Learning to Prune Image Captioning Models

PDF | ArXiv

Official pytorch implementation of the paper: "End-to-End Supermask Pruning: Learning to Prune Image Captioning Models"

Published at Pattern Recognition, Elsevier

Released on July 20, 2021

Description

This work explores model pruning for image captioning task at the first time. Empirically, we show that 80% to 95% sparse networks can either match or even slightly outperform their dense counterparts. In order to promote Green Computer Vision, we release the pre-trained sparse models for UD and ORT that are capable of achieving CIDEr scores >120 on MS-COCO dataset; yet are only 8.7 MB (reduction of 96% compared to dense UD) and 14.5 MB (reduction of 94% compared to dense ORT) in model size.

Figure 1: We show that our deep captioning networks with 80% to 95% sparse are capable of either matching or even slightly outperforming their dense counterparts.

Get Started

Please refer to the documentation.

Features

• Captioning models built using PyTorch
  • Up-Down LSTM
  • Object Relation Transformer
  • A Compact Object Relation Transformer (ACORT)
• Unstructured weight pruning
  • Supermask Pruning (SMP, end-to-end pruning)
  • Gradual magnitude pruning
  • Lottery ticket
  • One-shot magnitude pruning (paper 1, paper 2)
  • Single-shot Network Pruning (SNIP)
• Self-Critical Sequence Training (SCST)]()
  • Random sampling + Greedy search baseline: vanilla SCST
  • Beam search sampling + Greedy search baseline: à la Up-Down
  • Random sampling + Sample mean baseline: arxiv paper
  • Beam search sampling + Sample mean baseline: à la M2 Transformer
  • Optimise CIDEr and/or BLEU scores with custom weightage
  • Based on ruotianluo/self-critical.pytorch
• Multiple captions per image during teacher-forcing training
  • Reduce training time: run encoder once, optimize on multiple training captions
• Incremental decoding (Transformer with attention cache)
• Data caching during training
  • Training examples will be cached in memory to reduce disk I/O
  • With sufficient memory, the entire training set can be loaded from memory after the first epoch
  • Memory usage can be controlled via cache_min_free_ram flag
• coco_caption in Python 3
  • Based on salaniz/pycocoevalcap
• Tokenizer based on sentencepiece
  • Word
  • Radix encoding
  • (untested) Unigram, BPE, Character
• Datasets
  • MS-COCO
  • (contributions welcome) Flickr8k, Flickr30k, InstaPIC-1.1M

Pre-trained Sparse and ACORT Models

The checkpoints are available at this repo.

Soft-attention models implemented in TensorFlow 1.9 are available at this repo.

CIDEr score of pruning methods

Up-Down (UD)

Sparsity	NNZ	Dense Baseline	SMP	Lottery ticket (class-blind)	Lottery ticket (class-uniform)	Lottery ticket (gradual)	Gradual pruning	Hard pruning (class-blind)	Hard pruning (class-distribution)	Hard pruning (class-uniform)	SNIP
0.950	2.7 M	111.3	112.5	-	107.7	109.5	109.7	-	110.0	110.2	38.2
0.975	1.3 M	111.3	110.6	-	103.8	106.6	107.0	-	105.9	105.4	34.7
0.988	0.7 M	111.3	109.0	-	99.3	102.2	103.4	-	101.3	100.5	32.6
0.991	0.5 M	111.3	107.8

Object Relation Transformer (ORT)

Sparsity	NNZ	Dense Baseline	SMP	Lottery ticket (gradual)	Gradual pruning	Hard pruning (class-blind)	Hard pruning (class-distribution)	Hard pruning (class-uniform)	SNIP
0.950	2.8 M	114.7	113.7	115.7	115.3	4.1	112.5	113.0	47.2
0.975	1.4 M	114.7	113.7	112.9	113.2	0.7	106.6	106.9	44.0
0.988	0.7 M	114.7	110.7	109.8	110.0	0.9	96.9	59.8	37.3
0.991	0.5 M	114.7	109.3	107.1	107.0

Acknowledgements

• SCST, Up-Down: ruotianluo/self-critical.pytorch
• Object Relation Transformer: yahoo/object_relation_transformer
• coco_caption in Python 3: salaniz/pycocoevalcap

Citation

If you find this work useful for your research, please cite

@article{tan2021end,
  title={End-to-End Supermask Pruning: Learning to Prune Image Captioning Models},
  author={Tan, Jia Huei and Chan, Chee Seng and Chuah, Joon Huang},
  journal={Pattern Recognition},
  pages={108366},
  year={2021},
  publisher={Elsevier},
  doi={10.1016/j.patcog.2021.108366}
}

Feedback

Suggestions and opinions on this work (both positive and negative) are greatly welcomed. Please contact the authors by sending an email to tan.jia.huei at gmail.com or cs.chan at um.edu.my.

License and Copyright

The project is open source under BSD-3 license (see the LICENSE file).

©2021 Universiti Malaya.

Dev Info

Run Black linting:

black --line-length=120 --safe sparse_caption
black --line-length=120 --safe tests
black --line-length=120 --safe scripts