Interval-Insensitive-Loss

Performed ordinal classification on partially annotated examples.

Goal

Goal was to estimate the age (inter- val/range) of a person’s face.

1. A numbered list

Structure of Code

The following steps include the basic structure of code :

1. Read the entire data from the file.
2. Calculate the total number of classes or bins based on the provided bin size and size of data. number_of_classes = data_size / bin_size
3. Define the ranges (left extreme and right extreme) for each bin.
4. Segregate the data into files as “train” and “test” using Kfold algorithm.
5. Calculate values of different variables in the dual form and provide it to QP solver to find the Lagrangian constants .
6. Find weight vector using Lagrangian constants.
7. Classify the input samples using the obtained weight vector and determine accuarcy and mean absolute error loss.
8. Plot the graph based on the above results.

There are two files namely : binning.py and computation.py binning.py : steps 1 to 4 computation.py : steps 5 to 8

Execute the code

python computation.py bin_size Example: python computation.py 50

Output Format

bin_size
accuracy : case_1 case_2
MAE : case_1 case_2

Example:

10
Accuracy :  0.00963995354239 0.0720092915215
MAE :  89.9709639954 89.4425087108

50
Accuracy :  0.349361207898 0.49756097561
MAE :  69.6225319396 63.8850174216

100
Accuracy :  0.473403019744 0.583739837398
MAE :  54.6109175377 49.837398374

150
Accuracy :  0.564576074332 0.655400696864
MAE :  42.4390243902 34.5644599303

Refernce

http://proceedings.mlr.press/v39/antoniuk14.pdf