DataCV Challenge @ CVPR 2023

Welcome to DataCV Challenge 2023!

This is the development kit repository for the 1st DataCV Challenge. This repository includes details on how to download datasets, run baseline models, and organize your result as answer.zip. The final evaluation will occur on the CodeLab evaluation server, where all competition information, rules, and dates can be found.

Overview

Label-free model evaluation is the competition task. It is different from the standard evaluation that calculates model accuracy based on model outputs and corresponding test labels. Label-free model evaluation (AutoEval), on the other hand, has no access to test labels. In this competition, participants need to design a method that can estimate the model accuracies on test sets without ground truths.

DataCV Challenge @ CVPR 2023

Competition Submission

In total, the test set comprises 100 datasets. As a result, each model's accuracy should be predicted 100 times. Given that there are two models to be evaluated, the expected number of lines in the "answer.txt" file is 200.

The first 100 lines represent the accuracy predictions of the ResNet model, while the second 100 lines represent those of the RepVGG model. Each of the 100-line predictions is the accuracy prediction for the model on xxx.npy dataset, where xxx goes from 001 to 100.

To prepare your submission, you need to write your predicted accuracies into a plain text file named "answer.txt", with one prediction (e.g., 0.876543) per line. For example,

Then, zip the text file and it submit to the competition website.

How to organize an answer.txt file for validation evaluation?

Please refer to Organize Results for Submission.

In the competition, you are only required to submit the zipped prediction results named the "answer.txt". We give an example for this txt file at answer.txt demo.

Challenge Data

The training dataset consists of 1,000 transformed datasets from the original CIFAR-10 test set, using the transformation strategy proposed by Deng et al. (2021). The validation set was composed of CIFAR-10.1, CIFAR-10.1-C (add corruptions (Hendrycks et al., 2019) to CIFAR-10.1 dataset), and CIFAR-10-F (real-world images collected from Flickr)

The CIFAR-10.1 dataset is a single dataset. In contrast, CIFAR-10.1-C and CIFAR-10-F contain 19 and 20 datasets, respectively. Therefore, the total number of datasets in the validation set is 40.

The training datasets share a common label file named labels.npy, and images files are named new_data_xxx.npy, where xxx is a number from 000 to 999. For every dataset in the validation set, the image file and their labels are stored as two separate Numpy array files named "data.npy" and "labels.npy". The PyTorch implementation of the Dataset class for loading the data can be found in utils.py.

Download the training datasets: link

Download the validation datasets: link

Download the training datasets' accuracies on the ResNet-56 model: link

Download the training datasets' accuracies on the RepVGG-A0 model: link

NOTE: To access the test datasets and participate in the competition, please fill in the Datasets Request Form and send the signed form to the competition organiser. Failing to provide the form will lead to the revocation of the CodaLab account in the competition.

Classifiers being Evaluated

In this competition, the classifiers being evaluated are ResNet-56 and RepVGG-A0. Both implementations can be accessed in the public repository at https://github.com/chenyaofo/pytorch-cifar-models. To utilize the models and load their pretrained weights, use the code provided.

import torch

model = torch.hub.load("chenyaofo/pytorch-cifar-models", "cifar10_resnet56"， pretrained=True)
model = torch.hub.load("chenyaofo/pytorch-cifar-models", "cifar10_repvgg_a0", pretrained=True)

Organize Results for Submission

As we use automated evaluation scripts for submissions, the format of submitted files is very important. So, there is a function to store the accuracy predictions into the required format, named store_ans in code/utils.py and results_format/get_answer_txt.py.

Read the results_format/get_answer_txt.py file to comprehend the function usage. Execute the code below to see the results.

python3 results_format/get_answer_txt.py

Several Baselines

The necessary dependencies are specified in the requirements.txt file and the experiments were conducted using Python version 3.10.8, with a single GeForce RTX 2080 Ti GPU.

The table presented below displays the results of the foundational measurements using root-mean-square error (RMSE). Accuracies are converted into percentages prior to calculation.

Baseline Results

ResNet-56

Method	CIFAR-10.1	CIFAR-10.1-C	CIFAR-10-F	Overall
Rotation	7.285	6.386	7.763	7.129
ConfScore	2.190	9.743	2.676	6.985
Entropy	2.424	10.300	2.913	7.402
ATC	11.428	5.964	8.960	7.766
FID	7.517	5.145	4.662	4.985

RepVGG-A0

Method	CIFAR-10.1	CIFAR-10.1-C	CIFAR-10-F	Overall
Rotation	16.726	17.137	8.105	13.391
ConfScore	5.470	12.004	3.709	8.722
Entropy	5.997	12.645	3.419	9.093
ATC	15.168	8.050	7.694	8.132
FID	10.718	6.318	5.245	5.966

Code-Execution

To install required Python libraries, execute the code below.

pip3 install -r requirements.txt

The above results can be replicated by executing the code provided below in the terminal.

cd code/
chmod u+x run_baselines.sh && ./run_baselines.sh

To run one specific baseline, use the code below.

cd code/
python3 get_accuracy.py --model <resnet/repvgg> --dataset_path DATASET_PATH
python3 baselines/BASELINE.py --model <resnet/repvgg> --dataset_path DATASET_PATH

Baseline Description

The following succinctly outlines the methodology of each method, as detailed in the appendix of "Predicting Out-of-Distribution Error with the Projection Norm" paper (Yu et al., 2022).

Rotation. The Rotation Prediction (Rotation) (Deng et al., 2021) metric is defined as

$$\text{Rotation} = \frac{1}{m}\sum_{j=1}^m\Big\{\frac{1}{4}\sum_{r \in \{0^\circ, 90^\circ, 180^\circ, 270^\circ\}}\mathbf{1}\{C^r(\tilde{x}_j; \hat{\theta}) \neq y_r\}\Big\},$$

where $y_r$ is the label for $r \in \lbrace 0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ} \rbrace$, and $C^r(\tilde{x}_j; \hat{\theta})$ predicts the rotation degree of an image $\tilde{x}_j$.

ConfScore. The Averaged Confidence (ConfScore) (Hendrycks & Gimpel, 2016) is defined as $$\text{ConfScore} = \frac{1}{m}\sum_{j=1}^m \max_{k} \text{Softmax}(f(\tilde{x}_j; \hat{\theta}))_k,$$ where $\text{Softmax}(\cdot)$ is the softmax function.

Entropy. The Entropy (Guillory et al., 2021) metric is defined as

$$\text{Entropy} = \frac{1}{m}\sum_{j=1}^m \text{Ent}(\text{Softmax}(f(\tilde{x}_j; \hat{\theta}))),$$ where $\text{Ent}(p)=-\sum^K_{k=1}p_k\cdot\log(p_k)$.$$

ATC. The Averaged Threshold Confidence (ATC) (Garg et al., 2022) is defined as

$$\text{ATC} = \frac{1}{m}\sum_{j=1}^m\mathbf{1}\{s(\text{Softmax}(f(\tilde{x}_j; \hat{\theta}))) < t\},$$ where $s(p)=\sum^K_{j=1}p_k\log(p_k)$, and $t$ is defined as the solution to the following equation, $$\frac{1}{m^{\text{val}}} \sum_{\ell=1}^{m^\text{val}}\mathbf{1}\{s(\text{Softmax}(f(x_{\ell}^{\text{val}}; \hat{\theta}))) < t\} = \frac{1}{m^{\text{val}}}\sum_{\ell=1}^{m^\text{val}}\mathbf{1}\{C(x_\ell^{\text{val}}; \hat{\theta}) \neq y_\ell^{\text{val}}\}$$

where $(x_\ell^{\text{val}}, y_\ell^{\text{val}}), \ell=1,\dots, m^{\text{val}}$, are in-distribution validation samples.

FID. The Frechet Distance (FD) between datasets (Deng et al., 2020) is defined as

$$\text{FD}(\mathcal{D}_{ori}, \mathcal{D}) = \lvert \lvert \mu_{ori} - \mu \rvert \rvert_2^2 + Tr(\Sigma_{ori} + \Sigma - 2(\Sigma_{ori}\Sigma)^\frac{1}{2}),$$

where $\mu_{ori}$ and $\mu$ are the mean feature vectors of $\mathcal{D}_{ori}$ and $\mathcal{D}$, respectively. $\Sigma_{ori}$ and $\Sigma$ are the covariance matrices of $\mathcal{D}_{ori}$ and $\mathcal{D}$, respectively. They are calculated from the image features in $\mathcal{D}_{ori}$ and $\mathcal{D}$, which are extracted using the classifier $f_{\theta}$ trained on $\mathcal{D}_{ori}$.

The Frechet Distance calculation functions utilized in this analysis were sourced from a publicly available repository by Weijian Deng.

xingjianleng/autoeval_baselines