ChipNet: Budget-Aware Pruning with Heaviside Continuous Approximations

Authors: Rishabh Tiwari, Udbhav Bamba, Arnav Chavan, Deepak Gupta

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Chip Net outperforms state-of-the-art structured pruning methods by remarkable margins of up to 16.1% in terms of accuracy.
Researcher Affiliation Collaboration Transmute AI Research, The Netherlands Indian Institute of Technology, ISM Dhanbad, India Informatics Institute, University of Amsterdam, The Netherlands
Pseudocode Yes Algorithm 1: Chip Net Pruning Approach
Open Source Code Yes 1Code is publicly available at https://github.com/transmute AI/Chip Net
Open Datasets Yes For datasets, we have chosen CIFAR-10/100 (Krizhevsky, 2009) and Tiny Image Net (Wu et al.).
Dataset Splits Yes Finally, the model with best performance on the validation set is chosen for fine tuning.
Hardware Specification Yes All experiments were run on a Google Cloud Platform instance with a NVIDIA V100 GPU (16GB), 16 GB RAM and 4 core processor.
Software Dependencies No The paper mentions optimizers like 'Adam W' and 'SGD' but does not provide specific version numbers for these or any other software dependencies like deep learning frameworks or programming languages.
Experiment Setup Yes For the combined loss L in Eq. 1, weights α1 and α2 are set to 10 and 30, respectively, across all experiments. [...] WRN-26-12, Mobile Net V2, Res Net-50, Res Net-101, Res Net-110 were trained with batch size of 128 at initial learning rate of 5 10 2 using SGD optimizer with momentum 0.9 and weight decay 10 3. We use step learning rate strategy to decay learning rate by 0.5 after every 30 epochs.