anikristo / decaf Goto Github PK

Check datasets/README.md for details on the directory structure etc.

Check datasets/README.md for details of the directory structure etc.

1. Caltech-101 dataset (Object recognition)
2. Office dataset (Domain Adaptation)
3. Caltech-UCSD birds dataset (Subcategory recognition)
4. SUN-397 (Scene Recognition)
5. ILSVRC-2012 (Initial training) -> We need this only in case we can't get the model to work directly in TF

Only yield 30 image samples per class

Use inherited classes instead of the private method selection strategy.

Target accuracy for SVM: ~83%
Target accuracy for SVM with Dropout: ~85% [I think we don't do Dropout right now]

NOTE: The paper uses 30 training examples per class, while we are using all. Since the dataset is a bit skewed (e.g. there are 55 training examples for accordion and 800 for airplanes), there may be a certain bias, which affects the overall accuracy. We need to adjust the preprocessor accordingly (#9).

Check datasets/README.md for details on the directory structure etc.
Check: https://github.com/Yangqing/decaf/blob/86b39770476ab2865018ca976321ac1e99e286da/decaf/layers/data/cub.py