madlag/

bert-base-uncased-squadv1-x1.96-f88.3-d27-hybrid-filled-opt-v1

This model was created using the nn_pruning python library: the linear layers contains 27.0% of the original weights.

This model CANNOT be used without using nn_pruning optimize_model function, as it uses NoNorms instead of LayerNorms and this is not currently supported by the Transformers library.

It uses ReLUs instead of GeLUs as in the initial BERT network, to speedup inference. This does not need special handling, as it is supported by the Transformers library, and flagged in the model config by the "hidden_act": "relu" entry.

The model contains 43.0% of the original weights overall (the embeddings account for a significant part of the model, and they are not pruned by this method).

With a simple resizing of the linear matrices it ran 1.96x as fast as bert-base-uncased on the evaluation. This is possible because the pruning method lead to structured matrices: to visualize them, hover below on the plot to see the non-zero/zero parts of each matrix.

In terms of accuracy, its F1 is 88.33, compared with 88.5 for bert-base-uncased, a F1 drop of 0.17.

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 This model is case-insensitive: it does not make a difference between english and English.

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

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

Install nn_pruning: it contains the optimization script, which just pack the linear layers into smaller ones by removing empty rows/columns.

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