resnet18-approval-regressor

ResNet18 Approval Regressor (1–10)

Author: Dan Jung and Scotty McGee

Model Description

• Architecture: ResNet18 backbone, final fully connected layer outputs a scalar rating.
• Framework: PyTorch (see requirements.txt for dependencies).
• Task: Image regression. Predicts clothing "approval" ratings on a continuous scale from 1 to 10.
• Input: RGB clothing images, resized to 224x224.
• Output: Scalar approval score.

Use

import torch
from PIL import Image
from torchvision import transforms
from model import load_model

cfg = {"image_size": 224, "mean": [0.485, 0.456, 0.406], "std": [0.229, 0.224, 0.225], "label_range": [1.0, 10.0], "task": "image_regression", "backbone": "resnet18", "target": "approval(1-10)"}
tf = transforms.Compose([
    transforms.Resize((cfg["image_size"], cfg["image_size"])),
    transforms.ToTensor(),
    transforms.Normalize(mean=cfg["mean"], std=cfg["std"]),
])

m = load_model("pytorch_model.bin")
img = Image.open("your_image.jpg").convert("RGB")
x = tf(img).unsqueeze(0)
with torch.no_grad():
    y = m(x).squeeze(1).clamp(1.0, 10.0).item()
print(y)

Training Dataset and Procedure

• Dataset: 24679-Project-1
• Preprocessing: Images resized, normalized with standard Imagenet parameters.
• Validation: MAE and RMSE
• Config: Batch size 32, 4 Epochs, and Adam optimizer

Performance

• Validation MAE: 0.438 after 4 epochs
• Validation RMSE: 0.599

Intended Uses

• Automated clothing quality/style rating based on Dan's and Scotty's preferences.

Ethical Considerations

• Ratings are subjectivel ensures fairness and representation in labeling/clothing types.