Bert Base Uncased Squadv1 X2.01 F89.2 D30 Hybrid Rewind Opt V1

🚀 BERT-base uncased model fine-tuned on SQuAD v1

This project presents a fine-tuned BERT-base uncased model on the SQuAD v1 dataset. It leverages pruning techniques to reduce the model size and speed up inference while maintaining high accuracy.

🚀 Quick Start

To start using this model, you first need to install the nn_pruning library. This library contains the optimization script that packs the linear layers into smaller ones by removing empty rows/columns.

pip install nn_pruning

Then, you can use the transformers library almost as usual. You just have to call optimize_model when the pipeline has loaded.

from transformers import pipeline
from nn_pruning.inference_model_patcher import optimize_model

qa_pipeline = pipeline(
    "question-answering",
    model="madlag/bert-base-uncased-squadv1-x2.01-f89.2-d30-hybrid-rewind-opt-v1",
    tokenizer="madlag/bert-base-uncased-squadv1-x2.01-f89.2-d30-hybrid-rewind-opt-v1"
)

print("bert-base-uncased parameters: 200.0M")
print(f"Parameters count (includes only head pruning, not feed forward pruning)={int(qa_pipeline.model.num_parameters() / 1E6)}M")
qa_pipeline.model = optimize_model(qa_pipeline.model, "dense")

print(f"Parameters count after complete optimization={int(qa_pipeline.model.num_parameters() / 1E6)}M")
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", predictions)

✨ Features

• Pruning: The linear layers of the model contain 30.0% of the original weights, and the model contains 45.0% of the original weights overall. This significantly reduces the model size and speeds up inference.
• Activation Function: It uses ReLUs instead of GeLUs to speed up inference, which is supported by the Transformers library.
• Accuracy: The model achieves an F1 score of 89.19, a gain of 0.69 compared to the original bert-base-uncased model.

📦 Installation

Install the nn_pruning library using the following command:

pip install nn_pruning

💻 Usage Examples

Basic Usage

from transformers import pipeline
from nn_pruning.inference_model_patcher import optimize_model

qa_pipeline = pipeline(
    "question-answering",
    model="madlag/bert-base-uncased-squadv1-x2.01-f89.2-d30-hybrid-rewind-opt-v1",
    tokenizer="madlag/bert-base-uncased-squadv1-x2.01-f89.2-d30-hybrid-rewind-opt-v1"
)

print("bert-base-uncased parameters: 200.0M")
print(f"Parameters count (includes only head pruning, not feed forward pruning)={int(qa_pipeline.model.num_parameters() / 1E6)}M")
qa_pipeline.model = optimize_model(qa_pipeline.model, "dense")

print(f"Parameters count after complete optimization={int(qa_pipeline.model.num_parameters() / 1E6)}M")
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", predictions)

📚 Documentation

Fine-Pruning details

This model was fine-tuned from the HuggingFace model checkpoint on SQuAD1.1, and distilled from the model bert-large-uncased-whole-word-masking-finetuned-squad. It is case-insensitive.

A side-effect of the block pruning is that some of the attention heads are completely removed: 55 heads were removed on a total of 144 (38.2%). You can visualize the distribution of the remaining heads in the network after pruning by hovering on the plot below.

Details of the SQuAD1.1 dataset

Fine-tuning

• Python: 3.8.5
• Machine specs:

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

Results

Pytorch model file size: 374MB (original BERT: 420MB)

🔧 Technical Details

This model was created using the nn_pruning python library. It uses NoNorms instead of LayerNorms, which requires the use of the optimize_model function from the nn_pruning library. The pruning method leads to structured matrices, which allows the model to run 2.01x as fast as bert-base-uncased on the evaluation.

📄 License

This model is released under the MIT license.