A Multi-task learning model with two prediction heads

First prediction head classifies between keyword sentences vs statements/questions
Second prediction head corresponds to classifier for statements vs questions

Article

Demo Notebook

Colab Notebook Multi-task Query classifiers

Clone the model repo

git clone https://huggingface.co/shahrukhx01/bert-multitask-query-classifiers

%cd bert-multitask-query-classifiers/

Load models

from multitask_model import BertForSequenceClassification
from transformers import AutoTokenizer
import torch
model = BertForSequenceClassification.from_pretrained(
        "shahrukhx01/bert-multitask-query-classifiers",
        task_labels_map={"quora_keyword_pairs": 2, "spaadia_squad_pairs": 2},
    )
tokenizer = AutoTokenizer.from_pretrained("shahrukhx01/bert-multitask-query-classifiers")

Run inference on both Tasks

from multitask_model import BertForSequenceClassification
from transformers import AutoTokenizer
import torch
model = BertForSequenceClassification.from_pretrained(
        "shahrukhx01/bert-multitask-query-classifiers",
        task_labels_map={"quora_keyword_pairs": 2, "spaadia_squad_pairs": 2},
    )
tokenizer = AutoTokenizer.from_pretrained("shahrukhx01/bert-multitask-query-classifiers")

## Keyword vs Statement/Question Classifier
input = ["keyword query", "is this a keyword query?"]
task_name="quora_keyword_pairs"
sequence = tokenizer(input, padding=True, return_tensors="pt")['input_ids']
logits = model(sequence, task_name=task_name)[0]
predictions = torch.argmax(torch.softmax(logits, dim=1).detach().cpu(), axis=1)
for input, prediction in zip(input, predictions):
  print(f"task: {task_name}, input: {input} \n prediction=> {prediction}")
  print()