config.md

Installation

It's recommended that you install the requirements for these t5 models in a new environment since they are known to conflict with common Python package requirements in the scientific Python Stack.

Conda ENV Setup

Create

conda create -n nlp-t5

Activate

conda activate nlp-t5

Packages

conda install python==3.8
conda install -c conda-forge python-multipart pandas scikit-learn transformers protobuf

At least on an M1 Macbookpro, conda managed to get these dependencies right. However, you may need to brew install a few system-level packages and/or deal with gcc or xcode modules if you don't have some of the lower-level prerequisites installed.

References

Original Notebook https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb

Dataset https://github.com/dair-ai/emotion_dataset

Whitepaper https://arxiv.org/pdf/1910.10683.pdf

t5-demo.py

import pandas as pd
from transformers import AutoTokenizer, AutoModelWithLMHead

tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-emotion")
model = AutoModelWithLMHead.from_pretrained("mrm8488/t5-base-finetuned-emotion")

def get_emotion(text):
  input_ids = tokenizer.encode(text + '</s>', return_tensors='pt')

  output = model.generate(input_ids=input_ids,
               max_length=2)
  
  dec = [tokenizer.decode(ids) for ids in output]
  label = dec[0]
  return label
  
data = [(1, "I never drink the tap water."), (2, "This customer service is like an enraged wombat."), (3, "Sometimes it's lonely taking a survey.")]
df = pd.DataFrame(data, columns = ["id", "text"])
df['T5_prediction'] = df['text'].map(lambda text: get_emotion(text))

# Print results to console or Jupyter cell (no need to use print function)
print(df.head())