Flopping for FLOPs: Leveraging Equivariance for Computational Efficiency
Authors: Georg Bökman, David Nordström, Fredrik Kahl
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the effectiveness of flopping-equivariant networks. We first briefly discuss the setting of the experiments. Then we compare the efficiency and accuracy of equivariant versions of Res MLPs, Vi Ts and Conv Ne Xts from Section 4 to their non-equivariant counterparts. For a given architecture X, the equivariant version is E(X). E(X) always has around half the number of trainable parameters and FLOPs of X due to the block-diagonalization of the weight matrices (2). We will release code and weights at github.com/georg-bn/flopping-for-flops. |
| Researcher Affiliation | Academia | 1Chalmers University of Technology. Correspondence to: Georg B okman <EMAIL>. |
| Pseudocode | No | The paper describes the mathematical foundations and architectural changes in detail, but it does not include any explicit pseudocode blocks or algorithms labeled as such. |
| Open Source Code | Yes | We will release code and weights at github.com/georg-bn/flopping-for-flops. |
| Open Datasets | Yes | Dataset. We benchmark our model implementations on the Image Net-1K dataset (Deng et al., 2009; Russakovsky et al., 2015; Recht et al., 2019), which includes 1.2M images evenly spread over 1,000 object categories. |
| Dataset Splits | Yes | Dataset. We benchmark our model implementations on the Image Net-1K dataset (Deng et al., 2009; Russakovsky et al., 2015; Recht et al., 2019), which includes 1.2M images evenly spread over 1,000 object categories. |
| Hardware Specification | Yes | Hardware. All experiments were run on NVIDIA A100-40GB. The per GPU batch size ranged between 64 (for larger models) to 256 (for smaller models). The biggest model requires training on 32 A100 GPUs for c. 2 days. |
| Software Dependencies | Yes | Software versioning. Our experiments build upon Py Torch (Paszke et al., 2019) and the timm (Wightman, 2019) library. We enable mixed-precision training using the deprecated NVIDIA library Apex, this is to mirror the training recipes of the benchmarks as closely as possible. To enable Py Torch s compiler, we use a modern version ( 2.0). Specifically, we use Py Torch 2.5.1 with CUDA 11.8. |
| Experiment Setup | Yes | Hyperparameters. We use the same training recipes as the baselines. The complete set of hyperparameters can be found in Table 2 in the appendix. ... Table 2: Training recipes for different model architectures. We try to, as closely as possible, replicate the training recipe of the baselines. |