STrucks/RS-Internship

RS-Internship

ZSL with hyper-spectral data:

Data distribution:

\	% of whole dataset
1	0.00448824275539077
2	0.13933066640647868
3	0.08098351058639867
4	0.023124207239730705
5	0.047126548931603084
6	0.07122646111815786
7	0.002731973851107425
8	0.04663869645819104
9	0.0019514098936481608
10	0.09483852083130062
11	0.23953556444531174
12	0.057859303346667966
13	0.020001951409893647
14	0.12342667577324618
15	0.03766221094740951
16	0.009074056005463947

Data Preprocessing:

Data normalization works. Balancing also works.

Data distribution:

To make the data really balanced, we only take the classes with 100 or more samples.

Lets see how well the autoencoder encodes and decodes:

Losses:

Typical loss of the classifier:

Experiments:

Approach	left-out class	mean accuracy on test set	std of accuracy	mean accuracy on left-out classes
idea 1	-	30.5%	2.45%	-
idea 1	2	28.9%	1.53%	0.0%
idea 1	2,3	30.2%	1.97%	0.0%
idea 1	2,3,4	30.2%	1.97%	0.0%
idea 1	2,3,4,5	37.3%	2.66%	0.0%
idea 2	-	34%	1.37%	-
idea 2	2	29.7%	4.8%	0.0%
idea 2	2,3	34%	1,37%	0.0%
idea 2	2,3,4	32.4%	1.37%	0.0%
idea 2	2,3,4,5	36.4%	2.91%	0.0%
idea 2	2,3,4,5	36.4%	2.91%	0.0%
auto encoder features	-	39.6%	4.14%	0.0%
auto encoder features	2	27.8%	1.3%	0.0%
auto encoder features	2,3	36.7%	3.31%	0.0%
auto encoder features	2,3,4	-%	-%	0.0%
auto encoder features	2,3,4,5	-%	-%	0.0%
variantional auto encoder features	-	30.9%	1.53%	0.0%
variantional auto encoder features	2	32.5%	1.4%	0.0%
variantional auto encoder features	2,3	35.1%	4.9%	0.0%
variantional auto encoder features	2,3,4	28.9%	3.1%	0.0%
variantional auto encoder features	2,3,4,5	29.7%	2.7%	0.0%