Awesome

DeeSIL

new code version of DeeSIL. Unofficial but reproduces the results.

set up your env

using conda, install the required packages (of package-list.txt)

conda create -n deesil --file package-list.txt

and add utils/ to your env path conda develop path_to_utils/

train the first state

edit configs/scratch_ilsvrc_s10.cf to set your storing paths (e.g. path_to_model_dir) then train the first state model using:

python code/scratch.py configs/scratch_ilsvrc_s10.cf

(I put a trained model in the models/ folder in order to facilitate the results reproductions)

extract all the features

edit configs/features_extraction_ilsvrc_s10.cf to set your storing paths (e.g. path_to_model_dir and the root_path (that will be used then to store the features/, svms/ and predictions/ folder)) then extract the required features using:

python code/features_extraction.py configs/features_extraction_ilsvrc_s10.cf

train the DeeSIL classifiers

run the following commands, it will train the classifiers of DeeSIL for each state and store them in the right folder (replace root_path accordingly to the previous step)

python codes/train_deesil.py 1 0.0001 1 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 2 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 3 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 4 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 5 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 6 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 7 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 8 100 ilsvrc s10 root_path
python codes/train_deesil.py 1 0.0001 9 100 ilsvrc s10 root_path

(1 and 0.0001 stands for the C and tol parameters and are already optimized)

compute the predictions files

run the following commands, it will compute the predictions files of DeeSIL store them in the right folder (replace root_path accordingly to the previous steps)

python codes/compute_predictions_deesil.py ilsvrc s10 root_path

evaluate DeeSIL

run the following commands, it will evaluate DeeSIL accordingly to the predictions files of DeeSIL computed at the previous step (replace root_path accordingly to the previous steps)

python codes/eval_deesil.py ilsvrc s10 root_path

Results:

You should get for ILSVRC s10:

batch 1, top1 = 0.840, top5 = 0.949
batch 2, top1 = 0.654, top5 = 0.885
batch 3, top1 = 0.564, top5 = 0.844
batch 4, top1 = 0.501, top5 = 0.793
batch 5, top1 = 0.457, top5 = 0.755
batch 6, top1 = 0.427, top5 = 0.728
batch 7, top1 = 0.405, top5 = 0.705
batch 8, top1 = 0.380, top5 = 0.682
batch 9, top1 = 0.366, top5 = 0.666
batch 10, top1 = 0.351, top5 = 0.648
=================================================
===================  TOTAL  =====================
top1 = 0.456, top5 = 0.745