Pull Request

Author: avinash-ravichandran

README.md

@@ -1 +1,49 @@
-# MetaOptNet
+# Meta-learning with differentiable convex optimization
+The following are the miniImageNet few-shot accuracies of ProtoNet (Snell et al., NIPS 2017), Meta-learning with differentiable closed-form solvers — R2D2 (Bertinetto et al., arXiv 2018), and MetaOptNet (ours) ran on our codebase.
+
+miniImageNet 5-way accuracies
+                                                                             
+| Method                                     | 1-shot            | 5-shot          |
+|--------------------------------------------| ----------------- | --------------- |
+| ProtoNet (reported in paper)               | 49.4 ± 0.8 %      | 68.2 ± 0.7 %    |
+| ProtoNet (our implementation)              | 52.0 ± 0.6 %      | 68.5 ± 0.5 %    |
+| R2D2 (reported in paper)                   | 51.2 ± 0.6 %      | 68.2 ± 0.6 %    |
+| R2D2 (our implementation)                  | 54.0 ± 0.6 %      | 70.0 ± 0.5 %    |
+| MetaOptNet (ours)                          | 57.7 ± 0.6 %      | 73.6 ± 0.5 %    |
+
+For ProtoNet and R2D2, we used the network architectures used in the original publication. For MetaOptNet, we used ResNet12.
+As a note, we use a different training episode composition from the ProtoNet and R2D2 papers: we always trained with 5-way and each class contains 6 query points per episode.
+In order to replicate the results, execute the following commands:
+
+
+Prototypical Networks 1-shot
+```
+python train.py --gpu 0 --save-path "./experiments/exp_proto_1_shot" --train-way 5 --shot 1 --query 6 --head ProtoNet --network ProtoNet --episodes-per-batch 8
+python test.py --gpu 0 --load ./experiments/exp_proto_1_shot/best_model.pth --episode 1000 --way 5 --shot 1 --query 15 --head ProtoNet --network ProtoNet
+```
+Prototypical Networks 5-shot
+```
+python train.py --gpu 0 --save-path "./experiments/exp_proto_5_shot" --train-way 5 --shot 5 --query 6 --head ProtoNet --network ProtoNet --episodes-per-batch 8
+python test.py --gpu 0 --load ./experiments/exp_proto_5_shot/best_model.pth --episode 1000 --way 5 --shot 5 --query 15 --head ProtoNet --network ProtoNet
+```
+R2D2 1-shot
+```
+python train.py --gpu 0 --save-path "./experiments/exp_R2D2_1_shot" --train-way 5 --shot 1 --query 6 --head R2D2 --network R2D2 --episodes-per-batch 8
+python test.py --gpu 0 --load ./experiments/exp_R2D2_1_shot/best_model.pth --episode 1000 --way 5 --shot 1 --query 15 --head R2D2 --network R2D2
+```
+
+R2D2 5-shot
+```
+python train.py --gpu 0 --save-path "./experiments/exp_R2D2_5_shot" --train-way 5 --shot 5 --query 6 --head R2D2 --network R2D2 --episodes-per-batch 8
+python test.py --gpu 0 --load ./experiments/exp_R2D2_5_shot/best_model.pth --episode 1000 --way 5 --shot 5 --query 15 --head R2D2 --network R2D2
+```
+MetaOptNet 1-shot
+```
+python train.py --gpu 0 --save-path "./experiments/exp_MetaOptNet_1_shot" --train-way 20 --shot 1 --query 6 --head SVM --network ResNet --episodes-per-batch 2
+python test.py --gpu 0 --load ./experiments/exp_MetaOptNet_1_shot/best_model.pth --episode 1000 --way 5 --shot 1 --query 15 --head SVM --network ResNet
+```
+MetaOptNet 5-shot
+```
+python train.py --gpu 0 --save-path "./experiments/exp_MetaOptNet_5_shot" --train-way 15 --shot 5 --query 6 --head SVM --network ResNet --episodes-per-batch 1
+python test.py --gpu 0 --load ./experiments/exp_MetaOptNet_5_shot/best_model.pth --episode 1000 --way 5 --shot 5 --query 15 --head SVM --network ResNet
+```

__init__.py

@@ -0,0 +1 @@
+# Implement your code here.