hw2-tops-fit-automl

Tops Fit Classifier (FLAML AutoML)

Purpose

Predicts how a top fits one specific wearer (tight, regular, baggy) from its brand, fabric, garment type and flat-lay measurements. Built for CMU 24-679 Homework 2 to practice AutoML — not a size-recommendation tool for other people.

Data

• Source: ypolatog/tops-fit-dataset (CC-BY-4.0), created by classmate ypolatog.
• 30 real tops + 347 synthetic rows (jitter, scaling, mixup, categorical perturbation).
• Split: 8 real tops held out as test (stratified by fit, seed 42). Training used the other 22 real tops plus 235 synthetic rows whose parent tops were all in training (112 synthetic rows dropped to prevent leakage).
• Features: brand, main_fabric, garment_type (categorical); chest_width_cm, body_length_cm, sleeve_length_cm (numeric, cm).
• Target: fit.

Preprocessing

Categorical columns cast to pandas category with a shared vocabulary; no scaling (tree models).

Training setup and search space

• FLAML AutoML, time budget 120s, seed 42, metric macro-F1.
• Validation: 5-fold group cross-validation (group = parent top) so synthetic copies never cross folds.
• Families searched: lgbm, rf, xgboost, extra_tree, xgb_limitdepth, sgd, lrl1.
• Best model: xgboost
• Best hyperparameters: {'n_estimators': 72, 'max_leaves': 8, 'min_child_weight': np.float64(19.50627616863098), 'learning_rate': np.float64(0.020232966686106597), 'subsample': np.float64(0.7004758809403493), 'colsample_bylevel': 1.0, 'colsample_bytree': 1.0, 'reg_alpha': 0.0009765625, 'reg_lambda': np.float64(0.018706743166653998)}

Results

Uncertainty: with only 8 test tops, one prediction = 12.5 accuracy points, so these numbers are very noisy.

Limitations and ethics

• Labels describe one wearer's body and taste; the model does not generalize to other people.
• Only 30 real tops (19 t-shirts); most brands appear once or twice, so brand carries little reliable signal.
• Synthetic rows are near-copies of real tops and add little new information.
• Fit labels are subjective, and the boundary between regular and the other classes is fuzzy.

How to use

import pickle, pandas as pd
model = pickle.load(open("model.pkl", "rb"))
model.predict(pd.DataFrame([...]))   # same 6 feature columns, categorical columns as pandas category

Compute

Google Colab CPU, ~120s search.

AI usage disclosure

Code and this card were drafted with Claude (Anthropic) and reviewed, run, and interpreted by the author.

Acknowledgments

Dataset by ypolatog (CMU 24-679, Fall 2026).

File safety note

model.pkl stores the full FLAML AutoML object using Python's pickle format. Hugging Face's scanner flags it because pickle files can execute code when loaded, and this file references libraries outside the scanner's safe list (FLAML, XGBoost, NumPy). It contains no malicious code, but as a general rule, only load pickle files from sources you trust. In future work I would also export the model in a safe format, such as XGBoost's native JSON (save_model("model.json")), alongside the pickle.