---

Vision, Simplified.

Small models can recognize more than their size suggests.

GVM explores efficient computer vision using lightweight architectures, fast inference, and practical deployment.

Designed to run almost anywhere.

---

Classification Performance

---

Quick Start

import torch
import torchvision.transforms as transforms
import timm
import requests
import json
from PIL import Image

config = json.loads(
    requests.get(
        "https://huggingface.co/WhirlwindAI/GVM/resolve/main/config.json"
    ).text
)

model = timm.create_model(
    "mobilenetv2_100",
    pretrained=False,
    num_classes=config["num_classes"]
)

state = torch.hub.load_state_dict_from_url(
    "https://huggingface.co/WhirlwindAI/GVM/resolve/main/model.pth",
    map_location="cpu"
)

model.load_state_dict(state)
model.eval()

transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485,0.456,0.406],
        std=[0.229,0.224,0.225]
    )
])

image = Image.open("image.jpg").convert("RGB")
tensor = transform(image).unsqueeze(0)

prediction = model(tensor).argmax(1).item()

print(config["class_names"][prediction])

---

Highlights

---

Repository Contents

model.pth
config.json
README.md

---

Current Limitations

• Trained for only 3 epochs
• Frozen backbone during training
• CIFAR-100 is considerably harder than CIFAR-10
• Intended as an efficient baseline rather than a state-of-the-art classifier

---

Roadmap

• Higher resolution training
• Full backbone fine-tuning
• Improved augmentation
• ONNX export
• TensorRT support
• Interactive demo

---