search

hytseng0509 / CrossDomainFewShot

Cross-Domain Few-Shot Classification via Learned Feature-Wise Transformation (ICLR 2020 spotlight)
329 stars 62 forks source link

train_baseline result and provided pre-trained model on baseline++ with ResNet10 #4

Open TianyuanYu opened 4 years ago

TianyuanYu commented 4 years ago

Hi, thanks for sharing your code. When I use the pertrained baseline++ with ResNet10 (downloaded according to Readme), I can always get a very good result combined with gnn (around 63% for 1 shot on MiniImageNet). But when I try training ResNet10 by myself using train_baseline.py, the result combined with gnnnet will drop 4-5% (I tried 3 times with different baseline++ and only to get 58%-59%). Do you have any idea about the possible reason?

hytseng0509 commented 4 years ago

Hmm interesting. Can you show the screenshot of the loss curve (baseline++ training) displayed on the tensorboard? The path of the saved tensorboard file should be indicated in this line.

TianyuanYu commented 4 years ago

baseline++1

Thanks for your reply. This is my loss curve for baseline++.

hytseng0509 commented 4 years ago

Sorry for the late reply. image

Seems like your loss is a bit higher than mine. What's the Pytorch version you use?

TianyuanYu commented 4 years ago

Thanks. I am using python3.6.7 and pytorch1.4.0