poster-font-classifier

Fine-tuned distilbert-base-uncased that maps a natural-language poster brief to a font ('Font' column), for poster font recommendation.

• Task: single-label (softmax) text classification
• Labels (15): Avenir, Baskerville, Bebas Neue, Comic Sans MS, Courier New, Futura, Garamond, Helvetica, Impact, Lobster, Montserrat, Papyrus, Playfair Display, Roboto, Times New Roman
• Training data: 788 train / 198 validation examples
• Hyperparameters: 5 epochs, lr 5e-05, batch size 16, max length 128

Validation metrics

metric	value
accuracy	0.8939
f1_macro	0.8937

Usage

import json, torch
from huggingface_hub import hf_hub_download
from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "CXu0630/poster-font-classifier"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo).eval()
cfg = json.load(open(hf_hub_download(repo, "inference_config.json")))

enc = tokenizer(["retro jazz night poster"], truncation=True, max_length=cfg["max_len"], return_tensors="pt")
with torch.no_grad():
    logits = model(**enc).logits
probs = torch.softmax(logits, -1)[0]
print(sorted(((model.config.id2label[i], float(p)) for i, p in enumerate(probs)), key=lambda x: -x[1]))