Bert Base Uncased Sst2 Acc91.1 D37 Hybrid

🚀 bert-base-uncased model fine-tuned on SST-2

This project presents a fine - tuned bert-base-uncased model on the SST - 2 dataset. It leverages the nn_pruning library to optimize the model's weights, achieving a balance between performance and model size.

🚀 Quick Start

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

pip install nn_pruning

Then, you can use the transformers library as usual, but remember to call optimize_model after the pipeline has loaded.

from transformers import pipeline
from nn_pruning.inference_model_patcher import optimize_model

cls_pipeline = pipeline(
    "text-classification",
    model="echarlaix/bert-base-uncased-sst2-acc91.1-d37-hybrid",
    tokenizer="echarlaix/bert-base-uncased-sst2-acc91.1-d37-hybrid",
)

print(f"Parameters count (includes only head pruning, no feed forward pruning)={int(cls_pipeline.model.num_parameters() / 1E6)}M")
cls_pipeline.model = optimize_model(cls_pipeline.model, "dense")
print(f"Parameters count after optimization={int(cls_pipeline.model.num_parameters() / 1E6)}M")
predictions = cls_pipeline("This restaurant is awesome")
print(predictions)

✨ Features

• Weight Optimization: The linear layers of the model contain 37% of the original weights, and the model contains 51% of the original weights overall.
• Performance: It achieves an accuracy of 91.17 on the SST - 2 dataset.
• Case - Insensitive: The model does not distinguish between different cases of English words.

📦 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

cls_pipeline = pipeline(
    "text-classification",
    model="echarlaix/bert-base-uncased-sst2-acc91.1-d37-hybrid",
    tokenizer="echarlaix/bert-base-uncased-sst2-acc91.1-d37-hybrid",
)

print(f"Parameters count (includes only head pruning, no feed forward pruning)={int(cls_pipeline.model.num_parameters() / 1E6)}M")
cls_pipeline.model = optimize_model(cls_pipeline.model, "dense")
print(f"Parameters count after optimization={int(cls_pipeline.model.num_parameters() / 1E6)}M")
predictions = cls_pipeline("This restaurant is awesome")
print(predictions)

📚 Documentation

Fine - Pruning details

This model was fine - tuned from the HuggingFace model checkpoint on the task and distilled from the model textattack/bert-base-uncased-SST-2.

A side - effect of the block pruning method is that some of the attention heads are completely removed: 88 heads were removed out of a total of 144 (61.1%).

Details of the SST - 2 dataset

Results

Pytorch model file size: 351MB (original BERT: 420MB)

🔧 Technical Details

The model uses the nn_pruning python library to optimize the linear layers. The embeddings, which account for a significant part of the model, are not pruned by this method. The block pruning method also affects the attention heads in the model.

📄 License

This project is licensed under the Apache - 2.0 license.