That means that with the block-sparse runtime it ran 1.92x faster than an dense networks on the evaluation, at the price of some impact on the accuracy (see below).

This model was fine-tuned from the HuggingFace BERT base uncased checkpoint on SQuAD1.1, and distilled from the equivalent model csarron/bert-base-uncased-squad-v1. This model is case-insensitive: it does not make a difference between english and English.

Pruning details

A side-effect of the block pruning is that some of the attention heads are completely removed: 106 heads were removed on a total of 144 (73.6%).

Here is a detailed view on how the remaining heads are distributed in the network after pruning.

Density plot

Details

Dataset	Split	# samples
SQuAD1.1	train	90.6K
SQuAD1.1	eval	11.1k

Fine-tuning

Python: 3.8.5
Machine specs:

CPU:

Memory: 64 GiB
GPUs: 1 GeForce GTX 3090, with 24GiB memory
GPU driver: 455.23.05, CUDA: 11.1

Results

Pytorch model file size: 335M (original BERT: 438M)

Metric	# Value	# Original (Table 2)
EM	71.88	80.8
F1	81.36	88.5

Example Usage

python

from transformers import pipeline

qa_pipeline = pipeline(
    "question-answering",
    model="madlag/bert-base-uncased-squad1.1-block-sparse-0.07-v1",
    tokenizer="madlag/bert-base-uncased-squad1.1-block-sparse-0.07-v1"
)

predictions = qa_pipeline({
    'context': "Frédéric François Chopin, born Fryderyk Franciszek Chopin (1 March 1810 – 17 October 1849), was a Polish composer and virtuoso pianist of the Romantic era who wrote primarily for solo piano.",
    'question': "Who is Frederic Chopin?",
})

print(predictions)