SlimSAM: 0.1% Data Makes Segment Anything Slim

Authors: Zigeng Chen, Gongfan Fang, Xinyin Ma, Xinchao Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive assessments across performance metrics, efficiency, and training data requirements reveal that Slim SAM markedly enhances compression performance, concurrently achieving superior lightweight and efficiency with markedly reduced training data requirements.
Researcher Affiliation Academia Zigeng Chen, Gongfan Fang, Xinyin Ma, Xinchao Wang National University of Singapore zigeng99@u.nus.edu, xinchao@nus.edu.sg
Pseudocode No The paper describes the method conceptually and with diagrams (Figure 1, Figure 2) but does not include formal pseudocode blocks or algorithms.
Open Source Code Yes Code is available at https://github.com/czg1225/Slim SAM
Open Datasets Yes Our Slim SAM has been implemented in Py Torch [41] and trained on a single Nvidia Titan RTX GPU using only 0.1% (10,000 images) of the SA-1B [25] dataset.
Dataset Splits No The paper mentions 'validation performance' for early stopping, but it does not provide specific details on how the 10k training images are split for validation or if a separate validation set is used from SA-1B.
Hardware Specification Yes Our Slim SAM has been implemented in Py Torch [41] and trained on a single Nvidia Titan RTX GPU
Software Dependencies No The paper mentions 'PyTorch' as the implementation framework, but it does not specify its version number or any other software dependencies with their versions.
Experiment Setup Yes The model s parameters were optimized through the ADAM [24] algorithm with a batch size of 4. Training settings for both bottleneck aligning and embedding aligning are identical. The pruned models undergo distillation with an initial learning rate of 1e 4, which will be reduced by half if validation performance does not improve for 4 consecutive epochs. The total training duration is 40 epochs for Slim SAM-50 (with a 50% pruning ratio) and 80 epochs for Slim SAM-77 (with a 77% pruning ratio).