SNIP: SINGLE-SHOT NETWORK PRUNING BASED ON CONNECTION SENSITIVITY

Authors: Namhoon Lee, Thalaiyasingam Ajanthan, Philip Torr

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on MNIST, CIFAR-10, and Tiny-Image Net classification tasks with a variety of network architectures. Our results show that SNIP yields extremely sparse models with minimal or no loss in accuracy across all tested architectures, while being much simpler than other state-of-the-art alternatives.
Researcher Affiliation Academia Namhoon Lee, Thalaiyasingam Ajanthan & Philip H. S. Torr University of Oxford {namhoon,ajanthan,phst}@robots.ox.ac.uk
Pseudocode Yes Algorithm 1 SNIP: Single-shot Network Pruning based on Connection Sensitivity
Open Source Code Yes The code can be found here: https://github.com/namhoonlee/snip-public.
Open Datasets Yes We evaluate our method on MNIST, CIFAR-10, and Tiny-Image Net classification tasks with a variety of network architectures.
Dataset Splits Yes We spare 10% of the training data as a validation set and used only 90% for training.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions optimizers like SGD and Adam but does not provide specific version numbers for any software dependencies or deep learning frameworks.
Experiment Setup Yes Specifically, we train the models using SGD with momentum of 0.9, batch size of 100 for MNIST and 128 for CIFAR experiments and the weight decay rate of 0.0005, unless stated otherwise. The initial learning rate is set to 0.1 and decayed by 0.1 at every 25k or 30k iterations for MNIST and CIFAR, respectively.