{"id":19382854,"url":"https://github.com/locuslab/flyp","last_synced_at":"2025-10-09T13:06:47.403Z","repository":{"id":96318609,"uuid":"573049340","full_name":"locuslab/FLYP","owner":"locuslab","description":"Code for Finetune like you pretrain: Improved finetuning of zero-shot vision models","archived":false,"fork":false,"pushed_at":"2023-08-13T21:30:05.000Z","size":2923,"stargazers_count":98,"open_issues_count":6,"forks_count":14,"subscribers_count":5,"default_branch":"main","last_synced_at":"2025-04-23T20:47:52.415Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"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/locuslab.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":"2022-12-01T15:35:17.000Z","updated_at":"2025-02-28T12:52:11.000Z","dependencies_parsed_at":"2024-11-10T09:25:33.144Z","dependency_job_id":null,"html_url":"https://github.com/locuslab/FLYP","commit_stats":null,"previous_names":[],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/locuslab/FLYP","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/locuslab%2FFLYP","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/locuslab%2FFLYP/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/locuslab%2FFLYP/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/locuslab%2FFLYP/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/locuslab","download_url":"https://codeload.github.com/locuslab/FLYP/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/locuslab%2FFLYP/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":279001541,"owners_count":26083102,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","status":"online","status_checked_at":"2025-10-09T02:00:07.460Z","response_time":59,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"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":[],"created_at":"2024-11-10T09:23:37.010Z","updated_at":"2025-10-09T13:06:47.386Z","avatar_url":"https://github.com/locuslab.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# FLYP: Finetune Like You Pretrain\n\nCode for the paper Finetune like you pretrain: Improved finetuning of zero-shot vision models.\n\nCREDITS: Our code is heavily based on https://github.com/mlfoundations/wise-ft and https://github.com/mlfoundations/open_clip. We thank the authors for open sourcing their code.\n\n## Setting up conda env\n```bash\nconda create -n flyp python=3.10\nconda activate flyp\npip install open_clip_torch\npip install wilds braceexpand webdataset h5py\npip install git+https://github.com/modestyachts/ImageNetV2_pytorch\nmkdir checkpoints\n```\n\n### Add directory to PYTHONPATH:\n\n```bash\ncd FLYP\nexport PYTHONPATH=\"$PYTHONPATH:$PWD\"\n```\n\n### Datasets\nAll the datasets we use are available publicly.\nRefer to the DATA.md for the respective dataset directory strucutre.\n\n### Script to reproduce on ImageNet\n```bash\nln -s PATH_TO_YOUR_ILSVRC2012_DATASET ./datasets/data/ILSVRC2012\n\npython datacreation_scripts/imagenet_csv_creator.py\n\npython src/main.py --train-dataset=ImageNet --epochs=10 --lr=1e-5 --wd=0.1 --batch-size=512 --model=ViT-B/16 --eval-datasets=ImageNet,ImageNetV2,ImageNetR,ImageNetA,ImageNetSketch,ObjectNet  --template=openai_imagenet_template  --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/imagenet.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=ImageNet/flyp_loss\n```\n\nImageNet Finetuned CLIP ViT-B-16 checkpoint can be found [here](https://drive.google.com/drive/folders/1oRPXybgzTp4lmY66XNq1n_LfN7MJHNeM).\n\n### Script to reproduce on iWILDCam\n```bash\nln -s PATH_TO_YOUR_iWILDCam_DATASET ./datasets/data/iwildcam_v2.0\n\npython datacreation_scripts/iwildcam.py\n\npython src/main.py --train-dataset=IWildCamIDVal --epochs=20 --lr=1e-5 --wd=0.2 --batch-size=256 --model=ViT-B/16 --eval-datasets=IWildCamIDVal,IWildCamID,IWildCamOOD --template=iwildcam_template  --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/iwildcam_v2.0/train.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=iwildcam/flyp_loss\n```\n\n### Script to reproduce on FMOW\n```bash\nln -s PATH_TO_YOUR_FMOW_DATASET ./datasets/data/fmow_v1.1\n\npython datacreation_scripts/fmow_csv_creator.py\n\npython src/main.py --train-dataset=FMOWIDVal --epochs=20 --lr=1e-5 --wd=0.2 --batch-size=256 --model=ViT-B/16 --eval-datasets=FMOWIDVal,FMOWID,FMOWOOD --template=fmow_template --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/fmow.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=fmow/flyp_loss\n```\n\n### Few shot on SST2\n\n```bash\n\nln -s PATH_TO_YOUR_SST2_DATASET ./datasets/data/sst2\n\npython datacreation_scripts/sst2.py\n\narch=\"ViT-B/16\"\nk=16\n\npython src/few_shot.py --train-dataset=sst2Val --epochs=20 --lr=1e-5 --wd=0.2 --batch-size=256 --model=$arch --warmup_length 0 --eval-datasets=sst2Val,sst2Test --template=sst2_template  --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/sst2/train.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=sst2/\"flyp_loss_\"$k\"shot\" --k=$k \n```\n\nFor VitL\n```\narch=\"ViT-L/14\"\n```\n\n### Few shot on PatchCamelyon\n\n```bash\nln -s PATH_TO_YOUR_PATCH_CAM_DATASET ./datasets/data/patchcamelyon\n\npython datacreation_scripts/patchcamelyon.py\n\narch=\"ViT-B/16\"\nk=16\n\npython src/few_shot.py --train-dataset=PatchCamelyonVal --epochs=20 --lr=1e-6 --wd=0.0 --batch-size=256 --model=$arch --warmup_length 0 --eval-datasets=PatchCamelyonVal,PatchCamelyonTest --template=patchcamelyon_template  --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/patchcamelyon/train.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=patchcamelyon/\"flyp_loss_\"$k\"shot\" --k=$k \n```\n\n### Transfer Learning on Caltech\n\n```bash\nln -s PATH_TO_YOUR_CALTECH_DATASET ./datasets/data/caltech-101\n\npython datacreation_scripts/caltech101.py\n\npython src/main.py --train-dataset=Caltech101Val --epochs=100 --lr=1e-5 --wd=0.0 --batch-size=256 --model=ViT-B/16 --warmup_length 500 --eval-datasets=Caltech101Val,Caltech101Test --template=caltech101_template  --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/caltech101/train.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=caltech101/flyp_loss\n```\n\n### Transfer Learning on StanfordCars\n\n```bash\nln -s PATH_TO_YOUR_STANFORD_CARS_DATASET ./datasets/data/StanfordCars\n\npython datacreation_scripts/stanfordCars.py\n\npython src/main.py --train-dataset=StanfordCarsVal --epochs=100 --lr=1e-5 --wd=0.0 --batch-size=256 --model=ViT-B/16 --warmup_length 500 --eval-datasets=StanfordCarsVal,StanfordCarsTest --template=stanfordcars_template  --save=./checkpoints/ --data-location=./datasets/data/ --ft_data=\"./datasets/csv/StanfordCars/train.csv\" --csv-img-key filepath --csv-caption-key title --exp_name=stanfordcars/flyp_loss\n```\n\n### Cross Entropy Ablation on ImageNet\nThis refers to the cross-entropy ablation, where we use language representations as a linear head over the image representations, projecting the image representations to class probabilities. Simply add a flag by --ce_ablation to any of the above cmd line scripts. Here we provide the cmd line script for ImageNet.\n\n```bash\npython src/main.py --train-dataset=ImageNet --epochs=10 --lr=1e-5 --wd=0.1 --batch-size=512 --model=ViT-B/16 --eval-datasets=ImageNet,ImageNetV2,ImageNetR,ImageNetA,ImageNetSketch,ObjectNet  --template=openai_imagenet_template  --save=./checkpoints/ --data-location=./datasets/data/ --exp_name=ImageNet/ce_ablation --ce_ablation\n```\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flocuslab%2Fflyp","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Flocuslab%2Fflyp","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flocuslab%2Fflyp/lists"}