Skip to content

Latest commit

 

History

History
43 lines (33 loc) · 1.26 KB

Experiments.md

File metadata and controls

43 lines (33 loc) · 1.26 KB

Experimental Setup

  • Download PASCAL VOC 2012 dataset
  • Download SBD dataset + Use train.txt provided by SBD instead of PASCAL.
  • Modify config.json path to your SBD_PATH
  • In all Experiments modify in the config files "path" to the PASCAL_VOC_PATH

Reproducing experiments in Table 1, 2, 3

  1. Unzip runs.zip in the same code folder

  2. Run:

./run_exps.sh
  1. Saved logs corresponding to each experiment:
    • logs_1shot: AMP-2 in Table 1
    • logs_5shot: AMP-2 in Table 2
    • logs_1shot_fgbg: AMP-2 in Table 3 1-shot
    • logs_5shot_fgbg: AMP-2 in Table 3 5-shot
    • logs_5shot_fgbg: AMP-2+FT in Table 3 5-shot

Reproducing iPASCAL results

  1. Checkout the branch for continual learning mode pascal_multirun
  2. Run:
./run_exps_cl.sh
  1. Plot the results
python plot_cl_multiruns.py cl_results_

Running Hyperparameter Search

Based on Tune, follow the instructions for installation here

git checkout hyperopt
python hyperparam_search.py --config CONFIG_FILE --model_path PATH --binary BINARY_FLAG