Equivariant Adaptation of Large Pretrained Models
Authors: Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Rajeswar Mudumba, Siamak Ravanbakhsh
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present experimental results on images and point clouds to evaluate our method of achieving equivariance with minimal modifications to pretrained networks. |
| Researcher Affiliation | Collaboration | Arnab Kumar Mondal Mila, Mc Gill University Service Now Research Siba Smarak Panigrahi Mila, Mc Gill University Sékou-Oumar Kaba Mila, Mc Gill University Sai Rajeswar Service Now Research Siamak Ravanbakhsh Mila, Mc Gill University |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/arnab39/Equivariant Adaptation |
| Open Datasets | Yes | We finetune these models on several benchmark natural image classification datasets, including CIFAR10 [31], CIFAR100 [31], and STL10 [42]. Moreover, we incorporate four different strategies to finetune the pretrained models, namely: 1) Vanilla, 2) Rotation Augmentation, 3) Learn Canonicalization, and 4) Prior-Regularized Learned Canonicalization. |
| Dataset Splits | No | The paper mentions evaluating on 'validation (val) set' for COCO and specifies training/testing splits for ModelNet40 ('9,843 models were allocated for training, while the remaining models were reserved for testing'), but it does not provide explicit percentages, sample counts, or detailed methodology for general train/test/validation splits across all datasets, nor does it refer to a standard split citation for validation. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, or cloud instances) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'escnn library [15, 50]' and that PyTorch models were used, but it does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | Our total loss is given by Ltotal = Lfine tune + β Lprior, where Lprior is defined in Eq. 3, Lfine tune refers to the cross-entropy loss for classification, and β is a hyperparameter which is set to 100. We jointly train the canonicalization function and fine-tune the pretrained image classification networks. |