Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Probabilistic Neural Pruning via Sparsity Evolutionary Fokker-Planck-Kolmogorov Equation
Authors: Zhanfeng Mo, Haosen Shi, Sinno Jialin Pan
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our SFPK-pruner exhibits competitive performance in various pruning scenarios. The code is available on Git Hub1. ... 5 EXPERIMENTS We evaluate our SFPK-pruner for unstructured and structured pruning in one-shot and gradual pruning settings across various deep vision models. The evaluation process of each pruning method involves 3 steps: 1. Prune the well-trained target network θ to a target sparsity level. 2. Retrain the pruned model for a fixed number of epochs until convergence. 3. Report the averaged prediction accuracy over the last few retraining epochs. The experiment settings are detailed in Appendix A.2. ... 5.3 ABLATION STUDIES |
| Researcher Affiliation | Academia | Zhanfeng Mo Haosen Shi Sinno Jialin Pan Nanyang Technological University, Singapore The Chinese University of Hong Kong EMAIL; EMAIL; EMAIL |
| Pseudocode | Yes | As outlined in Algorithm 1 (a detailed version is presented in Algorithm 2 and Algorithm 3 in Appendix A.1), the SFPK-pruner aims to simulate the evolution of t 7 πt using the evolution of the uniform distribution of n interacting mask particles. |
| Open Source Code | Yes | The code is available on Git Hub1. 1https://github.com/mzf666/SFPK-main |
| Open Datasets | Yes | 5.1 EXPERIMENTS ON CIFAR-100 We conduct a comparison of the one-shot pruning performance of our SFPK-pruner ... on CIFAR-100 (Krizhevsky, 2009). ... 5.2 EXPERIMENTS ON IMAGENET-1K ... on the Image Net-1K benchmark (Deng et al., 2009) |
| Dataset Splits | No | The paper mentions using CIFAR-100 and ImageNet-1K datasets, which typically have standard train/validation/test splits. However, it does not explicitly state the specific percentages, sample counts, or methodology used for these splits within the provided text. It details pruning steps and retraining epochs, but not how the datasets themselves were partitioned for training, validation, and testing. |
| Hardware Specification | Yes | Each run takes around 20 to 30 minutes on a NVIDIA 40G A100 GPU, with each SFPK update step (line 7 to line 11 of Algorithm 2) taking approximately 0.2 seconds. ... On an NVIDIA 40G A100 GPU, each run on Mobile Net-1K takes about 1.5 days, each run on Res Net-50 takes about 2 days, and each run on Dei T-T takes about 4 days. |
| Software Dependencies | No | The paper mentions |
| Experiment Setup | Yes | For all experiments, we set λ = 0.2, n = 10, and K = 100 for SFPK-pruners. ... We ran the SFPK-pruner on all the experiments with λ = 0.2, n = 10, and K = 150. ... Table 3: Experiment settings of one-shot pruning on CIFAR-100. Here, lr denotes learning rate , and Cos denotes the cosine annealing learning rate schedule, with an initial learning rate of 1e4 and 5 warm-up epochs. ... Retrain setting: Batch size 64, Epochs 100, lr 0.01, lr schedule Cos, Optimizer SGD, Momentum 0.875, Weight decay 0. SFPK setting: Mini-batch size 512, Simulation steps K 100, # mask particles n 10, Regularization rate λ 0.2, Localization radius rt 1.1, Mask polarizer Pε One-hot. |