app.py

import torch
import gradio as gr
import numpy as np
from PIL import Image
import torchvision.transforms as T
from model import DepthSTAR
import matplotlib.cm as cm

# Load model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DepthSTAR().to(device)
model.load_state_dict(torch.load("depth_model_all.pth", map_location=device))
model.eval()

# Preprocessing
transform = T.Compose([
    T.Resize((32, 32)),
    T.ToTensor(),
])

# Larger output display
image_output_size = 256

def predict_depth(image):
    img = transform(image).unsqueeze(0).to(device)
    with torch.no_grad():
        pred = model(img)[0, 0].cpu().numpy()

    # Normalize prediction
    pred_normalized = (pred - pred.min()) / (pred.max() - pred.min() + 1e-8)

    # Apply inferno colormap
    cmap = cm.get_cmap('inferno')
    pred_colored = cmap(pred_normalized)[:, :, :3]  # Drop alpha channel
    pred_colored = (pred_colored * 255).astype(np.uint8)
    pred_pil = Image.fromarray(pred_colored)

    # Upscale for display
    upscale_size = (256, 256)
    image_resized = image.resize(upscale_size, resample=Image.NEAREST)
    pred_resized = pred_pil.resize(upscale_size, resample=Image.NEAREST)

    return [image_resized, pred_resized]


# Gradio UI
examples = [["img_000.png"],["img_001.png"],["img_002.png"],["img_003.png"],["img_004.png"],["img_005.png"],]
# examples = [["example.png"]]

demo = gr.Interface(
    fn=predict_depth,
    inputs=gr.Image(type="pil", label="Input RGB Image", height=image_output_size),
    outputs=[
        gr.Image(type="pil", label="Original Image", height=image_output_size),
        gr.Image(type="pil", label="Predicted Depth Map", height=image_output_size),
    ],
    title="🔭 DepthStar: Light-weight Depth Estimation",
    description="Upload an RGB image and get the depth map predicted by our tiny DepthStar model.",
    examples=examples,
    theme="darkdefault",
)

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