/diht

Filtering, Distillation, and Hard Negatives for Vision-Language Pre-Training

Primary LanguagePythonOtherNOASSERTION

DiHT

"Filtering, Distillation, and Hard Negatives for Vision-Language Pre-Training" [Paper]

Installation

git clone https://github.com/facebookresearch/diht
cd diht
pip install -r requirements.txt
pip install -e .  # To install as an editable package

You can use pip install . to install the codebase as a non-editable package.

Example

import torch

from diht import model_zoo
from PIL import Image


text_tokenizer, image_transform, model = model_zoo.load_model(
    "diht_vitl14_336px", is_train=False
)

image = Image.open("infer_image.png").convert("RGB")
image = image_transform(image).unsqueeze(0)
text_captions = ["a mountain", "a beach", "a desert"]
text = text_tokenizer(text_captions)

with torch.no_grad():
    image_features, text_features, logit_scale = model(image, text)
    logits_per_image = logit_scale * image_features @ text_features.T
    probs = logits_per_image.softmax(dim=-1).numpy()

print(f"text captions: {text_captions}")
print(f"text caption probs: {probs}")

The above code snippet should output

text captions: ['a mountain', 'a beach', 'a desert']
text caption probs: [[0.99370664 0.00514017 0.00115326]]

Running on GPU/CPU

By default the model runs on CPU, to run on GPU you can do model = model.to(torch.device("cuda")). The image and text tensors will also have to be transferred accordingly.

Available Models

import diht


print(diht.available_models())

Example ImageNet-1K zero-shot evaluation

A simple image classification zero-shot evaluation using a single GPU can be performed by running:

Note: Download ImageNet-1K dataset from the original website. Edit IMAGENET_ROOT in example_imagenet_eval.py to match the location on your machine.

python example_imagenet_eval.py

For DiHT-L/14@336 the output should look like:

ImageNet1K acc@1 for diht_vitl14_336px: 77.9

Example retrieval zero-shot evaluation

A simple retrieval zero-shot evaluation using a single GPU can be performed by running:

Note: Download COCO and Flickr30K datasets from the original websites. Json files (coco_test.json and flickr30k_test.json) can be downloaded from https://github.com/salesforce/ALBEF#download. Edit COCO_ROOT and FLICKR30K_ROOT in example_retrieval_eval.py to match the locations on your machine.

python example_retrieval_eval.py

For DiHT-L/14@336 the output should look like:

COCO T2I r@1 for diht_vitl14_336px: 49.3
COCO I2T r@1 for diht_vitl14_336px: 65.3

Flickr30K T2I r@1 for diht_vitl14_336px: 78.2
Flickr30K I2T r@1 for diht_vitl14_336px: 91.1

Zero-shot model performance

Model ImageNet-1K COCO T2I COCO I2T Flickr30K T2I Flickr30K I2T
Accuracy@1 Recall@1 Recall@1 Recall@1 Recall@1
diht_vitb32_224px 68.0 40.6 59.3 68.6 84.4
diht_vitb16_224px 72.2 43.3 60.3 72.9 89.8
diht_vitl14_224px 77.0 48.0 65.1 76.7 92.0
diht_vitl14_336px 77.9 49.3 65.3 78.2 91.1

Citation

If you find this model useful, please consider citing our preprint using the citation below.

@article{rdk+23,
  title = {Filtering, Distillation, and Hard Negatives for Vision-Language Pre-Training},
  author = {Radenovic, Filip and Dubey, Abhimanyu and Kadian, Abhishek and Mihaylov, Todor and Vandenhende, Simon and Patel, Yash and Wen, Yi and Ramanathan, Vignesh and Mahajan, Dhruv},
  journal = {arXiv:2301.02280},
  year = {2023}
}

License

Copyright (c) Meta Platforms, Inc. and affiliates.
All rights reserved.
This source code is licensed under the license found in the
LICENSE file in the root directory of this source tree.