Understanding and accurately classifying fungi is crucial for ecological studies, food safety, and public health. In this paper, I present my approach to the FungiCLEF 2024 challenge, which aims to classify images of fungi, identify open-set “unknown” fungi species in the test data, and reduce the confusion between edible and poisonous fungi. This method leverages a combination of Metaformer-0, Metaformer-2, and CAFormer-S18 models, chosen for their strong classification performance relative to their computational eficiency. The models utilize metadata, while CAFormer-S18 does not, yet all belong to the same family of models known as Metaformers and employ convolutional blocks followed by multi-headed self-attention transformer blocks. My primary novel contribution is the application of OpenGAN to detect unknown fungi species, enhanced by incorporating WGAN-GP to improve training stability, resulting in a new open-set classifier training paradigm I term OpenWGAN-GP. This approach enables a lightweight discriminator to utilize the latent representations from the closed-set classifier for binary classification of open-set vs. closed-set species. My best-performing ensemble achieved public leaderboard scores of 0.2394 for Track 1, 0.1681 for Track 2, and 0.4075 for Track 3, along with a macro-averaged F1 score of 49.81%. Track 1 represents the classification loss with unknowns, Track 2 represents the edible-poisonous confusion loss (weighted heavily for poisonous to edible misclassifications), and the sum of Track 1 and Track 2. My method secured 1st place in the FungiCLEF 2024 competition for Track 1, F1, and Accuracy on the private leaderboard. Code is available at https://github.com/Jack-Etheredge/fungiclef2024.
Correctly identifying mushroom species and distinguishing between poisonous and edible varieties
are critical for public health. In 2023 alone China had 1,303 reported cases of mushroom poisonings
traced to 97 species of mushroom, of which 12 were newly discovered species [
CEUR
ceur-ws.org
Fine-grained classification involves classifying data that belongs to the same greater category.
Finegrained classification is challenging due to the large intra-set variation and low inter-set variation, in
contrast to standard coarse grained recognition. Fine-grained data, like open-set data, are ubiquitous in
the real world, and addressing the unique challenges posed by both has become the subject of more
intense academic interest recently. Metaformer [
Most works on open-set recognition primarily utilize coarse-grained classification datasets such as
CIFAR-10 [
FungiCLEF presents an opportunity to address this challenge by applying open-set recognition
techniques to fine-grained data, a task where many methods have shown promise [
In this work, I utilized OpenGAN [
The FungiCLEF 2024 challenge [
FungiCLEF represents two distinct challenges. One challenge is fine-grained and long-tailed
classification. Seesaw loss [
∑ ≠
+
.
is a balancing coeficient between diferent classes.
mitigation factor and a compensation factor :
= ⋅
is determined by a combination of a
mitigates the penalty on tail classes based on their instance ratio compared to head classes by
decreasing the penalty on class relative to the ratio of instance counts between the less abundant tail
class and the more abundant head class . Conversely, increases the penalty on class whenever
a misclassification occurs from class to class . This dual-factor approach in allows Seesaw loss
to dynamically adjust penalties based on both instance distribution and misclassification behavior,
optimizing the learning process in long-tailed multi-class classification tasks. The loss function is
explained in greater detail in the original paper [
The closed-set image classification models used to classify the images belong to a family of hybrid
convolutional and self-attention transformer models known as Metaformers [
The other challenge is the recognition of the open-set “unknown” class. OpenWGAN-GP was used to classify images as belonging to the closed-set or open-set datasets. The architecture of the open-set discriminator and the OpenWGAN-GP training methodology are described in greater detail in Section 3.6.
The public leaderboard for the competition reported multiple metrics for each submission. Track 1 was a classification loss that included unknowns.
×100 weight for poisonous → edible misclassification. Track 3 was the sum of the Track 1 and Track 2 losses. Additionally, the macro-averaged F1 score and accuracy were reported. The accuracy has been ignored for all experimental results reported here. Apart from the macro-averaged F1 score, none of the metrics correct for class imbalance. Thus the impact of classification performance on each class impacts final performance for Tracks 1-3 proportional to the number of observations for that class.
Track 1 loss is a standard classification error with an additional “unknown” class: (1) (2) (3) (4) (5) 1 = ∑ 1( , ( )),
1( , ()) = { 0 if () =
. for class predictions () for observations from a classifier and true labels . The cost function 1 is defined as
Track 2 loss penalizes the confusion of edible and poisonous species. Consider a function that indicates poisonous species as ( ) = 1
if species is poisonous, and ( ) = 0 denote the cost for poisonous → edible misclassification (a poisonous observation was predicted as otherwise. Let edible) and is defined as: the cost for edible → poisonous misclassfication. = 1 and for class predictions () for observations from a classifier and true labels as in 1. The cost function 2 is defined as 2( , ()) = 0 ⎧ ⎨
⎩ if ( ) = (()) if ( ) = 1 and (()) = 0 otherwise
Track 3 (the “user-focused loss”) is simply the sum of Track 1 and Track 2 losses:
weighted binary cross entropy loss.
A custom poison loss was used for all of the final models. The poison loss was formulated as a class
Given the set of poisonous classes and the set of edible classes , sum the probabilities of each class independently where is the softmax probability output for class .
3 = 1 + 2. = ∑ .
∈ = ∑ .
∈ (6) (7) (8) (9) (10) (11)
Let ∈ {0, 1} be the binary ground truth label for an image, where 1 indicates a poisonous class and 0 indicates an edible class. = 100 is the weight assigned to the edible class to penalize edible → poisonous misclassifications. Thus, the weighted cross-entropy loss is as follows: ,ℎ = −[ log( ) + (1 − ) log( )]
Since the softmax probabilities output by the model sum to 1, the probabilities for all the poisonous classes and all the edible classes were summed independently and used as the prediction for the binary cross entropy criterion. A weight of 100 was assigned to the edible class, since edible → poisonous misclassifications (true label is edible, predicted label is poisonous) was penalized ×100 in the Track 2 loss. The total training loss was the sum of the seesaw loss and the custom poison loss. This approach ensures that the training process emphasizes correctly classifying edible species as edible, thus reducing the risk of mistakenly classifying edible species as poisonous, which is heavily penalized in the evaluation metrics.
All experiments were performed on a machine with a single NVIDIA RTX 3090 graphics card and all models were trained using PyTorch [22]. Given that I was working on a single computer for this competition, eficient use of the limited compute available was of critical importance. Ensembling and test-time augmentations were used to increase performance while keeping training eficiency in check. An ensemble of computationally lean models have been shown to outperform a single larger model with respect to both training and inference cost [23]. Model architectures were chosen based on their performance on ImageNet and/or iNaturalist relative to the computational complexity of the models in TFLOPs, aiming for a final ensemble of at least two models. Test time augmentations allow for much greater performance without requiring any additional training of the models, making them a particularly attractive target for optimization when compute and time are both limited.
Metaformers [
I fine-tuned CAFormer-S18 with weights pretrained on ImageNet-21K [25] while Metaformer-0 and
Metaformer-2 models were pretrained on iNaturalist2021 [
OpenGAN proposes selection of the discriminator against a validation set of open- and closed-set examples, selecting the discriminator with the best validation ROC-AUC. However, since the ROC-AUC is calculated using a range of classification thresholds, the best classification threshold would also need to be determined for each OpenGAN discriminator. As such, the macro-averaged F1 was used as the selection metric instead of ROC-AUC for OpenWGAN-GP. Additionally, if the OpenWGAN-GP discriminator is selected based on macro-F1, an ensemble can be averaged without the need to calibrate the classification threshold.
Another diference from the original OpenGAN paper is that models were selected based on their macro-F1 performance rather than the proposed ROC-AUC. Since models would be used in an ensemble and the OpenWGAN-GP probabilities would be averaged, it was important to assume that the classification threshold for all of the individual OpenWGAN-GP models would be the same. An alternative strategy would have been voting, which would have allowed for diferent classification thresholds per model, but this was not explored in this study.
OpenGAN is a methodology for training a lightweight discriminator that utilizes the intermediate representation of an image to generate a binary classification for open-set recognition. Several related methods were proposed, but the one that performed best in their experiments and the one that I focus on in this work is OpenGAN fea with the inclusion of open-set training data, which I will simply refer to as OpenGAN. This paradigm allows for training a classification without initial consideration of the open set data. The discriminator is a multilayer perceptron consisting of fully connected layers with sizes → × 8 → × 4 → × 2 → → 1 , where represents the dimension of the intermediate representation from the closed-set classifier and is a hidden dimension multiplier. = 64 unless otherwise specified. The output layer uses a sigmoid activation function. Batch normalization [27] and LeakyRELU [28] are used between each dense layer. During training, the generator generates a feature vector of length from a 100-dimensional input vector with each value sampled independently from a standard normal distribution (mean 0, variance 1). The generator is also a multilayer perceptron with batch normalization and LeakyRELU. It has a similar architecture, but there are some critical diferences. The output dimension of the generator must match the input dimension closed-set images open-set images closed-set classifier generator intermediate
features fake open-set features
A) supervised closed (0) vs open (1)
discriminator real (1) vs fake (0)
B) adversarial of the discriminator, both of which correspond to the dimension of the intermediate representation of the closed-set classifier. As such, the generator is a multilayer perceptron with fully connected layers of sizes → × 8 → × 4 → × 2 → × 4 → , where = 100 . The input dimension is arbitrary, but this work utilizes a 100-dimensional input vector. As with the discriminator, = 64 unless otherwise specified. Additionally, the output activation used by the generator is Tanh instead of sigmoid activation. During training, the discriminator is trained to classify the closed and open set data in a supervised manner with binary labels using binary cross entropy. The generator is used to generate additional open set data and both the generator and discriminator are trained using the standard GAN training paradigm [29]. As such, the discriminator is updated twice per update of the generator since the discriminator is trained adversarially against the generator (real vs fake) as well as supervised (openvs closed-set). This training paradigm is illustrated in Figure 1. The Adam optimizer [30] was used with a learning rate of 1e-4 for the discriminator and 2e-4 for the generator. The higher learning rate is used for the generator to account for the 2:1 updates of the discriminator vs the generator.
Metaformer-0 and Metaformer-2 allow the fusion of metadata information with the vision
information. Metadata was utilized to provide the model with information concerning location, local growth
conditions, and temporal information by including the country code, substrate, and habitat, and
observation date (month and day). Example substrates include “fruits”, “wood”, “cones”, “soil”, and “peat
mosses”, while example habitats include “bog”, “dune”, “meadow”, and “roof”. There are 34, 32, and
31 categories for country code, substrate, and habitat, respectively. Metadata was preprocessed for
Metaformer-0 and Metaformer-2 according to [
I used the AdamW optimizer [31] with an initial learning rate of 1e-3 on only the classification output dense layer with the pretrained model frozen for the first 5 epochs, then reduced to 5e-5 for subsequent epochs. CAFormer-S18 models were trained with a batch size of 40, while Metaformer-0 models were trained with a batch size of 32, and Metaformer-2 models were trained with a batch size of 12. A weight decay of 0.05 was used for all of these models. The learning rate was reduced by a factor of 0.1 if the model did not improve the validation loss for 5 consecutive epochs. Early stopping was also employed when training all models to prevent wasting compute time on models which were no longer improving in their generalization to the validation set as measured by the validation loss.
LogitNorm [32] describes a technique that applies an L2 norm to the logits during training (the norm is not applied during inference). It was included with the hope that it should improve the separation of the classes in the embedding space relative to standard seesaw loss, which might enable OpenWGAN-GP to leverage the embedding space for more accurate open-set recognition. Additionally, LogitNorm was shown to act similarly to temperature scaling [33] to create models that generate less overconfident predictions. This would be important for maximum softmax probability or entropy thresholding, which were explored as alternatives to OpenWGAN-GP.
Training was performed with a square random crop, TrivialAugment [34], horizontal flip with 50% probability, and GridMask [35] with a probability of 20%, applied in that order.
At test-time, all images were resized with bicubic interpolation along the shortest dimension to 384 or 576 depending on the model followed by a square center crop of the same size. Test-time augmentations and ensembling were instrumental techniques for the final inference performance. Using a larger image size for inference than training was shown to improve accuracy for multiple datasets in FixRes [36].
The strategies employed were averaging horizontal flips, multi-instance averaging, ensemble averaging, and inference at a higher resolution (576x576) for CAFormer-S18 relative to training (384x384). The overall inference pipeline is summarized in Figure 2. Fivecrop was also investigated, but could not be incorporated in the allowed compute budget. Since it did not yield as significant an improvement relative to a larger ensemble with horizontal flipping according to local evaluation and public leaderboard scores for Track 3, the chosen configuration was preferred. It could be useful in the future to experiment with ensembling techniques that are more sophisticated than simple averaging, but none were attempted in this study.
Due to the open set “unknown” class being implicitly edible, the penalty for misclassifying poisonous mushrooms as unknown was greater than the decrease in misclassification loss. To mitigate this in the proposed solution, if the top prediction of the classification network was a poisonous mushroom, the prediction from the OpenGAN open set classifier was ignored for that observation.
My best-performing ensemble and OpenWGAN-GP combinations achieved 1st place on Track 1, F1, and Accuracy on the private leaderboard. All results reported are for the public leaderboard test set unless explicitly stated otherwise.
Diferent open set detection methods were evaluated against the public leaderboard and these results can be seen in Table 1. Experiments with a local validation set showed that temperature scaling improved the performance of maximum softmax probability (MSP) thresholding as well as softmax closed-set classifier penultimate
layer features
open-set discriminator horizontal flipping closed-set multiclass softmax probabilities open-vs-closed probabilities g n i g a r e v a e c n a t s n i i t l u m
x6 averaged closed-set probabilities x6 averaged open-vs-closed probabilities x6 final averaged closed-set probabilities x6 final averaged open-vs-closed probabilities x6 x6 ensemble averaging a) Inference pipeline b) Classification decision entropy thresholding (experiments not shown). OpenWGAN-GP consistently performed better on the leaderboard than softmax thresholding or entropy thresholding, even after temperature scaling [33] the probabilities. As can be seen in Table 1, the performance of either of these methods depends on optimizing a threshold. The optimal entropy threshold for local validation was 6, which did not appear to be optimal for the public leaderboard. This suggests that this method may not generalize well between test sets. OpenWGAN-GP is a binary classifier that is selected using the macro- F1 score with a classification threshold of 0.5, which means that no additional thresholding should be needed to generalize between test sets. Table 1 shows that despite not tuning the classification threshold, OpenWGAN-GP shows the best Track 1 and Track 3 performance while maintaining a similar Track 2 performance to MSP and entropy thresholding after avoiding poisonous → unknown misclassification as explained below.
It can be seen from the results in Table 1 that ignoring OpenWGAN-GP predictions for open-set recognition in the cases when the highest predicted probability belongs to a poisonous species (“ignore poison pred”) is critical to preventing the open-set recognition from degrading performance on Track 3. OpenWGAN-GP without ignore poison pred achieves a better Track 1 score than OpenWGAN-GP with ignore poison pred, but a much higher Track 2 score. This suggests that in many cases OpenWGAN-GP is correctly identifying unknowns that the classification network is predicting to be poisonous, but that the poisonous → edible cost for the poisonous closed → open misclassifications overwhelms the improvement in classification loss. This reinforces how challenging it is to simultaneously optimize classification performance, identification of unknown species, and avoidance of misclassifying poisonous species as edible.
Following [
Multiple architectures were evaluated for this study. Eficientnet-B0 [37] and Eficientnetv2-S [38] were both experimented with but their results were not as promising as Metaformer and CAFormer against a local validation set in early experiments. Results are not shown for these experiments since they are not directly comparable to the experiments reported. CAFormer-S18 has better performance than Metaformer-0 when the image resolution is increased at inference time despite belonging to the same family of models as Metaformer-0 and Metaformer-2, which leverage metadata information. CAFormers performed almost as well as Metaformer despite not utilizing information from the metadata. Future work could evaluate merging the two architectures into a CAFormer with a head for the metadata information. An ensemble of three CAFormer-S18 models that vary only by their training and validation data split performs nearly as well as ensembles of Metaformer-0, Metaformer-2, and CAFormer-S18. The best performing ensemble was Metaformer-0, Metaformer-2, CAFormer-S18 split C. Split C performs better than the other two CAFormer-S18 data splits, which provides a likely explanation as to why this ensemble outperformed Metaformer-0, Metaformer-2, and CAFormer-S18 split A.
OpenWGAN-GP training was optimized with respect to the hidden dimension size, ratio of closed and open set samples, and whether training augmentations were applied. Table 3 shows that the best Track 3 performance is achieved when using training augmentations and oversampling the open-set data to roughly the same number of samples as the closed-set data. This case is represented as weighted undersampling (w.u.) for closed-set sampling and “3x all” sampling for open-set sampling, which represents the case of oversampling the open-set dataset completely 3 times with training augmentations to increase the diversity of representations of the limited open-set data. These settings are used for all results shown in Table 2 and the OpenWGAN-GP results in Table 1. For sampling the closed-set data, weighted undersampling outperforms random undersampling and balanced undersampling. In cases where there is a >5% disparity between the number of samples in the open and closed sets, training is performed with balanced sampling between the open and closed sets. This pertains to all combinations except the 3x oversampling of the open set and weighted undersampling of the closed set. Local experimentation suggested that using the entire closed set dataset could yield a slight increase in Track 3 performance, but this dramatically increases the training time for the OpenWGAN-GP classifier (experiments not shown). While the Track 3 performance does not appear to be particularly sensitive to the hidden dimension size, the trend suggests that a smaller hidden dimension may have slightly improved performance, as shown in Table 4. More exhaustive combinations could not be performed due to competition submission limitations.
Several test-time augmentations were evaluated. The performance of each of these augmentations is shown in Table 5. Multi-instance averaging, averaging of horizontal flips of the same image, averaging of multiple crops of the same image, and averaging multiple image sizes are explored with Metaformer-0 trained on split D. Since Metaformer-0 does not support inference at a diferent resolution than the training resolution (in this case 384x384), the image size in Table 5 refers to the image resolution of the shorter dimension before a square crop of 384. For example, if the image size is 441, then the image is resized to 441 along the shorter dimension (assuming it is a rectangular image) and then a square center crop of 384 is taken. As such, image size must be at least 384 for Metaformer-0. Each augmentation improves performance individually and in combination. Of the test-time augmentations that were experimented with, multi-instance averaging has the greatest impact of any individual transformation.
Data split open WGAN -GP
i.p.p. Track 3 performance is best for multi-instance averaging in combination with horizontal flip averaging of the combinations shown in Table 5.
Despite improved performance on local validation, increasing the image resolution to 576 relative to the training resolution of 384 appears to have mixed results on the leaderboard performance as shown in Table 6. For data split A, the Track 1 score is improved with a higher resolution, but Track 2, Track 3, and F1 performance are better with an image resolution of 384. Image resolution 384 seems favored overall.
Table 7 shows that the removal of poison loss degrades the performance of the model across all metrics. Track 2 has a larger percent change than Track 1 or F1, which is sensible given that Track 2 corresponds to the edible → poisonous confusion loss.
Removal of LogitNorm from the training decreased performance for Track 1 and F1 as shown in Table 8. Presumably this is due to improving the separation of the classes in the latent space which is used by OpenWGAN-GP for unknown classification. Future work could explore whether LogitNorm also improves classification of fine-grained datasets in cases for which open-set recognition is not a consideration. Interestingly, removal of LogitNorm increases performance in Track 2 as shown in Table 8. The gain in performance in Track 2 from the removal of LogitNorm is great enough that Track 3 (the sum of Track 1 and Track 2 losses) is improved. This may suggest that LogitNorm is incompatible with the the poison loss used in this work. 4.5. OpenWGAN-GP The training stability and overall performance of OpenWGAN-GP is demonstrated relative to the original OpenGAN in Table 9. The identical performance of OpenGAN with and without these training optimizations suggests that the particular failure state observed is OpenGAN classifying none of the test set observations as open-set. The data sampling optimizations explored in Table 3 were not suficient to overcome the failure of OpenGAN to learn a meaningful representation of the data for classification. The switch from ROC-AUC to F1 as the discriminator selection metric against the validation set apparently also did not make a diference in light of the training failure. Since the same CAFormer-S18 classification model was used to generate the embeddings used by the OpenGAN variants shown above, it appears that the improved Track 1 (and consequently Track 3) performance is the result of the addition of the WGAN-GP training paradigm to OpenGAN.
Public leaderboard performance is shown in Table 10 for teams that have selected models for private leaderboard evaluation. Private leaderboard performance for the selected models for each team are shown in Table 11. The best performance for each metric is independently reported for each team, which means that results for each team may represent distinct solutions for each metric. My models achieved the best performance for Track 1 and accuracy in both the public and private leaderboards and the best F1 for the private leaderboard. My models placed 3rd on the private leaderboard for Track 2 and 2nd for Track 3, indicating that the poisonous → edible misclassification could be improved for the methods presented here.
jack-etheredge (ours) chirmy
IES TingTing1999 upupup DS@GT
The proposed methodology for open-set recognition of fungi species addresses the critical challenge of distinguishing between edible and poisonous mushrooms while efectively identifying unknown species. This study demonstrates the potential of combining Metaformer and CAFormer models to achieve robust classification performance. The integration of metadata in Metaformer models significantly enhances the model’s ability to leverage additional contextual information, thereby improving classification accuracy. However, one notable challenge is the current evaluation metrics, which assume unknown species are edible. This assumption may not be ideal if the models are intended to be used for foraging contexts, where new poisonous species of mushrooms are continually discovered. Re-evaluating these metrics to consider unknown species as potentially poisonous could mitigate the tension between open-set classification and the misclassification of poisonous species, thereby enhancing the practical applicability of the models in real-world scenarios. The current structure of the metrics which treats unknown species as edible puts open-set recognition and poisonous species identification in direct opposition with each other since misclassifying a closed-set poisonous species as unknown is heavily penalized, making their joint optimization challenging. If the intention is to build a system which displays both high detection rates for unknown species and high recall for poisonous mushrooms, the high penalty for poisonous → edible misclassification would work in favor of rather than against identification of unknown species if unknowns were assumed poisonous instead of edible. This may ultimately improve the model’s performance in both aspects.
During OpenWGAN-GP training of the open-set discriminator, the same label was used for real features and open-set features. All results shown utilizing OpenWGAN-GP in Section 4 represent cases when the open-set label and real label are shared during training of the open-set discriminator. By assigning the same label to real features and open-set features during the supervised and adversarial phases of each update respectively, the generator is incentivized to generate features that are indistinguishable from features created by the closed-set classifier for open-set observations. This is the opposite of the mapping used in the implementation of OpenGAN, which used the same closed-set label as the real label, which would have the efect of generating supplemental closed-set features instead. Initial experiments showed that a higher macro-F1 score was achieved between open- and closed-set validation examples for the FungiCLEF 2024 dataset when the real label was shared with the open-set label rather than the closed-set label. If the open-set label is the same as the real label, as in this work, the generator generates fake open-set features and the discriminator predicts less realistic features as closed-set features. Intuitively, the open-set should be more diverse and thus sharing the label with the inherently less realistic fake generated data seems the more logical choice in most scenarios when greater diversity is expected to be observed in the open-set data.
Future research should explore the redefinition of evaluation metrics to account for the possibility of unknown species being poisonous. This adjustment could reduce the conflict between optimizing for open-set classification and minimizing poisonous species misclassification. Additionally, investigating more sophisticated ensembling techniques and incorporating advanced data augmentation strategies could further improve model performance. Exploring the use of few-shot learning techniques might address the challenge posed by classes with very few observations. Finally, expanding the application of the proposed OpenWGAN-GP framework to other domains with similar classification challenges could validate its versatility and robustness.
This paper presents a novel method for open-set recognition of fungi species. The integration of WGAN-GP training optimizations into OpenGAN, resulting in OpenWGAN-GP, enhances training stability and enables lightweight discriminators to efectively identify unknown fungi species. An ensemble of Metaformer and CAFormer models is leveraged to classify fungi accurately while avoiding the misclassification of poisonous mushrooms as edible. The application of carefully chosen testtime augmentations, such as image resolution adjustments, horizontal flipping, and multi-instance averaging, dramatically improves classification performance. These techniques collectively contributed to achieving 1st place in the FungiCLEF 2024 competition for Track 1, F1, and Accuracy and 2nd place for the final ranking metric Track 3, which combines edible → poisonous confusion loss Track 2 with standardard misclassification loss including the unknown class Track 1.
The author would like to thank Jillian Etheredge for constructive criticism of the manuscript. [22] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: An Imperative Style, High-Performance Deep Learning Library, in: Advances in Neural Information Processing Systems, volume 32, Curran Associates, Inc., 2019. URL: https://papers.nips.cc/paper_files/paper/2019/hash/ bdbca288fee7f92f2bfa9f7012727740-Abstract.html. [23] X. Wang, D. Kondratyuk, E. Christiansen, K. M. Kitani, Y. Alon, E. Eban, Wisdom of Committees: An Overlooked Approach To Faster and More Accurate Models, 2022. URL: http://arxiv.org/abs/ 2012.01988. doi:10.48550/arXiv.2012.01988, arXiv:2012.01988 [cs]. [24] W. Yu, C. Si, P. Zhou, M. Luo, Y. Zhou, J. Feng, S. Yan, X. Wang, MetaFormer Baselines for Vision, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (2024) 896–912. URL: http://arxiv.org/abs/2210.13452. doi:10.1109/TPAMI.2023.3329173, arXiv:2210.13452 [cs]. [25] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, ImageNet: A large-scale hierarchical image database, in: 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255. URL: https://ieeexplore.ieee.org/document/5206848. doi:10.1109/CVPR.2009.5206848, iSSN: 1063-6919. [26] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, A. Courville, Improved Training of Wasserstein GANs, 2017. URL: http://arxiv.org/abs/1704.00028. doi:10.48550/arXiv.1704.00028, arXiv:1704.00028 [cs, stat]. [27] S. Iofe, C. Szegedy, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, 2015. URL: http://arxiv.org/abs/1502.03167. doi:10.48550/arXiv.1502. 03167, arXiv:1502.03167 [cs]. [28] B. Xu, N. Wang, T. Chen, M. Li, Empirical Evaluation of Rectified Activations in Convolutional Network, 2015. URL: http://arxiv.org/abs/1505.00853. doi:10.48550/arXiv.1505.00853, arXiv:1505.00853 [cs, stat]. [29] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative Adversarial Networks, 2014. URL: http://arxiv.org/abs/1406.2661. doi:10.48550/arXiv. 1406.2661, arXiv:1406.2661 [cs, stat]. [30] D. P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization, CoRR abs/1412.6980 (2014).
URL: https://api.semanticscholar.org/CorpusID:6628106. [31] I. Loshchilov, F. Hutter, Decoupled Weight Decay Regularization, 2019. URL: http://arxiv.org/abs/ 1711.05101. doi:10.48550/arXiv.1711.05101, arXiv:1711.05101 [cs, math]. [32] H. Wei, R. Xie, H. Cheng, L. Feng, B. An, Y. Li, Mitigating Neural Network Overconfidence with Logit Normalization, 2022. URL: http://arxiv.org/abs/2205.09310. doi:10.48550/arXiv.2205. 09310, arXiv:2205.09310 [cs]. [33] C. Guo, G. Pleiss, Y. Sun, K. Q. Weinberger, On Calibration of Modern Neural Networks, 2017. URL: http://arxiv.org/abs/1706.04599, arXiv:1706.04599 [cs]. [34] S. G. Müller, F. Hutter, TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentation, 2021.
URL: http://arxiv.org/abs/2103.10158. doi:10.48550/arXiv.2103.10158, arXiv:2103.10158 [cs]. [35] P. Chen, S. Liu, H. Zhao, X. Wang, J. Jia, GridMask Data Augmentation, 2024. URL: http://arxiv.
org/abs/2001.04086. doi:10.48550/arXiv.2001.04086, arXiv:2001.04086 [cs]. [36] H. Touvron, A. Vedaldi, M. Douze, H. Jégou, Fixing the train-test resolution discrepancy, 2022.
URL: http://arxiv.org/abs/1906.06423. doi:10.48550/arXiv.1906.06423, arXiv:1906.06423 [cs]. [37] M. Tan, Q. V. Le, EficientNet: Rethinking Model Scaling for Convolutional Neural Networks,
ArXiv (2019). URL: https://api.semanticscholar.org/CorpusID:167217261. [38] M. Tan, Q. V. Le, EficientNetV2: Smaller Models and Faster Training, in: International Conference on Machine Learning, 2021. URL: https://api.semanticscholar.org/CorpusID:232478903.