Review — Fixing the train-test resolution discrepancy: FixEfficientNet

Apply FixRes onto EfficientNet for Additional Results

FixEfficientNet (orange curve) surpasses all EfficientNet models, including the models trained with Noisy student (red curve) and adversarial examples (blue curve). The sws models are from [2].

Fixing the train-test resolution discrepancy: FixEfficientNet
FixEfficientNet, by Facebook AI Research
2020 arXiv v5, Over 200 Citations. (Sik-Ho Tsang @ Medium)

• There is lack of results for EfficientNet using FixRes in FixRes. FixRes is applied onto EfficientNet for additional results, and better performance is obtained.

Outline

1. FixEfficientNet
2. Experimental Results

1. FixEfficientNet

• FixRes is a simple but efficient fine-tuning strategy.
• First, EfficientNet is trained using a smaller input image size (train res).
• Then, EfficientNet is re-trained or a few top layers at the target resolution (test res).
• The only difference is that FixRes data augmentation is combined with label smoothing (in Inception-v3) during the fine-tuning.
• (Please feel free to read FixRes for more details if interested.)

2. Experimental Results

2.1. ImageNet

Results on ImageNet without external data (single Crop evaluation)

Results on ImageNet with extra training data

FixEfficientNet-L2 surpasses all other results reported in the literature.

• It achieves 88.5% Top-1 accuracy and 98.7% Top-5 accuracy on the ImageNet-2012 validation benchmark.

2.2. ImageNet-Real

Results on ImageNet Real labels

• There are some incorrect labels in ImageNet, ImageNet clean labels are labels cleaned by Beyer et all. [5].

With 90.9% Top-1 accuracy and 98.8% Top-5 accuracy, FixEfficientNet-L2 surpasses all other results reported in the literature with this labels.

2.3. ImageNet-V2

Results on ImageNet-V2 [17] Matched Frequency with extra-training data.

Results on ImageNet-V2 [17] Matched Frequency without external data (single Crop evaluation).

• ImageNet-V2 [17] dataset was introduced to overcome the lack of a test split in the Imagenet dataset. ImageNet-V2 consists of 3 novel test sets that replace the ImageNet test set, which is no longer available.

Performance comparison and state of the art on ImageNet-v2, single crop with external data, sorted by top-1 accuracy. NS: Noisy Student [8]. BS: Billion-scale [2].

FixEfficientNet-L2 that fine-tuned from EfficientNet establishes the new state of the art with additional data on this benchmark.

Hope I can review Noisy Student [8], and Billion-scale [2] in the coming future.

Reference

[2020 arXiv] [FixEfficientNet]
Fixing the train-test resolution discrepancy: FixEfficientNet

Image Classification

1989–2018 … 2019: [ResNet-38] [AmoebaNet] [ESPNetv2] [MnasNet] [Single-Path NAS] [DARTS] [ProxylessNAS] [MobileNetV3] [FBNet] [ShakeDrop] [CutMix] [MixConv] [EfficientNet] [ABN] [SKNet] [CB Loss] [AutoAugment, AA] [BagNet] [Stylized-ImageNet] [FixRes] [Ramachandran’s NeurIPS’19]
2020: [Random Erasing (RE)] [SAOL] [AdderNet] [FixEfficientNet]
2021: [Learned Resizer]

My Other Previous Paper Readings