Generation of the most likely answers

[seonwoo-min]

Hi Lorenz,

Thanks for the great work and codes!
While I look through the codes, I notice a possible discrepancy between the paper and the codes.
The paper states that you used beam search with num_beams=5 and do_sample=True to obtain the most likely answers. On the other hand, the generate.py script is using the beam search with do_sample=False instead. Am I missing something here?

semantic_uncertainty/code/generate.py

Lines 113 to 130 in 27adbf0

	if args.decoding_method == 'beam_search':
	most_likely_generation = model.generate(input_ids,
	num_beams=5,
	num_return_sequences=2,
	do_sample=False,
	max_length=input_ids.shape[1] +
	max_length_of_generated_sequence,
	eos_token_id=period_token_id,
	bad_words_ids=question_framing_ids)
	elif args.decoding_method == 'greedy':
	most_likely_generation = model.generate(input_ids,
	num_beams=1,
	do_sample=False,
	max_length=input_ids.shape[1] +
	max_length_of_generated_sequence,
	eos_token_id=period_token_id,
	bad_words_ids=question_framing_ids)