Changing legend.

b0c32a5 over 3 years ago

No virus

4.03 kB

	---
	language: en
	thumbnail:
	license: mit
	tags:
	- question-answering
	- bert
	- bert-base
	datasets:
	- squad
	metrics:
	- squad
	widget:
	- text: "Where is the Eiffel Tower located?"
	context: "The Eiffel Tower is a wrought-iron lattice tower on the Champ de Mars in Paris, France. It is named after the engineer Gustave Eiffel, whose company designed and built the tower."
	- text: "Who is Frederic Chopin?"
	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."
	---

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

	This model was created using the [nn_pruning](https://github.com/huggingface/nn_pruning) python library: the linear layers contains 64.0% of the original weights.

	The model contains 36.0% of the original weights overall (remaining weights are mostly in the input/ouput embeddings).

	With a simple resizing of the linear matrices it ran 1.84x as fast as BERT-base 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.

	<div class="graph"><script src="/madlag/bert-base-uncased-squadv1-x1.84-f88.7-d36-hybrid-filled-v1/raw/main/model_card/density_info.js" id="6a33cbcb-db1a-4e7a-a6dc-1c353babcf55"></script></div>

	In terms of accuracy, its F1 is 88.72, compared with 88.5 for BERT-base, a F1 gain of 0.22.

	## Fine-Pruning details
	This model was fine-tuned from the HuggingFace [BERT](https://www.aclweb.org/anthology/N19-1423/) base uncased checkpoint on [SQuAD1.1](https://rajpurkar.github.io/SQuAD-explorer), and distilled from the equivalent model [csarron/bert-base-uncased-squad-v1](https://huggingface.co/csarron/bert-base-uncased-squad-v1).
	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: 48 heads were removed on a total of 144 (33.3%).
	Here is a detailed view on how the remaining heads are distributed in the network after pruning.
	<div class="graph"><script src="/madlag/bert-base-uncased-squadv1-x1.84-f88.7-d36-hybrid-filled-v1/raw/main/model_card/pruning_info.js" id="12c50081-be7a-4612-a727-67c9c9309cc9"></script></div>

	## Details of the SQuAD1.1 dataset

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

	### Fine-tuning
	- Python: `3.8.5`

	- Machine specs:

	```CPU: Intel(R) Core(TM) i7-6700K 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: `379M` (original BERT: `438M`)

	\| Metric \| # Value \| # Original ([Table 2](https://www.aclweb.org/anthology/N19-1423.pdf))\| Variation \|
	\| ------ \| --------- \| --------- \| --------- \|
	\| EM \| 81.69 \| 80.8 \| +0.89\|
	\| F1 \| 88.72 \| 88.5 \| +0.22\|

	## Example Usage

	```python
	from transformers import pipeline
	from nn_pruning.inference_model_patcher import optimize_model

	qa_pipeline = pipeline(
	"question-answering",
	model="madlag/bert-base-uncased-squadv1-x1.84-f88.7-d36-hybrid-filled-v1",
	tokenizer="madlag/bert-base-uncased-squadv1-x1.84-f88.7-d36-hybrid-filled-v1"
	)

	print("BERT-base parameters: 110M")
	print(f"Parameters count (includes head pruning)={int(qa_pipeline.model.num_parameters() / 1E6)}M")
	qa_pipeline.model = optimize_model(qa_pipeline.model, "dense")

	print(f"Parameters count after 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)
	```