{"id":25781307,"url":"https://github.com/ekellbuch/longtail_ensembles","last_synced_at":"2026-05-15T16:02:52.558Z","repository":{"id":207036134,"uuid":"638987112","full_name":"ekellbuch/longtail_ensembles","owner":"ekellbuch","description":"Evaluating ensemble performance in long-tailed datasets (Neurips 2023 Heavy Tails Workshop)","archived":false,"fork":false,"pushed_at":"2024-04-26T23:16:25.000Z","size":1659,"stargazers_count":1,"open_issues_count":0,"forks_count":0,"subscribers_count":2,"default_branch":"main","last_synced_at":"2025-02-27T07:41:26.914Z","etag":null,"topics":["class-imbalance","ensemble-learning","fairness-ml","imbalanced-classes","imbalanced-classification","imbalanced-data","imbalanced-learning"],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/ekellbuch.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null}},"created_at":"2023-05-10T14:16:09.000Z","updated_at":"2024-04-26T23:16:29.000Z","dependencies_parsed_at":"2023-11-13T19:59:27.983Z","dependency_job_id":"e316db2c-778a-4e50-bf18-a2f02ec1777b","html_url":"https://github.com/ekellbuch/longtail_ensembles","commit_stats":null,"previous_names":["ekellbuch/longtail_ensembles"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/ekellbuch/longtail_ensembles","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ekellbuch%2Flongtail_ensembles","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ekellbuch%2Flongtail_ensembles/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ekellbuch%2Flongtail_ensembles/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ekellbuch%2Flongtail_ensembles/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/ekellbuch","download_url":"https://codeload.github.com/ekellbuch/longtail_ensembles/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ekellbuch%2Flongtail_ensembles/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":33071582,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-15T11:35:32.926Z","status":"ssl_error","status_checked_at":"2026-05-15T11:35:31.362Z","response_time":103,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.5:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["class-imbalance","ensemble-learning","fairness-ml","imbalanced-classes","imbalanced-classification","imbalanced-data","imbalanced-learning"],"created_at":"2025-02-27T07:01:42.905Z","updated_at":"2026-05-15T16:02:52.514Z","avatar_url":"https://github.com/ekellbuch.png","language":"Python","funding_links":[],"categories":["2.2 Ensemble Learning"],"sub_categories":[],"readme":"# The Effects of Ensembling on Long-Tailed Data\n\nCode for the paper [\"The Effects of Ensembling on Long-Tailed Data\"](https://openreview.net/pdf?id=l4GYs60kre) where we perform a systematic comparison between logit and probability ensembling for a variety\nof models trained on balanced and imbalanced datasets.\n\n## Findings:\n- Adding more ensemble members continues to improve performance on imbalanced datasets.\n- No difference between logit and probability ensembles across a variety of balanced datasets. \n- There are differences between logit and probability ensembles on imbalanced datasets depending on the ensemble diversity and dependency. \n\n![Fig 1: Comparison between logit and probability ensembles of models trained on CIFAR10-LT](https://github.com/ekellbuch/longtail_ensembles/blob/main/results/figures/compare_M/compare_ensemble.png)\n\n\n![Table 3: Ensembles outperform common approaches to handle long-tails](https://github.com/ekellbuch/longtail_ensembles/blob/main/results/figures/model_performance/ensemble_cifar100.png)\n\n```\n@inproceedings{\nbuchanan2023the,\ntitle={The Effects of Ensembling on Long-Tailed Data},\nauthor={E. Kelly Buchanan and Geoff Pleiss and Yixin Wang and John Patrick Cunningham},\nbooktitle={NeurIPS 2023 Workshop Heavy Tails in Machine Learning},\nyear={2023}\n}\n```\nInstallation instructions in docs/README.md: [docs/README.md](docs/README.md)\n\n## Experiments:\n1. Train resnet32 model on CIFAR10 dataset\n```\npython scripts/run.py --config-name=\"run_gpu_cifar10\"\n```\n2. Train models on CIFAR10LT dataset across multiple losses\n```\nwandb sweep experiments/compare_loss/train_gpu_loss_cifar10.yaml\n```\n3. Train additional models on CIFAR10LT. \n```\nwandb sweep experiments/compare_loss/train_gpu_loss_cifar10_largeM.yaml\n```\n\n## Paper Experiments\n\n|                                                                                                 Wandb Experiment                                                                                                  |                                                        parameters                                                        |                                                                                                                                     comments                                                                                                                                      |\n|:-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|:------------------------------------------------------------------------------------------------------------------------:|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|\n| [nggmmw4m](https://wandb.ai/ekellbuch/uncategorized/sweeps/nggmmw4m) , [0itowy8a](https://wandb.ai/ekellbuch/uncategorized/sweeps/0itowy8a), [d4s9wp4v](https://wandb.ai/ekellbuch/uncategorized/sweeps/d4s9wp4v) |   train resnet32 and resnet110 models on CIFAR10-LT using multiple losses and for different seeds. (IMBALANCECIFAR10)    |     models trained using balanced softmax loss have best performance                                                                                                  \n|                       [9hwaytks](https://wandb.ai/ekellbuch/longtail_ensembles-scripts/sweeps/9hwaytks), [gv4bucon](https://wandb.ai/ekellbuch/longtail_ensembles-scripts/sweeps/gv4bucon)                        | train resnet32_cfa and resnet_110 on CIFAR100-LT using multiple losses and for difference seeds.  (IMBALANCECIFAR100Aug) |   models trained using balanced softmax loss have best performance        \n\n\n## Reproduce paper tables and figures:\n- [x] Fig: Ensemble size vs ensemble type across multiple losses\n```\npython scripts/vis_scripts/plot_results_metric_M.py --config-path=\"../../results/configs/comparison_baseline_cifar10lt\" --config-name=\"compare_M\"\n```\n- [x] Table: Ensemble performance of models trained on  CIFAR10-LT and CIFAR100-LT:\n```\npython scripts/compare_all_results.py --config-path=\"../results/configs/comparison_baseline_cifar10lt\" --config-name=\"default\"\npython scripts/compare_all_results.py --config-path=\"../results/configs/comparison_baseline_cifar100lt\" --config-name=\"default\"\n```\n- [x] Fig: Class ID vs avg. Disagreement:\n```\npython scripts/vis_scripts/plot_results_pclass.py \n```\n- [x] Fig: Class ID vs diversity/dependency:\n```\npython scripts/vis_scripts/plot_results_dkl_diff.py \n```\n- [x] Fig: performance of logit and probability ensembles on balanced datasets. \n```\npython scripts/vis_scripts/plot_single_metric_xy.py --datasets=base --metric=error\n```\n## References:\n-  Balanced Meta Softmax: [github.com/jiawei-ren/BalancedMetaSoftmax-Classification](https://github.com/jiawei-ren/BalancedMetaSoftmax-Classification)\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fekellbuch%2Flongtail_ensembles","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fekellbuch%2Flongtail_ensembles","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fekellbuch%2Flongtail_ensembles/lists"}