search

wyharveychen / CloserLookFewShot

source code to ICLR'19, 'A Closer Look at Few-shot Classification'
Other
1.13k stars 268 forks source link

Baseline and Baseline++ #way#shot method #38

Open BoostMom opened 5 years ago

BoostMom commented 5 years ago

Hi : I am wondering if I want to train baseline++ with few shot training method, I just change the dataloader class ,right? Because now in train.py, the baseline and baseline++ are not trained in few shot method.

wyharveychen commented 5 years ago

Hello, n-way and k-shot specify the parameter in meta-training methods in the meta-training stage. For Baseline and Baseline++, it always trains a classifier for all class with all data in training stage.

BoostMom commented 5 years ago

yes，I saw the content in code。However, there were five shot and one shot of baseline and baesline++ comparison in experiment result. How that happened

ycliu93 commented 5 years ago

yes，I saw the content in code。However, there were five shot and one shot of baseline and baesline++ comparison in experiment result. How that happened

For novel class which contains few samples per class, we do train the classifier of both baseline and baseline++ in the "few-shot method" (i.e., using the few samples from the novel class).

As both meta-learning method and our baseline methods use the same amount of training samples (abundant samples from the base class and few samples from the novel class), we thus present the comparison in our paper.

Please have a look at our paper.