Function eval_ppl

lib/eval.py:81–96

(args, model, tokenizer, device=torch.device("cuda:0"))

# Function to evaluate perplexity (ppl) on a specified model and tokenizer
def eval_ppl(args, model, tokenizer, device=torch.device("cuda:0")):
    # Set dataset
    dataset = "wikitext"

    # Print status
    print(f"evaluating on {dataset}")

    # Get the test loader
    _, testloader = get_loaders(
        dataset, seed=0, seqlen=model.seqlen, tokenizer=tokenizer
    )

    # Evaluate ppl in no grad context to avoid updating the model
    with torch.no_grad():
        ppl_test = eval_ppl_wikitext(model, testloader, 1, device)
    return ppl_test


# Function to evaluate perplexity (ppl) specifically on the wikitext dataset