Baseline

A Closer Look at Few-shot Classification

Introduction

Name:	Baseline
Embed.:	Conv64F/ResNet12/ResNet18
Type:	Fine-tuning
Venue:	ICLR'19
Codes:	CloserLookFewShot

• When reproduceingthis method with the same setting in the original paper, you should skip validation during training-stage and choose the last model it saves.
• Notice that baseline use N-cls-head to train, where N > num_base_classes.

Cite this work with:

@inproceedings{DBLP:conf/iclr/ChenLKWH19,
  author    = {Wei{-}Yu Chen and
               Yen{-}Cheng Liu and
               Zsolt Kira and
               Yu{-}Chiang Frank Wang and
               Jia{-}Bin Huang},
  title     = {A Closer Look at Few-shot Classification},
  booktitle = {7th International Conference on Learning Representations, {ICLR} 2019,
               New Orleans, LA, USA, May 6-9, 2019},
  year      = {2019},
  url       = {https://openreview.net/forum?id=HkxLXnAcFQ}
}

---

Results and Models

Classification

	Embedding	📖 miniImageNet (5,1)	💻 miniImageNet (5,1)	📖miniImageNet (5,5)	💻 miniImageNet (5,5)	📝 Comments
1	Conv64F	42.11	42.34 ± 0.31 ⬇️ 📋	62.53	62.18 ± 0.30 ⬇️ 📋	Reproduce
2	Conv64F	-	44.90 ± 0.32 ⬇️ 📋	-	63.96 ± 0.30 ⬇️ 📋	Table.2
3	ResNet12	-	56.39 ± 0.36 ⬇️ 📋	-	76.18 ± 0.27 ⬇️ 📋	Table.2
4	ResNet18	-	54.11 ± 0.35 ⬇️ 📋	-	74.44 ± 0.29 ⬇️ 📋	Table.2
5	ResNet18	-	51.18 ± 0.34 ⬇️ 📋	-	74.06 ± 0.28 ⬇️ 📋	Reproduce

	Embedding	📖 tieredImageNet (5,1)	💻 tieredImageNet (5,1)	📖tieredImageNet (5,5)	💻 tieredImageNet (5,5)	📝 Comments
1	Conv64F	-	48.20 ± 0.35 ⬇️ 📋	-	68.96 ± 0.33 ⬇️ 📋	Table2
2	ResNet18	-	64.65 ± 0.41 ⬇️ 📋	-	82.73 ± 0.29 ⬇️ 📋	Table2