Bert 预训练模型每次给出随机输出

Ash*_*vin 3 python-3.x pytorch bert-language-model huggingface-transformers

我试图在 Huggingface bert 变压器之后添加一个额外的层,所以我BertForSequenceClassification在我的nn.Module网络中使用。但是,与直接加载模型相比,我看到模型给了我随机输出。

模型 1:

from transformers import BertForSequenceClassification

model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels = 5) # as we have 5 classes

import torch
from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

input_ids = torch.tensor(tokenizer.encode(texts[0], add_special_tokens=True, max_length = 512)).unsqueeze(0)  # Batch size 1

print(model(input_ids))

Run Code Online (Sandbox Code Playgroud)

出去:

(tensor([[ 0.3610, -0.0193, -0.1881, -0.1375, -0.3208]],
        grad_fn=<AddmmBackward>),)
Run Code Online (Sandbox Code Playgroud)

模型 2:

import torch
from torch import nn

class BertClassifier(nn.Module):
    def __init__(self):
        super(BertClassifier, self).__init__()
        self.bert = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels = 5)
        # as we have 5 classes

        # we want our output as probability so, in the evaluation mode, we'll pass the logits to a softmax layer
        self.softmax = torch.nn.Softmax(dim = 1) # last dimension
    def forward(self, x):
        print(x.shape)
        x = self.bert(x)

        if self.training == False: # in evaluation mode
            pass
            #x = self.softmax(x)

        return x

# create our model

bertclassifier = BertClassifier()

print(bertclassifier(input_ids))
Run Code Online (Sandbox Code Playgroud)
torch.Size([1, 512])
torch.Size([1, 5])
(tensor([[-0.3729, -0.2192,  0.1183,  0.0778, -0.2820]],
        grad_fn=<AddmmBackward>),)
Run Code Online (Sandbox Code Playgroud)

应该是同款吧。我在这里发现了类似的问题,但没有合理的解释https://github.com/huggingface/transformers/issues/2770

  1. Bert 是否有一些随机化参数,如果有的话,如何获得可重现的输出?

  2. 为什么这两个模型给我不同的输出?有什么我做错了吗?

Zab*_*azi 5

原因是由于Bert的分类器层的随机初始化。如果你打印你的模型,你会看到

    (pooler): BertPooler(
      (dense): Linear(in_features=768, out_features=768, bias=True)
      (activation): Tanh()
    )
  )
  (dropout): Dropout(p=0.1, inplace=False)
  (classifier): Linear(in_features=768, out_features=5, bias=True)
)
Run Code Online (Sandbox Code Playgroud)

classifier在最后一层有一个,这层是在 之后添加的bert-base。现在,期望您将为下游任务训练该层。

如果您想获得更多见解:

model, li = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels = 5, output_loading_info=True) # as we have 5 classes
print(li)
Run Code Online (Sandbox Code Playgroud)
{'missing_keys': ['classifier.weight', 'classifier.bias'], 'unexpected_keys': ['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias'], 'error_msgs': []}
Run Code Online (Sandbox Code Playgroud)

您可以看到classifier.weightbias丢失了,因此每次调用时这些部分都会随机初始化BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels = 5)