def infer(args):
paddle.set_device(args.device)
set_seed(args.seed)
model = UnifiedTransformerLMHeadModel.from_pretrained(
args.model_name_or_path)
tokenizer = UnifiedTransformerTokenizer.from_pretrained(
args.model_name_or_path)
test_ds = load_dataset('duconv', splits='test_1')
test_ds, test_data_loader = create_data_loader(test_ds, tokenizer, args,
'test')
model.eval()
total_time = 0.0
start_time = time.time()
pred_responses = []
for step, inputs in enumerate(test_data_loader, 1):
input_ids, token_type_ids, position_ids, attention_mask, seq_len = inputs
output = model.generate(input_ids=input_ids,
token_type_ids=token_type_ids,
position_ids=position_ids,
attention_mask=attention_mask,
seq_len=seq_len,
max_length=args.max_dec_len,
min_length=args.min_dec_len,
decode_strategy=args.decode_strategy,
temperature=args.temperature,
top_k=args.top_k,
top_p=args.top_p,
num_beams=args.num_beams,
length_penalty=args.length_penalty,
early_stopping=args.early_stopping,
num_return_sequences=args.num_return_sequences,
use_fp16_decoding=args.use_fp16_decoding,
use_faster=args.faster)
total_time += (time.time() - start_time)
if step % args.logging_steps == 0:
print('step %d - %.3fs/step' %
(step, total_time / args.logging_steps))
total_time = 0.0
ids, scores = output
results = select_response(ids, scores, tokenizer, args.max_dec_len,
args.num_return_sequences)
pred_responses.extend(results)
start_time = time.time()
with open(args.output_path, 'w', encoding='utf-8') as fout:
for response in pred_responses:
fout.write(response + '\n')
print('\nSave inference result into: %s' % args.output_path)
target_responses = [example['response'] for example in test_ds]
calc_bleu_and_distinct(pred_responses, target_responses)
def interaction(args, model, tokenizer):
history = []
start_info = "Enter [EXIT] to quit the interaction, [NEXT] to start a new conversation."
cprint(start_info, "yellow", attrs=["bold"])
while True:
user_utt = input(colored("[Human]: ", "red", attrs=["bold"])).strip()
if user_utt == "[EXIT]":
break
elif user_utt == "[NEXT]":
history = []
cprint(start_info, "yellow", attrs=["bold"])
else:
history.append(user_utt)
inputs = tokenizer.dialogue_encode(
history,
add_start_token_as_response=True,
return_tensors=True,
is_split_into_words=False)
inputs['input_ids'] = inputs['input_ids'].astype('int64')
ids, scores = model.generate(
input_ids=inputs['input_ids'],
token_type_ids=inputs['token_type_ids'],
position_ids=inputs['position_ids'],
attention_mask=inputs['attention_mask'],
max_length=args.max_dec_len,
min_length=args.min_dec_len,
decode_strategy=args.decode_strategy,
temperature=args.temperature,
top_k=args.top_k,
top_p=args.top_p,
num_beams=args.num_beams,
length_penalty=args.length_penalty,
early_stopping=args.early_stopping,
num_return_sequences=args.num_return_sequences,
use_faster=True)
bot_response = select_response(ids,
scores,
tokenizer,
args.max_dec_len,
args.num_return_sequences,
keep_space=False)[0]
print(colored("[Bot]:", "blue", attrs=["bold"]),
colored(bot_response, attrs=["bold"]))
history.append(bot_response)
return