MCPcopy Create free account
hub / github.com/THUDM/AgentTuning / get_model_output

Function get_model_output

eval_heldout/science-world/eval_utils.py:843–864  ·  view source on GitHub ↗
(args, input_str, tokenizer, lm_model, device, logger)

Source from the content-addressed store, hash-verified

841
842
843def get_model_output(args, input_str, tokenizer, lm_model, device, logger):
844 input_ids = tokenizer(input_str, return_tensors="pt", max_length=args["max_input_len"] , truncation=True).input_ids
845
846 sample_outputs = lm_model.generate(
847 input_ids.to(device),
848 max_length=16,
849 num_return_sequences=args['beams'],
850 num_beams=args['beams'],
851 )
852
853 lm_pred = sample_outputs
854
855 # Take the first prediction that is not "look around"
856 logger.info("Top N Predictions:")
857 predStrs = []
858 for i, pred in enumerate(lm_pred):
859 text = tokenizer.decode(pred)
860 text = post_process_generation(text)
861 logger.info("\t" + str(i) + "\t" + str(text) )
862 predStrs.append(text)
863
864 return predStrs
865
866
867def post_process_generation(raw_pred):

Callers

nothing calls this directly

Calls 2

post_process_generationFunction · 0.85
decodeMethod · 0.80

Tested by

no test coverage detected