1% and 10% data selection

google-research / s4l

Tensorflow implementation of S4L: Self-Supervised Semi-Supervised Learning

https://arxiv.org/abs/1905.03670

Apache License 2.0

96 stars 21 forks source link

1% and 10% data selection #2

Open euwern opened 4 years ago

euwern commented 4 years ago

I have a question regarding to the imagenet2012 1% and 10% of data selected for training. In the script, it specified 13000 (1%) and 128000 (10%). Total trainval: 1281167.

Do you select the first 1% of images per class, as in first 13 images per class for 1% and first 128 images per class for 10%?

ghost commented 4 years ago

I have a question regarding to the imagenet2012 1% and 10% of data selected for training. In the script, it specified 13000 (1%) and 128000 (10%). Total trainval: 1281167.

Do you select the first 1% of images per class, as in first 13 images per class for 1% and first 128 images per class for 10%?

In short, Yes.

ghost commented 4 years ago

In the line with this question, I wonder how you sampled these 13000/128000 images from ILSVRC-2012. Are they chosen randomly? https://www.tensorflow.org/datasets/catalog/imagenet2012_subset How did you make these sets?