Evaluating Tokenizers

TokenizerEvaluator computes reconstruction and codebook-quality metrics over a dataloader.

from medtokenizers import load_tokenizer, TokenizerEvaluator

model = load_tokenizer("./path/to/checkpoint")
evaluator = TokenizerEvaluator(model, device="cuda")
results = evaluator.evaluate(test_loader)
evaluator.print_results(results)

Available metrics include PSNR, SSIM, LPIPS, MSE / MAE, perplexity, and codebook usage – see Evaluation.