This is the official PyTorch implementation of the ECCV 2024 paper : GalLoP: Learning Global and Local Prompts for Vision-Language Models
To install gallop and all the dependencies please use the following commands:
python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -e .
Please follow instructions in the CoOp repository to download the datasets used for the few-shot image classification and domain generalization experiments. Note that downloading the class names and splits (e.g. split_zhou_Caltech101.json) is not required as there are already included in the gallop/datasets folder. For the out-of-distribition experiments, we use the OOD datasets iNaturalist, SUN, Places, and Texture curated by Huang et al. 2021. PLease follow instructions from this repository to download the datasets.
The overall file structure is as follows:
The directory structure where all the datasets are stored should look like this:
$DATA_FOLDER/
└–––-Imagenet/
│ ├–––––train/
│ └––––val/
├––––imagenetv2-matched-frequency-format-val/
├––––imagenet-a/
├––––imagenet-r/
├––––sketch/
├––––caltech101/
│ └––––101_ObjectCategories/
├––––EuroSAT_RGB/
├––––dtddataset/
│ └–––dtd/
│ └––––images/
├––––fgvc-aircraft-2013b/
├––––flowers102/
├––––food-101/
│ └–––images/
├––––oxford-iiit-pet/
│ └–––images/
├––––stanford_cars/
├––––SUN397/
├––––UCF-101-midframes/
└––––ood_data/
├–––iNaturalist/
├–––SUN/
├–––Places/
└–––dtd/
└––––images/
We provide training scripts in the scripts folder. For instance, to launch the training on Imagenet with the ViT-B/16 backbone and with 16 shots do:
./scripts/run_imagenet.sh
Note that you may adapt the values of CUDA_VISIBLE_DEVICES and checkpointing_segments to launch on your own GPUs. For reference, we trained the Imagenet ViT-B/16 model on 2 NVIDIA RTX A6000 in approx. 6h 38 min.
Here is a simple snippet code illustrating how to perform inference with gallop:
import torch
from PIL import Image
from clip.clip import _transform
import gallop.lib as lib
from gallop.vlprompt import GalLoP
checkpoint_path = "your_ckpt_path" # Ajust the path and name to the checkpoint if needed.
device = "cuda" if torch.cuda.is_available() else "cpu"
val_transform = _transform(224)
model = GalLoP(clip_name="clip_vit_b16", n_global_prompts=4, n_local_prompts=4)
lib.load_checkpoint(model, checkpoint_path)
model.to(device)
image = val_transform(Image.open(".github/horses.png")).unsqueeze(0).to(device)
class_names = ["horses", "dogs", "cats"]
with torch.no_grad():
# Compute text features from global and local prompts:
text_features, local_text_features = model.encode_text(class_names)
text_features /= text_features.norm(dim=-1, keepdim=True)
local_text_features /= local_text_features.norm(dim=-1, keepdim=True)
# Compute global and local logits from global and local visual features:
global_logits, local_logits = model(image, text_features=text_features, local_text_features=local_text_features)
# Compute out-of-distribution score using Global-Local Maximum Concept Matching:
ood_scores = model.compute_scores(global_logits, local_logits, ood_method="GL-MCM")
# Perform ensembling of logits to obtain final class probabilities
probs, _, _ = model.create_prediction_scores(global_logits, local_logits)
print("Label probs:", probs.cpu().numpy())
print("OOD score:", ood_scores)To run the full evaluation of GalLoP (accuracy, ood detection and ood generalization) on the Imagenet ViT-B/16 checkpoint please run the following script (you may need to change the path and/or name to the checkpoint):
./scripts/eval_imagenet.sh "your_ckpt_path"
| Dataset | Imagenet | Caltech | Pets | Cars | Flowers | Food | Aircraft | SUN | DTD | Eurosat | UCF | Average |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CLIP | 66.7 | 92.2 | 88.4 | 65.5 | 70.7 | 84.8 | 24.8 | 62.3 | 44.1 | 48.3 | 64.7 | 75.7 |
| CoOp | 71.7 | 95.6 | 91.9 | 83.1 | 97.1 | 84.2 | 43.4 | 74.7 | 69.9 | 84.9 | 82.2 | 79.9 |
| PLOT | 72.6 | 96.0 | 93.6 | 84.6 | 97.6 | 87.1 | 46.7 | 76.0 | 71.4 | 92.0 | 85.3 | 82.1 |
| PromptSRC | 73.2 | 96.1 | 93.7 | 85.8 | 97.6 | 86.5 | 50.8 | 77.2 | 72.7 | 92.4 | 86.5 | 82.9 |
| GalLoP | 75.1 | 96.7 | 94.1 | 89.2 | 98.8 | 86.5 | 58.3 | 77.2 | 75.5 | 90.1 | 86.9 | 84.4 |
| Dataset | Imagenet | Imagenet-V2 | Imagenet-S | Imagenet-A | Imagenet-R | Average |
|---|---|---|---|---|---|---|
| CLIP | 66.7 | 60.8 | 46.2 | 47.8 | 74.0 | 57.2 |
| CoOp | 71.7 | 64.6 | 47.9 | 49.9 | 75.1 | 59.4 |
| PLOT | 72.6 | 64.9 | 46.8 | 48.0 | 73.9 | 58.4 |
| PromptSRC | 71.3 | 64.4 | 49.6 | 50.9 | 77.8 | 60.7 |
| GalLoP | 75.1 | 67.5 | 49.5 | 50.3 | 77.8 | 61.3 |
| Dataset | Imagenet Accuracy |
iNaturalist FPR95 / AUC |
SUN FPR95 / AUC |
Places FPR95 / AUC |
Textures FPR95 / AUC |
Average FPR95 / AUC |
|---|---|---|---|---|---|---|
| CLIP (MCM) | 66.7 | 30.9 / 94.6 | 37.7 / 92.6 | 44.8 / 89.8 | 57.9 / 86.1 | 42.8 / 90.8 |
| CLIP (GL-MCM) | 66.7 | 15.2 / 96.7 | 30.4 / 93.1 | 38.9 / 89.9 | 57.9 / 83.6 | 35.5 / 90.8 |
| CoOp (MCM) | 71.7 | 28.0 / 94.4 | 37.0 / 92.3 | 43.0 / 89.7 | 39.3 / 91.2 | 36.8 / 91.9 |
| CoOp (GL-MCM) | 71.7 | 14.6 / 96.6 | 28.5 / 92.7 | 36.5 / 90.0 | 43.1 / 88.0 | 30.7 / 91.8 |
| PLOT | 72.6 | 15.9 / 96.6 | 33.7 / 92.8 | 38.2 / 91.0 | 39.2 / 90.2 | 31.8 / 92.7 |
| LoCoOp (MCM) | 71.5 | 23.1 / 95.5 | 32.7 / 93.4 | 39.9 / 90.6 | 40.2 / 91.3 | 34.0 / 92.7 |
| LoCoOp (GL-MCM) | 71.5 | 16.1 / 96.9 | 23.4 / 95.1 | 32.9 / 92.0 | 42.3 / 90.2 | 28.7 / 93.5 |
| PromptSRC | 73.2 | 20.6 / 95.7 | 30.1 / 93.7 | 38.0 / 91.1 | 46.0 / 89.0 | 33.7 / 92.4 |
| LSN+CoOp | 72.9 | 23.5 / 95.5 | 29.8 / 93.5 | 36.4 / 90.9 | 38.2 / 89.5 | 32.0 / 92.3 |
| LSN+CoCoOp | 72.9 | 21.6 / 95.8 | 26.3 / 94.4 | 34.5 / 91.3 | 38.5 / 90.4 | 30.2 / 93.0 |
| GalLoP | 75.1 | 13.7 / 97.1 | 24.9 / 94.0 | 32.5 / 91.3 | 38.4 / 90.4 | 27.3 / 93.2 |
Here are some experimental details to reproduce the few-shot experimental results. When training with very low shots, we use fewer training epochs to avoid overfitting, as in previous works.
Here are the number of epochs we used for each number of shots:
- 1 shot: 50 epochs,
- 2 and 4 shots: 100 epochs,
- 8 and 16 shots: 200 epochs. Note that we train with 50 epochs for Imagenet and 200 epochs for SUN397, regardless of the number of shots.
In addition, the batch sizes are different for each dataset (depending on the number of classes in the dataset). The batch sizes for each dataset are:
- Imagenet: 128,
- Caltech101: 32,
- FGVCAircraft: 32,
- StanfordCars: 32,
- Flowers102: 32,
- Food101: 32,
- SUN397: 128,
- OxfordPets: 32,
- DTD: 32,
- Eurosat: 10,
- UCF101: 32.
If you found our paper and/or code usefull please consider cite our work:
@inproceedings{lafon2024gallop,
title={Gallop: Learning global and local prompts for vision-language models},
author={Lafon, Marc and Ramzi, Elias and Rambour, Cl{\'e}ment and Audebert, Nicolas and Thome, Nicolas},
booktitle={European Conference on Computer Vision},
pages={264--282},
year={2024},
organization={Springer}
}