app.py

import gradio as gr
from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
from PIL import Image
import torch
import numpy as np
import os 

# --- Documentation Strings ---

USAGE_GUIDELINES = """
## 1. Quick Start Guide: Generating a Segmentation Mask
This tool uses a powerful AI model (SegFormer) to analyze an image and identify different object boundaries, creating a map known as a segmentation mask.

1.  **Upload**: Click the 'Upload Blood Smear Image' box and select your image (JPG or PNG).
2.  **Test**: Optionally, click one of the 'Example Images' below to load sample data instantly.
3.  **Run**: Click the **"Submit"** button.
4.  **Review**: The 'Predicted Grayscale Mask' will appear, showing the boundaries the model detected.
"""

INPUT_EXPLANATION = """
## 2. Expected Inputs

| Input Field | Purpose | Requirement |
| :--- | :--- | :--- |
| **Upload Image** | The image containing the sample you want to analyze. | Must be a single image file (JPG, PNG). The input image is automatically resized to 512x512 pixels before processing. |

"""

OUTPUT_EXPLANATION = """
## 3. Expected Outputs (The Segmentation Mask)

The output is the **Predicted Grayscale Mask**. This is a map where every pixel is assigned a color based on the object class the model identified.

*   **Grayscale Representation:** Unlike color segmentation maps, this output uses different *shades of gray* to represent the 150 possible object categories (classes) the underlying model was trained on.
*   **Purpose:** The mask visually separates the detected boundaries (e.g., cell outlines, background, debris) from one another.
*   **Clarity:** The mask is automatically enlarged by **300% (3x)** to make the details easier to see.

### Sample Data for Testing
You can quickly test the application using the following provided example files. Click on the thumbnails (Sample blood smear image) below to upload the image automatically:

"""

# --------------------
# Core Pipeline Functions
# --------------------

# Load pretrained model
# Model: SegFormer B0 fine-tuned on ADE20K (150 classes)
processor = SegformerImageProcessor(do_reduce_labels=False)
model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")
model.eval()

# Prediction function
def segment_image(input_image):
    if input_image is None:
        gr.Warning("Please upload an image or select an example.")
        return None
        
    inputs = processor(images=input_image, return_tensors="pt")

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits

    # 1. Get the predicted class index for each pixel
    pred_mask = torch.argmax(logits, dim=1)[0].cpu().numpy()
    
    # 2. Normalize the mask to create a visible grayscale image (0-255 range)
    num_classes = logits.shape[1] # Should be 150 for ADE20K
    # normalized_mask = (pred_mask * (255 // num_classes)).astype(np.uint8)
    #change the segmented mask color to white (255) and the background to black (0)
    normalized_mask = np.where(pred_mask > 0, 255, 0).astype(np.uint8)
    output_image = Image.fromarray(normalized_mask)

    # 3. Bigger mask (3x) for better visualization
    scale_factor = 3
    new_size = (output_image.width * scale_factor, output_image.height * scale_factor)
    bigger_output = output_image.resize(new_size, resample=Image.NEAREST)

    return bigger_output

# --------------------
# Gradio UI
# --------------------

with gr.Blocks(title="Semantic Segmentation Tool") as demo:
    gr.Markdown("<h1 style='text-align: center;'> Malaria Blood Smear Segmentation (SegFormer) </h1>")
    gr.Markdown("Tool for analyzing image boundaries using a general purpose semantic segmentation model.")

    # 1. Guidelines Accordion
    with gr.Accordion(" Tips & Guidelines", open=False):
        gr.Markdown(USAGE_GUIDELINES)
        gr.Markdown("---")
        gr.Markdown(INPUT_EXPLANATION)
        gr.Markdown("---")
        gr.Markdown(OUTPUT_EXPLANATION)
        

    # 2. Interface Definition (Embedded)
    gr.Markdown("## Step 1: Upload the image")
    
    with gr.Row():
        with gr.Column(scale=1):
            # Define Input component directly inside the column (No .render() needed)
            input_image = gr.Image(type="pil", label="Step 1: Upload Blood Smear Image") 
    
    gr.Markdown("## Step 2: Click Submit for Segmentation")
    with gr.Row():
        submit_button = gr.Button("Submit for Segmentation", variant="primary")
    gr.Markdown("## Output")
    with gr.Row(scale=2):
         # Define Output component directly inside the column (No .render() needed)
        output_image = gr.Image(type="pil", label="Step 3: Predicted Grayscale Mask")


    # 3. Examples Section
    gr.Markdown("---")
    gr.Markdown("## Example Images")
    gr.Examples(
        examples=["data/1.png", "data/2.png", "data/3.png", "data/211.png"],
        inputs=[input_image],
        outputs=[output_image],
        fn=segment_image,
        cache_examples=False,
        label="Click to load and run a sample image",
    )
    
    # Event Handler
    submit_button.click(
        fn=segment_image,
        inputs=input_image,
        outputs=output_image
    )


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