GBML

A collection of Gradient-Based Meta-Learning Algorithms with pytorch

• MAML

python3 main.py --alg=MAML

• Reptile

python3 main.py --alg=Reptile

• CAVIA

python3 main.py --alg=CAVIA

Results on miniImagenet

• Without pre-trained encoder (Use 64 channels by default. The exceptions are in parentheses)

	5way 1shot	5way 1shot (ours)	5way 5shot	5way 5shot (ours)
MAML	48.70 ± 1.84%	49.00 %	63.11 ± 0.92%	65.18 %
Reptile	47.07 ± 0.26%	43.40 %	62.74 ± 0.37%	-
CAVIA	49.84 ± 0.68% (128)	50.07 % (64)	64.63 ± 0.53% (128)	64.21 % (64)
iMAML	49.30 ± 1.88%	-	-	-
Meta-Curvature	55.73 ± 0.94% (128)	-	70.30 ± 0.72% (128)	-

• With pre-trained encoder (To be implemented.)

	5way 1shot	5way 1shot (ours)	5way 5shot	5way 5shot (ours)
Meta-SGD	56.58 ± 0.21%	-	68.84 ± 0.19%	-
LEO	61.76 ± 0.08%	-	77.59 ± 0.12%	-
Meta-Curvature	61.85 ± 0.10%	-	77.02 ± 0.11%	-

Dependencies

• Python >= 3.6
• Pytorch >= 1.2
• Higher
• Torchmeta

To do

• Add ResNet and Pre-trained encoder
• Add iMAML, Meta-Curvature