app.py

import io

import gradio as gr
import requests
import torch
import torch.nn.functional as F
import torchvision.transforms as transforms
from PIL import Image

from constants import MAKES_MODELS, PRICE_BIN_LABELS, YEARS

print("downloading checkpoint...")
data = requests.get(
    "https://data.aqnichol.com/car-data/models/mobilenetv2_432000_calib_torchscript.pt",
    stream=True,
).content

print("creating model...")
model = torch.jit.load(io.BytesIO(data))
model.eval()
transform = transforms.Compose(
    [
        transforms.ToTensor(),
        transforms.Normalize(
            (0.48145466, 0.4578275, 0.40821073),
            (0.26862954, 0.26130258, 0.27577711),
        ),
    ]
)

print("done.")


def classify(img: Image.Image):
    in_tensor = transform(img)[None]
    outputs = model(in_tensor)

    price_bins = dict(
        zip(PRICE_BIN_LABELS, F.softmax(outputs["price_bin"], dim=-1)[0].tolist())
    )
    years = dict(
        zip(
            [str(year) for year in YEARS] + ["Unknown"],
            F.softmax(outputs["year"], dim=-1)[0].tolist(),
        )
    )
    make_models = dict(
        zip(
            ([f"{make} {model}" for make, model in MAKES_MODELS] + ["Unknown"]),
            F.softmax(outputs["make_model"], dim=-1)[0].tolist(),
        )
    )
    return (
        f"${int(round(outputs['price_median'].item()))}",
        price_bins,
        years,
        make_models,
        img,
    )


iface = gr.Interface(
    fn=classify,
    inputs=gr.Image(shape=(224, 224), type="pil"),
    outputs=[
        gr.Text(label="Price Prediction"),
        gr.Label(label="Price Bin", num_top_classes=5),
        gr.Label(label="Year", num_top_classes=5),
        gr.Label(label="Make/Model", num_top_classes=10),
        gr.Image(label="Cropped Input"),
    ],
)
iface.queue(concurrency_count=2).launch()