segformer(images=image, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        logits = model_segformer(**inputs).logits
    pred_mask = torch.argmax(logits, dim=1)[0].cpu().numpy()

    # Colorize
    colors = np.array([[0,0,0], [255,0,0], [0,255,0], [0,0,255], [255,255,0], [255,0,255], [0,255,255], [128,128,128]])
    color_mask = np.zeros((pred_mask.shape[0], pred_mask.shape[1], 3), dtype=np.uint8)
    for c in range(8):
        color_mask[pred_mask == c] = colors[c]
    return image, Image.fromarray(color_mask)

# StarDist Inference
def infer_stardist(image):
    image_gray = rgb2gray(np.array(image)) if image.mode == 'RGB' else np.array(image)
    labels, _ = model_stardist.predict_instances(normalize(image_gray))
    overlay = render_label(labels, img=image_gray)
    overlay = (overlay[..., :3] * 255).astype(np.uint8)
    return image, Image.fromarray(overlay)

# MEDIAR Inference
def infer_mediar(image, temp_dir="temp_mediar"):
    os.makedirs(temp_dir, exist_ok=True)
    input_path = os.path.join(temp_dir, "input_image.tiff")
    output_path = os.path.join(temp_dir, "input_image_label.tiff")
    image.save(input_path)

    model_args = {
        "classes": 3,
        "decoder_channels": [1024, 512, 256, 128, 64],
        "decoder_pab_channels": 256,
        "encoder_name": 'mit_b5',
        "in_channels": 3
    }

    model = MEDIARFormer(**model_args)
    weights = torch.load("from_phase1.pth", map_location=device)
    model.load_state_dict(weights, strict=False)
    model.to(device)
    model.eval()

    predictor = Predictor(model, device.type, temp_dir, temp_dir, algo_params={"use_tta": False})
    predictor.img_names = ["input_image.tiff"]
    _ = predictor.conduct_prediction()

    pred = imread(output_path)
    fig, ax = plt.subplots(figsize=(6, 6))
    ax.imshow(pred, cmap="cividis")
    ax.axis("off")

    buf = io.BytesIO()
    plt.savefig(buf, format="png")
    plt.close()
    buf.seek(0)

    return image, Image.open(buf)

# Cellpose Inference
def infer_cellpose(image, temp_dir="temp_cellpose"):
    os.makedirs(temp_dir, exist_ok=True)
    input_path = os.path.join(temp_dir, "input_image.tif")
    output_overlay = os.path.join(temp_dir, "overlay.png")

    image.save(input_path)
    img = cellpose_io.imread(input_path)
    masks, flows, styles = model_cellpose.eval(img, batch_size=1)

    fig = plt.figure(figsize=(12,5))
    cellpose_plot.show_segmentation(fig, img, masks, flows[0])
    plt.tight_layout()
    fig.savefig(output_overlay)
    plt.close(fig)

    return image, Image.open(output_overlay)

# Main segmentation dispatcher
def segment(model_name, image):
    ext = image.format.lower() if hasattr(image, 'format') and image.format else None
    if model_name == "Cellpose" and ext not in ["tif", "tiff", None]:
        return None, f"❌ Cellpose only supports `.tif` or `.tiff` images."

    if model_name == "SegFormer":
        return infer_segformer(image)
    elif model_name == "StarDist":
        return infer_stardist(image)
    elif model_name == "MEDIAR":
        return infer_mediar(image)
    elif model_name == "Cellpose":
        return infer_cellpose(image)
    else:
        return None, f"❌ Unknown model: {model_name}"

# === Gradio UI ===
with gr.Blocks(title="Cell Segmentation Explorer") as app:
    gr.Markdown("## Cell Segmentation Explorer")
    gr.Markdown("Choose a segmentation model, upload an appropriate image, and view the predicted mask.")

    with gr.Row():
        with gr.Column():
            model_dropdown = gr.Dropdown(
                choices=["SegFormer", "StarDist", "MEDIAR", "Cellpose"],
                label="Select Segmentation Model",
                value="SegFormer"
            )
            image_input = gr.Image(type="pil", label="Uploaded Image")
            description_box = gr.Markdown("Accepted formats: `.png`, `.jpg`, `.tif`, `.tiff`.")
            submit_btn = gr.Button("Submit")
            clear_btn = gr.Button("Clear")
        with gr.Column():
            output_image = gr.Image(label="Segmentation Result")

    def handle_submit(model_name, img):
        if img is None:
            return None
        _, result = segment(model_name, img)
        return result

    submit_btn.click(
        fn=handle_submit,
        inputs=[model_dropdown, image_input],
        outputs=output_image
    )

    clear_btn.click(
        lambda: [None, None],
        inputs=None,
        outputs=[image_input, output_image]
    )

    gr.Markdown("---")
    gr.Markdown("### Sample Images (click to use as input)")

    original_sample_paths = ["img1.png", "img2.png", "img3.png"]
    resized_sample_paths = []

    for idx, p in enumerate(original_sample_paths):
        img = PILImage.open(p).resize((128, 128))
        temp_path = f"/tmp/sample_resized_{idx}.png"
        img.save(temp_path)
        resized_sample_paths.append(temp_path)

    sample_image_components = []
    with gr.Row():
        for i, img_path in enumerate(resized_sample_paths):
            def load_full_image(idx=i):
                return PILImage.open(original_sample_paths[idx])

            sample_img = gr.Image(value=img_path, type="pil", interactive=True, show_label=False)
            sample_img.select(
                fn=load_full_image,
                inputs=[],
                outputs=image_input
            )
            sample_image_components.append(sample_img)

app.launch()