app.py

import spaces
import torch
from torchvision.transforms import v2
import gradio
from PIL import Image
from huggingface_hub import hf_hub_download
from models.linear_predictor import Predictor
import numpy
import os

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

LABELS = [
    "Adipose",
    "Background",
    "Debris",
    "Lymphocytes",
    "Mucus",
    "Smooth Muscle",
    "Normal Colon Mucosa",
    "Cancer-associated Stroma",
    "Colorectal Adenocarcinoma Epithelium",
]

model = Predictor(n_labels=len(LABELS))

model_file = hf_hub_download(
    repo_id="Hali5/Mae-Model-MedMNIST-Predictor",
    filename="checkpoints/model_linear_v2_epoch_100.pt"
)

model.load_state_dict(torch.load(model_file,map_location=device))
model.to(device)
model.eval()

tf = v2.Compose([
    v2.ToImage(),
    v2.Resize((64, 64), antialias=True),
    v2.ToDtype(torch.float32, scale=True),
])

dataset = numpy.load("test_samples.npz")
images = dataset["images"]
labels = dataset["labels"]

val_dataset = numpy.load("val_samples.npz")
val_images = val_dataset["images"]
val_labels = val_dataset["labels"]

example_rows_test = []
example_rows_val = []

number_of_examples = len(labels)

os.makedirs("ui_examples", exist_ok=True)

for i in range(number_of_examples):

    img_array = images[i]
    val_img_arr = val_images[i]

    label_index = int(labels[i].item() if hasattr(labels[i], 'item') else labels[i])
    val_label_index = int(val_labels[i].item() if hasattr(val_labels[i], 'item') else val_labels[i])

    if img_array.max() <= 1.0:
        img_array = (img_array * 255).astype(numpy.uint8)
        val_img_arr = (val_img_arr * 255).astype(numpy.uint8)
    else:
        img_array = img_array.astype(numpy.uint8)
        val_img_arr = val_img_arr.astype(numpy.uint8)
    
    truth_label_text = LABELS[label_index] if label_index < len(LABELS) else f"Class {label_index}"
    val_truth_label_text = LABELS[val_label_index] if val_label_index < len(LABELS) else f"Class {val_label_index}"
    
    file_path = f"ui_examples/sample_{i}.jpg"
    val_file_path = f"ui_examples/val_sample_{i}.jpg"

    Image.fromarray(img_array, "RGB").save(file_path)
    Image.fromarray(val_img_arr, "RGB").save(val_file_path)

    example_rows_test.append([file_path, truth_label_text])
    example_rows_val.append([val_file_path, val_truth_label_text])

@spaces.GPU
def predict(image,truth_labels=True):
    if image is None:
        return None

    # uplouded image
    img_tensor = tf(image).unsqueeze(0).to(device)

    with torch.no_grad():
        outputs = model(img_tensor)
        print(outputs.shape)
        probabilities = torch.nn.functional.softmax(outputs.squeeze(0), dim=0)
    
    return {LABELS[i]: float(probabilities[i]) for i in range(len(LABELS))}

custom_css = """
.tab img{
    object-fit: fill !important;
    width: 100% !important;
    height: 100% !important;
    image-rendering: pixelated !important; /* Forces crisp pixel lines */
}
"""

with gradio.Blocks(css=custom_css) as demo:
    gradio.Markdown("# PathMNIST Image Classification")

    with gradio.Tab("Predict", elem_classes="tab"):
        gradio.Markdown("## Upload a tissue image for classification")
        with gradio.Row():
            input_img = gradio.Image(height=512, width=512)
            with gradio.Column():
                output_lbl = gradio.Label(num_top_classes=9)
                btn = gradio.Button("Predict")
                btn.click(fn=predict, inputs=input_img, outputs=output_lbl)

    with gradio.Tab("Examples (Validation)", elem_classes="tab"):
        gradio.Markdown("## Select an example below to test the model against the PathMNIST validation dataset")
        
        with gradio.Row():
            with gradio.Column():
                input_img_val_ex = gradio.Image(label="Selected Test Image", height=512, width=512)
                truth_box_val = gradio.Textbox(label="Ground Truth Label", interactive=False)
            output_lbl_val_ex = gradio.Label(num_top_classes=9, label="Model Prediction")
        
        gradio.Examples(
            examples=example_rows_val, 
            inputs=[input_img_val_ex, truth_box_val],
            outputs=output_lbl_val_ex,
            fn=predict,
            cache_examples=True,
        )

    with gradio.Tab("Examples (Test)", elem_classes="tab"):
        gradio.Markdown("## Select an example below to test the model against the PathMNIST test dataset")
        
        with gradio.Row():
            with gradio.Column():
                input_img_test_ex = gradio.Image(label="Selected Test Image", height=512, width=512)
                truth_box_test = gradio.Textbox(label="Ground Truth Label", interactive=False)
            output_lbl_test_ex = gradio.Label(num_top_classes=9, label="Model Prediction")
        
        gradio.Examples(
            examples=example_rows_test, 
            inputs=[input_img_test_ex, truth_box_test],
            outputs=output_lbl_test_ex,
            fn=predict,
            cache_examples=True,
        )

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