app.py

import torch
from torch import nn
import torchvision.transforms as transforms
import torch.nn.functional as F
from pathlib import Path
import gradio as gr
from huggingface_hub import hf_hub_download

LABELS = Path("classes.txt").read_text().splitlines()
num_classes = len(LABELS)

model = nn.Sequential(
    nn.Conv2d(1, 64, 3, padding="same"),
    nn.ReLU(),
    nn.MaxPool2d(2),
    nn.Conv2d(64, 128, 3, padding="same"),
    nn.ReLU(),
    nn.MaxPool2d(2),
    nn.Conv2d(128, 256, 3, padding="same"),
    nn.ReLU(),
    nn.MaxPool2d(2),
    nn.Flatten(),
    nn.Linear(2304, 512),
    nn.ReLU(),
    nn.Linear(512, num_classes),
)

model_path = hf_hub_download(repo_id="jerilseb/quickdraw-small", filename="model.pth")
state_dict = torch.load(model_path, map_location="cpu")
model.load_state_dict(state_dict)
model.eval()


transform = transforms.Compose(
    [
        transforms.Resize((28, 28)),
        transforms.ToTensor(),
        transforms.Normalize((0.5,), (0.5,)),
    ]
)

def predict(image):
    image = image['composite']
    tensor = transform(image).unsqueeze(0)
    with torch.no_grad():
        out = model(tensor)

    probabilities = F.softmax(out[0], dim=0)
    values, indices = torch.topk(probabilities, 5)
    return {LABELS[i]: v.item() for i, v in zip(indices, values)}


inputs = gr.ImageEditor(
    type="pil",
    height=720,
    width=720,
    layers=False,
    image_mode="L",
    brush=gr.Brush(default_color="white", default_size=20),
    sources=[],
    label="Draw a shape",
)
demo = gr.Interface(predict, inputs=inputs, outputs="label", live=True)
demo.launch(debug=True)