转换成onnx格式模型后，推理结果出现偏差

[XinyuDu]

基于pytorch版本的bertspanner训练了一个NER模型。在pytorch下推理结果很好。但是通过torch.onnx.export将其转换成onnx模型后，再用onnx模型推理结果跟之前pytorch模型推理结果不一致。是我漏掉了什么？还是模型中有onnx不支持的op？谢谢！

pytorch 推理代码如下：

from processors.ner_span import InputExample
from processors.ner_span import ner_processors as processors
from processors.ner_span import convert_examples_to_features
from processors.utils_ner import bert_extract_item
from models.transformers import WEIGHTS_NAME,BertConfig,AlbertConfig
from models.bert_for_ner import BertSpanForNer
from processors.utils_ner import CNerTokenizer
from models.albert_for_ner import AlbertSpanForNer
import torch

label_list = ['O', 'TIM', 'UNI', 'TIT', 'COM', 'NAM', 'MAI', 'PHO', 'PNM','X',"[CLS]", "[SEP]"]
num_labels = len(label_list)
id2label = {i: label for i, label in enumerate(label_list)}
MODEL_CLASSES = {
    ## bert ernie bert_wwm bert_wwwm_ext
    'bert': (BertConfig, BertSpanForNer, CNerTokenizer),
    'albert': (AlbertConfig,AlbertSpanForNer,CNerTokenizer)
}
config_class, model_class, tokenizer_class = MODEL_CLASSES['bert']
tokenizer = tokenizer_class.from_pretrained('outputresumebert', do_lower_case=True)

sentence = '王建国,电话：13885699528,email：wjg@wjg.com,2000年9月-至今北京大学 本科'
example = InputExample(guid=0, text_a=sentence, subject=[])
feature = convert_examples_to_features([example], label_list, 128, tokenizer)
input_ids = torch.tensor([feature[0].input_ids], dtype=torch.long)
input_mask = torch.tensor([feature[0].input_mask], dtype=torch.long)

model = model_class.from_pretrained('outputresumebert')
model.to(torch.device("cpu"))
model.eval()
with torch.no_grad():  # CORE-2
    inputs = {"input_ids": input_ids, "attention_mask": input_mask,"start_positions": None,"end_positions": None}
    outputs = model(**inputs)
start_logits, end_logits = outputs[:2]
R = bert_extract_item(start_logits, end_logits)
label_entities = [[id2label[x[0]],x[1],x[2]] for x in R]
result=[]
for label in label_entities:
    result.append([label[0],sentence[label[1]:label[2]+1],[label[1],label[2]]])
print(result)

推理结果：
[['NAM', '王建国', [0, 2]], ['PHO', '13885699528', [7, 17]], ['MAI', 'wjg@wjg.com', [25, 35]], ['TIM', '2000年9月', [37, 43]], ['UNI', '北京大学', [47, 50]]]

onnx推理代码如下：

import onnxruntime
import numpy as np

ort_session = onnxruntime.InferenceSession("./convert_pytorch_to_tf/resumebert.onnx")

def to_numpy(tensor):
    return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()

# compute ONNX Runtime output prediction
sentence = '王建国,电话：13885699528,email：wjg@wjg.com,2000年9月-至今北京大学 本科'
example = InputExample(guid=0, text_a=sentence, subject=[])
feature = convert_examples_to_features([example], label_list, 128, tokenizer)
input_ids = torch.tensor([feature[0].input_ids], dtype=torch.long)
input_mask = torch.tensor([feature[0].input_mask], dtype=torch.long)
ort_inputs = {'input_ids': to_numpy(input_ids), 'input_mask': to_numpy(input_mask)}
ort_outs = ort_session.run(None, ort_inputs)
start_logits = torch.from_numpy(ort_outs[0])
end_logits = torch.from_numpy(ort_outs[1])
R = bert_extract_item(start_logits, end_logits)
label_entities = [[id2label[x[0]],x[1],x[2]] for x in R]
result=[]
for label in label_entities:
    result.append([label[0],sentence[label[1]:label[2]+1],[label[1],label[2]]])
print(result)

推理结果：
[['NAM', '王建国', [0, 2]], ['PHO', '13885699528', [7, 17]], ['TIM', '2000年9月-至今', [37, 46]]]

模型转换代码如下：

from processors.ner_span import InputExample
from processors.ner_span import ner_processors as processors
from processors.ner_span import convert_examples_to_features
from processors.utils_ner import bert_extract_item
from models.transformers import WEIGHTS_NAME,BertConfig,AlbertConfig
from models.bert_for_ner import BertSpanForNer
from processors.utils_ner import CNerTokenizer
from models.albert_for_ner import AlbertSpanForNer
import torch
import numpy as np

label_list = ['O', 'TIM', 'UNI', 'TIT', 'COM', 'NAM', 'MAI', 'PHO', 'PNM','X',"[CLS]", "[SEP]"]
num_labels = len(label_list)
id2label = {i: label for i, label in enumerate(label_list)}
MODEL_CLASSES = {
    ## bert ernie bert_wwm bert_wwwm_ext
    'bert': (BertConfig, BertSpanForNer, CNerTokenizer),
    'albert': (AlbertConfig,AlbertSpanForNer,CNerTokenizer)
}
config_class, model_class, tokenizer_class = MODEL_CLASSES['bert']
tokenizer = tokenizer_class.from_pretrained('outputresumebert', do_lower_case=True)

sentence = '王建国,电话：13885699528,email：wjg@wjg.com,2000年9月-至今北京大学 本科'

example = InputExample(guid=0, text_a=sentence, subject=[])
feature = convert_examples_to_features([example], label_list, 128, tokenizer)
input_ids = torch.tensor([feature[0].input_ids], dtype=torch.long)
input_mask = torch.tensor([feature[0].input_mask], dtype=torch.long)

model = model_class.from_pretrained('outputresumebert')
model.eval()

torch.onnx.export(model,
                  (input_ids,input_mask),
                  "convert_pytorch_to_tf/resumebert.onnx",
                  opset_version=10,
                  do_constant_folding=True,
                  input_names=['input_ids','input_mask'],
                  output_names=['start', 'end'],
                  dynamic_axes={'input_ids': {0 : 'batch_size'},    # variable lenght axes
                                'input_mask': {0 : 'batch_size'}}
                 )   

[tianzuishiwo]

你的问题解决了吗？现在遇到同样的问题。