repeated token is generated comparing to beam-search, when using fast_contrastive_search on T5 about simctg HOT 4 CLOSED

I used the fast_contrastive_search， cpoied from https://github.com/yxuansu/SimCTG/blob/main/SimCTGEncDec/SimCTGT5/simctgt5.py code ,as follows： but generated reapied tokens:

Hi @wuzhiye7,

Can you send the name of Chinese BART huggingface model and your inputs to me? I would like to test the instance and provide you some feedbacks.

from simctg.

model : t5-pegasus-base ， huggingface model hub: imxly/t5-pegasus

input_tokens :['[CLS]', '我', '不会', '贴', '假', '睫', '毛', '呀', '，', '好', '难', '！', '[SEP]']
input_ids :[[101, 1909, 6932, 4745, 463, 3466, 2644, 840, 5661, 1266, 5314, 5658, 102], [101, 32018, 1909, 7117, 7914, 4913, 3399, 179, 505, 1963, 3443, 26300, 2808, 6312, 135, 40959, 731, 31348, 15699, 5661, 24630, 1963, 463, 3466, 2644, 637, 199, 2374, 2106, 4866, 5661, 541, 1963, 26300, 4745, 198, 28756, 4745, 615, 26257, 198, 179, 102]]

@yxuansu

from simctg.

model : t5-pegasus-base ， huggingface model hub: imxly/t5-pegasus input_tokens :['[CLS]', '我', '不会', '贴', '假', '睫', '毛', '呀', '，', '好', '难', '！', '[SEP]'] input_ids :[[101, 1909, 6932, 4745, 463, 3466, 2644, 840, 5661, 1266, 5314, 5658, 102], [101, 32018, 1909, 7117, 7914, 4913, 3399, 179, 505, 1963, 3443, 26300, 2808, 6312, 135, 40959, 731, 31348, 15699, 5661, 24630, 1963, 463, 3466, 2644, 637, 199, 2374, 2106, 4866, 5661, 541, 1963, 26300, 4745, 198, 28756, 4745, 615, 26257, 198, 179, 102]]

@yxuansu

Hi @wuzhiye7,

I have tested the case on my end. Please follow the instructions below:

(1) First, install simctg from pip:

pip install simctg --upgrade

(2) Second, run the example below:

from simctg.simctgt5 import SimCTGT5
model_name = r'imxly/t5-pegasus'
# initialize tokenizer
from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained(model_name)
# initialize model
from transformers.models.mt5.modeling_mt5 import MT5ForConditionalGeneration
t5model = MT5ForConditionalGeneration.from_pretrained(model_name)
model = SimCTGT5(model_name, user_defined_model=t5model, user_defined_tokenizer=tokenizer, special_token_list=[])

print ('------------------------------------------')
# prepare input
text = '我不会贴假睫毛呀，好难！'
ids = tokenizer.encode(text, return_tensors='pt')
print ('The input text is: {}'.format(text))
print ('------------------------------------------')
# generate result
output = model.fast_contrastive_search(input_ids=ids, beam_width=5, alpha=0.5, decoding_len=30,
        start_of_sequence_token_id=tokenizer.cls_token_id, 
        end_of_sequence_token_id=tokenizer.sep_token_id, early_stop = True)
output_text = ''.join(tokenizer.convert_ids_to_tokens(output))
print ('The output text is: {}'.format(output_text))
'''
  ------------------------------------------
  The input text is: 我不会贴假睫毛呀，好难！
  ------------------------------------------
  The output text is: 如何贴假睫毛？我是女生
'''

P.S. If you are interested, the source code of simctg package is located here (https://github.com/yxuansu/SimCTG/tree/main/simctg).

Please let me know if you have any questions.

from simctg.

model : t5-pegasus-base ， huggingface model hub: imxly/t5-pegasus input_tokens :['[CLS]', '我', '不会', '贴', '假', '睫', '毛', '呀', '，', '好', '难', '！', '[SEP]'] input_ids :[[101, 1909, 6932, 4745, 463, 3466, 2644, 840, 5661, 1266, 5314, 5658, 102], [101, 32018, 1909, 7117, 7914, 4913, 3399, 179, 505, 1963, 3443, 26300, 2808, 6312, 135, 40959, 731, 31348, 15699, 5661, 24630, 1963, 463, 3466, 2644, 637, 199, 2374, 2106, 4866, 5661, 541, 1963, 26300, 4745, 198, 28756, 4745, 615, 26257, 198, 179, 102]]
@yxuansu

Hi @wuzhiye7,

I have tested the case on my end. Please follow the instructions below:

(1) First, install simctg from pip:

pip install simctg --upgrade

(2) Second, run the example below:

from simctg.simctgt5 import SimCTGT5
model_name = r'imxly/t5-pegasus'
# initialize tokenizer
from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained(model_name)
# initialize model
from transformers.models.mt5.modeling_mt5 import MT5ForConditionalGeneration
t5model = MT5ForConditionalGeneration.from_pretrained(model_name)
model = SimCTGT5(model_name, user_defined_model=t5model, user_defined_tokenizer=tokenizer, special_token_list=[])

print ('------------------------------------------')
# prepare input
text = '我不会贴假睫毛呀，好难！'
ids = tokenizer.encode(text, return_tensors='pt')
print ('The input text is: {}'.format(text))
print ('------------------------------------------')
# generate result
output = model.fast_contrastive_search(input_ids=ids, beam_width=5, alpha=0.5, decoding_len=30,
        start_of_sequence_token_id=tokenizer.cls_token_id, 
        end_of_sequence_token_id=tokenizer.sep_token_id, early_stop = True)
output_text = ''.join(tokenizer.convert_ids_to_tokens(output))
print ('The output text is: {}'.format(output_text))
'''
  ------------------------------------------
  The input text is: 我不会贴假睫毛呀，好难！
  ------------------------------------------
  The output text is: 如何贴假睫毛？我是女生
'''

P.S. If you are interested, the source code of simctg package is located here (https://github.com/yxuansu/SimCTG/tree/main/simctg).

Please let me know if you have any questions.

thanks ，its ok now

from simctg.