blurr

A library designed for fastai developers who want to train and deploy Hugging Face transformers

Named after the fastest transformer (well, at least of the Autobots), BLURR provides both a comprehensive and extensible framework for training and deploying 🤗 huggingface transformer models with fastai >= 2.0.

Utilizing features like fastai's new @typedispatch and @patch decorators, along with a simple class hiearchy, BLURR provides fastai developers with the ability to train and deploy transformers on a variety of tasks. It includes a high, mid, and low-level API that will allow developers to use much of it out-of-the-box or customize it as needed.

Supported Text/NLP Tasks:

• Sequence Classification (multiclassification and multi-label classification)
• Token Classification
• Question Answering
• Summarization
• Tranlsation
• Language Modeling (Causal and Masked)

Supported Vision Tasks:

• In progress

Supported Audio Tasks:

• In progress

Install

You can now pip install blurr via pip install ohmeow-blurr

Or, even better as this library is under very active development, create an editable install like this:

git clone https://github.com/ohmeow/blurr.git
cd blurr
pip install -e ".[dev]"

How to use

Please check the documentation for more thorough examples of how to use this package.

The following two packages need to be installed for blurr to work:

1. fastai (see http://docs.fast.ai/ for installation instructions)
2. huggingface transformers (see https://huggingface.co/transformers/installation.html for details)

Imports

import torch
from transformers import *
from fastai.text.all import *

from blurr.text.data.all import *
from blurr.text.modeling.all import *

Get your data

path = untar_data(URLs.IMDB_SAMPLE)

model_path = Path("models")
imdb_df = pd.read_csv(path / "texts.csv")

Get n_labels from data for config later

n_labels = len(imdb_df["label"].unique())

Get your 🤗 objects

model_cls = AutoModelForSequenceClassification

pretrained_model_name = "bert-base-uncased"

config = AutoConfig.from_pretrained(pretrained_model_name)
config.num_labels = n_labels

hf_arch, hf_config, hf_tokenizer, hf_model = get_hf_objects(pretrained_model_name, model_cls=model_cls, config=config)

Build your Data 🧱 and your DataLoaders

# single input
blocks = (TextBlock(hf_arch, hf_config, hf_tokenizer, hf_model), CategoryBlock)
dblock = DataBlock(blocks=blocks, get_x=ColReader("text"), get_y=ColReader("label"), splitter=ColSplitter())

dls = dblock.dataloaders(imdb_df, bs=4)

dls.show_batch(dataloaders=dls, max_n=2)

... and 🚂

# slow
model = BaseModelWrapper(hf_model)

learn = Learner(
    dls,
    model,
    opt_func=partial(Adam, decouple_wd=True),
    loss_func=CrossEntropyLossFlat(),
    metrics=[accuracy],
    cbs=[BaseModelCallback],
    splitter=blurr_splitter,
)

learn.freeze()

learn.fit_one_cycle(3, lr_max=1e-3)

# slow
learn.show_results(learner=learn, max_n=2)

Using the high-level Blurr API

Using the high-level API we can reduce DataBlock, DataLoaders, and Learner creation into a single line of code.

Included in the high-level API is a general BLearner class (pronouned "Blurrner") that you can use with hand crafted DataLoaders, as well as, task specific BLearners like BLearnerForSequenceClassification that will handle everything given your raw data sourced from a pandas DataFrame, CSV file, or list of dictionaries (for example a huggingface datasets dataset)

# slow
learn = BlearnerForSequenceClassification.from_data(imdb_df, pretrained_model_name, dl_kwargs={"bs": 4})

# slow
learn.fit_one_cycle(1, lr_max=1e-3)

# slow
learn.show_results(learner=learn, max_n=2)

⭐ Props

A word of gratitude to the following individuals, repos, and articles upon which much of this work is inspired from:

• The wonderful community that is the fastai forum and especially the tireless work of both Jeremy and Sylvain in building this amazing framework and place to learn deep learning.
• All the great tokenizers, transformers, docs, examples, and people over at huggingface
• FastHugs
• Fastai with 🤗Transformers (BERT, RoBERTa, XLNet, XLM, DistilBERT)
• Fastai integration with BERT: Multi-label text classification identifying toxicity in texts
• fastinference