transformer_onnx

transformers_onnx is a simple package which can use inside transformers pipeline.

Install

pip install transformers_onnx

Convert model into Onnx format

#for question-answering
python -m transformers.onnx --feature "question-answering" -m nlpconnect/roberta-base-squad2-nq ./qa/

#for text-classification or zeroshot classification
python -m transformers.onnx --feature "sequence-classification" -m cross-encoder/nli-roberta-base ./classifier/

#for feature-extraction (last_hidden_state or pooler_output)
python -m transformers.onnx --feature "default" -m nlpconnect/dpr-ctx_encoder_bert_uncased_L-2_H-128_A-2 ./feature/

#for token-classification
python -m transformers.onnx --feature "token-classification" -m dslim/bert-base-NER ./ner/

Use transformers_onnx to run transformers pipeline

Question Answering

from transformers import pipeline, AutoTokenizer, AutoConfig
from transformer_onnx import OnnxModel

model = OnnxModel("qa/model.onnx", task="question-answering")
model.config = AutoConfig.from_pretrained("nlpconnect/roberta-base-squad2-nq")
tokenizer = AutoTokenizer.from_pretrained("nlpconnect/roberta-base-squad2-nq")
qa = pipeline("question-answering", model=model, tokenizer=tokenizer)

# Input data
context = ["Released on 6/03/2021",
        "Release delayed until the 11th of August",
        "Documentation can be found here: huggingface.com"]
# Define column queries
queries = ["What is Released date?", "till when delayed?", "What is the url?"]
qa(context=context, question=queries)

Text Classification/ Zero shot classification

from transformers import pipeline, AutoTokenizer, AutoConfig
from transformer_onnx import OnnxModel

model = OnnxModel("classifier/model.onnx", task="sequence-classification")
model.config = AutoConfig.from_pretrained("cross-encoder/nli-roberta-base")
tokenizer = AutoTokenizer.from_pretrained("cross-encoder/nli-roberta-base")
zero_shot = pipeline("zero-shot-classification", model=model, tokenizer=tokenizer)
zero_shot(sequences=["Hello Hiiii", "I am playing football"], candidate_labels=["Greeting", "Sports"])

Feature Extraction

from transformers import pipeline, AutoTokenizer, AutoConfig
from transformer_onnx import OnnxModel

# for last_hidden_state
model = OnnxModel("feature/model.onnx", task="last_hidden_state")
tokenizer = AutoTokenizer.from_pretrained("nlpconnect/dpr-ctx_encoder_bert_uncased_L-2_H-128_A-2")
feature_extractor = pipeline("feature-extraction", model=model, tokenizer=tokenizer)
feature_extractor(["Hello Hiiii", "I am playing football"])

# for pooler_output
model = OnnxModel("feature/model.onnx", task="pooler_output")
tokenizer = AutoTokenizer.from_pretrained("nlpconnect/dpr-ctx_encoder_bert_uncased_L-2_H-128_A-2")
feature_extractor = pipeline("feature-extraction", model=model, tokenizer=tokenizer)
feature_extractor(["Hello Hiiii", "I am playing football"])

NER

from transformers import pipeline, AutoTokenizer, AutoConfig
from transformer_onnx import OnnxModel

model = OnnxModel("ner/model.onnx", task="token-classification")
model.config = AutoConfig.from_pretrained("dslim/bert-base-NER")
tokenizer = AutoTokenizer.from_pretrained("dslim/bert-base-NER")
ner = pipeline("token-classification", model=model, tokenizer=tokenizer)
ner("My name is transformers and I live in github/huggingface")