Fixing the train-test resolution discrepancy: FixEfficientNet

18 Mar 2020  ·  Hugo Touvron, Andrea Vedaldi, Matthijs Douze, Hervé Jégou ·

This paper provides an extensive analysis of the performance of the EfficientNet image classifiers with several recent training procedures, in particular one that corrects the discrepancy between train and test images. The resulting network, called FixEfficientNet, significantly outperforms the initial architecture with the same number of parameters. For instance, our FixEfficientNet-B0 trained without additional training data achieves 79.3% top-1 accuracy on ImageNet with 5.3M parameters. This is a +0.5% absolute improvement over the Noisy student EfficientNet-B0 trained with 300M unlabeled images. An EfficientNet-L2 pre-trained with weak supervision on 300M unlabeled images and further optimized with FixRes achieves 88.5% top-1 accuracy (top-5: 98.7%), which establishes the new state of the art for ImageNet with a single crop. These improvements are thoroughly evaluated with cleaner protocols than the one usually employed for Imagenet, and particular we show that our improvement remains in the experimental setting of ImageNet-v2, that is less prone to overfitting, and with ImageNet Real Labels. In both cases we also establish the new state of the art.

PDF Abstract

Datasets


Results from the Paper


Ranked #9 on Image Classification on ImageNet ReaL (using extra training data)

     Get a GitHub badge
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Image Classification ImageNet FixEfficientNet-B6 Top 1 Accuracy 86.7% # 134
Number of params 43M # 739
Image Classification ImageNet FixEfficientNet-B8 Top 1 Accuracy 85.7% # 212
Image Classification ImageNet FixEfficientNet-B4 Top 1 Accuracy 85.9% # 193
Number of params 19M # 563
Image Classification ImageNet FixEfficientNet-L2 Top 1 Accuracy 88.5% # 49
Number of params 480M # 1008
Hardware Burden None # 1
Operations per network pass None # 1
GFLOPs 585 # 538
Image Classification ImageNet FixEfficientNet-B0 Top 1 Accuracy 80.2% # 716
Number of params 5.3M # 452
GFLOPs 1.60 # 135
Image Classification ImageNet FixEfficientNet-B1 Top 1 Accuracy 82.6% # 522
Number of params 7.8M # 496
Image Classification ImageNet FixEfficientNet-B2 Top 1 Accuracy 83.6% # 413
Number of params 9.2M # 505
Image Classification ImageNet FixEfficientNet-B3 Top 1 Accuracy 85% # 271
Number of params 12M # 531
Image Classification ImageNet FixEfficientNet-B5 Top 1 Accuracy 86.4% # 152
Number of params 30M # 686
Image Classification ImageNet FixEfficientNetB4 Top 1 Accuracy 84.0% # 368
Number of params 19M # 563
Image Classification ImageNet FixEfficientNet-B7 Top 1 Accuracy 87.1% # 109
Number of params 66M # 831
GFLOPs 82 # 491
Image Classification ImageNet ReaL FixEfficientNet-L2 Accuracy 90.9% # 9
Params 480M # 50
Image Classification ImageNet ReaL FixEfficientNet-B8 Accuracy 90.0% # 21
Params 87M # 46

Methods