This repository hosts the code for the experiments in the CVPR 2021 paper How Well Do Self-Supervised Models Transfer?
This codebase has been tested with the following package versions:
python=3.6.8
torch=1.2.0
torchvision=0.4.0
PIL=7.1.2
numpy=1.18.1
scipy=1.2.1
pandas=1.0.3
tqdm=4.31.1
sklearn=0.22.2
In the paper we evaluate 14 pre-trained ResNet50 models, 13 self-supervised and 1 supervised. To download and prepare all models in the same format, run:
python download_and_prepare_models.py
This will prepare the models in the same format and save them in a directory named models.
Note 1: For SimCLR-v1 and SimCLR-v2, the TensorFlow checkpoints need to be downloaded manually (using the links in the table below) and converted into PyTorch format (using https://github.com/tonylins/simclr-converter and https://github.com/Separius/SimCLRv2-Pytorch, respectively).
Note 2: In order to convert BYOL, you may need to install some packages by running:
pip install jax jaxlib dill git+https://github.com/deepmind/dm-haiku
Below are links to the pre-trained weights used.
The datasets files (images, labels, splits etc.) can be downloaded from the following links:
CIFAR10/100, Aircraft, Caltech101, Cars, DTD, Flowers, Food, Pets, SUN397, VOC, CropDiseases, EuroSAT, ISIC2018, ChestX, NYUv2, ADE20K
To use our Caltech101 splits move both the train.txt and test.txt files found in caltech101-splits in to your Caltech101 data location (e.g. ../data/Caltech101/).
There are several classes defined in the datasets directory. The data is expected in a directory name data, located on the same level as this repository. Below is an outline of the expected file structure:
data/
CIFAR10/
DTD/
...
ssl-transfer/
datasets/
models/
readme.md
...
We provide the code for our linear evaluation in linear.py.
To evaluate DeepCluster-v2 on CIFAR10 given our pre-computed best regularisation hyperparameter, run:
python linear.py --dataset cifar10 --model deepcluster-v2 --C 0.316
The test accuracy should be close to 94.07%, the value reported in Table 1 of the paper.
To evaluate the Supervised baseline, run:
python linear.py --dataset cifar10 --model supervised --C 0.056
This model should achieve close to 91.47%.
To search for the best regularisation hyperparameter on the validation set, exclude the --C argument:
python linear.py --dataset cifar10 --model supervised
Finally, when using SimCLR-v1 or SimCLR-v2, always use the --no-norm argument:
python linear.py --dataset cifar10 --model simclr-v1 --no-norm
We provide code for finetuning in finetune.py.
To finetune DeepCluster-v2 on CIFAR10, run:
python finetune.py --dataset cifar10 --model deepcluster-v2
This model should achieve close to 97.06%, the value reported in Table 1 of the paper.
We provide the code for our few-shot evaluation in few_shot.py.
To evaluate DeepCluster-v2 on EuroSAT in a 5-way 5-shot setup, run:
python few_shot.py --dataset eurosat --model deepcluster-v2 --n-way 5 --n-support 5
The test accuracy should be close to 88.39% ± 0.49%, the value reported in Table 2 of the paper.
Or, to evaluate the Supervised baseline on ChestX in a 5-way 50-shot setup, run:
python few_shot.py --dataset chestx --model supervised --n-way 5 --n-support 50
This model should achieve close to 32.34% ± 0.45%.
We use the detectron2 framework to train our models on PASCAL VOC object detection.
Below is an outline of the expected file structure, including config files, converted models and the detectron2 framework:
detectron2/
tools/
train_net.py
...
...
ssl-transfer/
detectron2-configs/
finetune/
byol.yaml
...
frozen/
byol.yaml
...
models/
detectron2/
byol.pkl
...
...
...
To set it up, perform the following steps:
- Install detectron2 (requries PyTorch 1.5 or newer). We expect the installed framework to be located at the same level as this repository, see outline of expected file structure above.
- Convert the models into the format used by detectron2 by running
python convert_to_detectron2.py. The converted models will be saved in a directory calleddetectron2inside themodelsdirectory.
We include the config files for the frozen training in detectron2-configs/frozen and for full finetuning in detectron2-configs/finetune.
In order to train models, navigate into detectron2/tools/. We can now train e.g. BYOL with a frozen backbone on 1 GPU by running:
./train_net.py --num-gpus 1 --config-file ../../ssl-transfer/detectron2-configs/frozen/byol.yaml OUTPUT_DIR ./output/byol-frozen
This model should achieve close to 82.01 AP50, the value reported in Table 3 of the paper.
The code for running the surface normal estimation experiments is given in the surface-normal-estimation. We use the MIT CSAIL Semantic Segmentation Toolkit, but there is also a docker configuration file that can be used to build a container with all the dependencies installed. One can train a model with a command like:
./scripts/train_finetune_models.sh <pretrained-model-path> <checkpoint-directory>
and the resulting model can be evaluated with
./scripts/test_models.sh <checkpoint-directory>
We also use the same framework performing semantic segmentation. As per the surface normal estimation experiments, we include a docker configuration file to make getting dependencies easier. Before training a semantic segmentation model you will need to change the paths in the relevant YAML configuration file to point to where you have stored the pre-trained models and datasets. Once this is done the training script can be run with, e.g.,
python train.py --gpus 0,1 --cfg selfsupconfig/byol.yaml
where selfsupconfig/byol.yaml is the aforementioned configuration file. The resulting model can be evaluated with
python eval_multipro.py --gpus 0,1 --cfg selfsupconfig/byol.yaml
If you find our work useful for your research, please consider citing our paper:
@inproceedings{Ericsson2021HowTransfer,
title = {{How Well Do Self-Supervised Models Transfer?}},
year = {2021},
booktitle = {CVPR},
author = {Ericsson, Linus and Gouk, Henry and Hospedales, Timothy M.},
url = {http://arxiv.org/abs/2011.13377},
arxivId = {2011.13377}
}If you have any questions, feel welcome to create an issue or contact Linus Ericsson ([email protected]).