Question on the Amount of Labeled Data Used to Train the ResNet Feature Extractor

[cpphoo]

Hi, thanks for your excellent work!

After reading your paper, I was wondering whether the ResNet backbone was trained on the labelled portion of the base dataset (i.e. 10% of examples in each base class for tieredImageNet or 40% of examples in each base class in MiniImageNet as reported in Ren et. al.) or 100% of the images from each base class in the dataset.

[Yikai-Wang]

We use 90% of examples in each base class in all experiments for training set, and the residual 10% are for validation set.
Details can be found in

ICI-FSL/datasets.py

Lines 151 to 161 in f2537b7

	for c in range(len(keys)):
	num = 0
	num_train = int(len(self.data[keys[c]]) * 9 / 10)
	for file in self.data[keys[c]]:
	if type == 'train' and num <= num_train:
	self.Files.append(file)
	self.belong.append(c)
	elif type=='val' and num>num_train:
	self.Files.append(file)
	self.belong.append(c)
	num = num+1

[cpphoo]

Got it! Thanks for your prompt reply!