This repository contains the implementation of the slowly changing regression problem, previously known as the bit-flipping problem.
The first step to replicate the results is to create the data.
cd lop/slowly_changing_regression
mkdir env_temp_cfg temp_cfg
python3.8 multi_param_expr.py -c cfg/prob.json This step creates 100 temporary cfg files in env_temp_cfg,
as cfg/prob.json specifies 100 runs.
Each of the new temporary cfg files can then be used to generate data using slowly_changing_regression.py.
The command below creates data for one run.
python3.8 slowly_changing_regression.py -c env_temp_cfg/0.json The next step is to test a learning network that uses backprop to learn.
This command produces 100 temporary cfg files in temp_cfg.
python3.8 multi_param_expr.py -c cfg/sgd/bp/relu.json To test continual backprop, use a different cfg file, like cfg/sgd/cbp/relu.json
Each of the new temporary cfg files can then be used to do one run of backprop. Each run takes about 15 CPU-minutes on a normal laptop.
python3.8 expr.py -c temp_cfg/0.json Finally, after completing all the runs, the command below can be used to generate the plot for the ReLU network.
cd plots
python3.8 online_performance.py -c ../cfg/sgd/bp/relu.json The command above will generate the figure in the left pane of the plot below. The plot below shows the performance averaged over 100 runs. I recommend doing at least 30 runs, as there can be a high variance among individual runs. The plot below bins the error into bins of size 20000.
