File size: 2,694 Bytes
b84549f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 |
from transformers import BertTokenizer, BertModel, BertConfig
from utils.dl.common.model import set_module
from torch import nn
import torch
from utils.common.log import logger
bert_model_tag = 'bert-base-multilingual-cased'
class BertForSenCls(nn.Module):
def __init__(self, num_classes):
super(BertForSenCls, self).__init__()
logger.info(f'init bert for sen cls (using {bert_model_tag})')
self.bert = BertModel.from_pretrained(bert_model_tag)
self.classifier = nn.Linear(768, num_classes)
def forward(self, **x):
x['return_dict'] = False
pool_output = self.bert(**x)[-1]
return self.classifier(pool_output)
class BertForTokenCls(nn.Module):
def __init__(self, num_classes):
super(BertForTokenCls, self).__init__()
logger.info(f'init bert for token cls (using {bert_model_tag})')
self.bert = BertModel.from_pretrained(bert_model_tag)
self.classifier = nn.Linear(768, num_classes)
def forward(self, **x):
x['return_dict'] = False
pool_output = self.bert(**x)[0]
return self.classifier(pool_output)
class BertForTranslation(nn.Module):
def __init__(self):
super(BertForTranslation, self).__init__()
self.bert = BertModel.from_pretrained(bert_model_tag)
vocab_size = BertConfig.from_pretrained(bert_model_tag).vocab_size
self.decoder = nn.Linear(768, vocab_size)
logger.info(f'init bert for sen cls (using {bert_model_tag}), vocab size {vocab_size}')
# https://github.com/huggingface/transformers/blob/66954ea25e342fd451c26ec1c295da0b8692086b/src/transformers/models/bert_generation/modeling_bert_generation.py#L594
self.decoder.weight.data.normal_(mean=0.0, std=0.02)
def forward(self, **x):
x['return_dict'] = False
seq_output = self.bert(**x)[0]
return self.decoder(seq_output)
def bert_base_sen_cls(num_classes):
return BertForSenCls(num_classes)
def bert_base_token_cls(num_classes):
return BertForTokenCls(num_classes)
def bert_base_translation(no_bert_pooler=False):
# return BertForTranslation()
from transformers import BertTokenizer, BertModel, BertConfig, EncoderDecoderModel, BertGenerationDecoder
encoder = BertModel.from_pretrained(bert_model_tag)
model = BertGenerationDecoder.from_pretrained(bert_model_tag)
model.bert = encoder
if no_bert_pooler:
logger.info('replace pooler with nn.Identity()')
encoder.pooler = nn.Identity()
return model
|