IICGP

(Interactive) Interpretable Cartesian Genetic Programming

To build and install a Julia binding for OpenCV, we recommend the following tutorial.

CGP Functions computational costs

CPU used for measurement: Intel(R) Core(TM) i5-7440HQ CPU @ 2.80GHz

Scalar functions:

Function	@btime
f_add	73.729 ns (4 allocations: 64 bytes)
f_subtract	72.271 ns (4 allocations: 64 bytes)
f_mult	73.489 ns (4 allocations: 64 bytes)
f_div	76.209 ns (4 allocations: 64 bytes)
f_abs	71.767 ns (4 allocations: 64 bytes)
f_sqrt	73.844 ns (4 allocations: 64 bytes)
f_pow	169.109 ns (4 allocations: 64 bytes)
f_exp	80.602 ns (4 allocations: 64 bytes)
f_sin	73.807 ns (4 allocations: 64 bytes)
f_cos	77.198 ns (4 allocations: 64 bytes)
f_tanh	75.054 ns (4 allocations: 64 bytes)
f_sqrt_xy	72.361 ns (4 allocations: 64 bytes)
f_lt	72.076 ns (4 allocations: 64 bytes)
f_gt	71.393 ns (4 allocations: 64 bytes)
f_and	73.774 ns (4 allocations: 64 bytes)
f_or	75.554 ns (4 allocations: 64 bytes)
f_xor	77.220 ns (4 allocations: 64 bytes)
f_not	71.539 ns (4 allocations: 64 bytes)

Image functions on full size image (160x210 = 33600 pixels):

Function	@btime
f_dilate	129.781 μs (9 allocations: 65.88 KiB)
f_erode	129.623 μs (9 allocations: 65.88 KiB)
f_subtract	18.095 μs (3 allocations: 65.78 KiB)
f_remove_details	521.964 μs (33 allocations: 263.45 KiB)
f_make_boxes	2.521 ms (5896 allocations: 2.87 MiB)
f_felzenszwalb_segmentation	20.024 ms (136549 allocations: 15.51 MiB)
f_components_segmentation	990.960 μs (71 allocations: 1.74 MiB)
f_box_segmentation	2.224 ms (784 allocations: 3.24 MiB)
f_threshold	114.592 μs (10 allocations: 546.36 KiB)
f_binary	152.759 μs (17 allocations: 616.38 KiB)
f_corners	1.769 ms (14 allocations: 1.17 MiB)
f_gaussian	942.508 μs (128 allocations: 1.70 MiB)
f_laplacian	580.035 μs (31 allocations: 1.35 MiB)
f_sobel_x	557.711 μs (114 allocations: 1.68 MiB)
f_sobel_y	539.037 μs (114 allocations: 1.68 MiB)
f_canny	6.954 ms (567 allocations: 8.05 MiB)
f_edges	1.434 ms (197 allocations: 3.80 MiB)
f_opening	253.334 μs (15 allocations: 65.97 KiB)
f_closing	253.296 μs (15 allocations: 65.97 KiB)
f_tophat	272.295 μs (17 allocations: 131.73 KiB)
f_bothat	272.230 μs (17 allocations: 131.73 KiB)
f_morphogradient	277.672 μs (19 allocations: 197.50 KiB)
f_morpholaplace	594.862 μs (29 allocations: 2.31 MiB)
f_bitwise_not	13.469 μs (3 allocations: 65.78 KiB)
f_bitwise_and	18.322 μs (3 allocations: 65.78 KiB)
f_bitwise_or	18.547 μs (3 allocations: 65.78 KiB)
f_bitwise_xor	18.309 μs (3 allocations: 65.78 KiB)
f_motion_capture	179.721 μs (9 allocations: 197.41 KiB)
f_motion_distances	3.327 ms (66773 allocations: 3.48 MiB)

Image functions on halved-size image (80x105 = 8400 pixels):

Function	@btime
f_threshold	19.376 μs (8 allocations: 79.16 KiB)
f_subtract	5.704 μs (1 allocation: 8.38 KiB)
f_binary	25.306 μs (14 allocations: 85.22 KiB)
f_erode	17.323 μs (7 allocations: 8.47 KiB)
f_dilate	17.316 μs (7 allocations: 8.47 KiB)
f_bitwise_not	4.199 μs (1 allocation: 8.38 KiB)
f_bitwise_and	5.803 μs (1 allocation: 8.38 KiB)
f_bitwise_or	5.653 μs (1 allocation: 8.38 KiB)
f_bitwise_xor	6.126 μs (1 allocation: 8.38 KiB)
f_motion_capture	25.473 μs (5 allocations: 25.22 KiB)

Features extraction

Mean-pooling	Centroid

GA - mutant evaluation

function fit(e::CGPInd, c::CGPInd, seed::Int64)
	p = 1.0
	lmin = min(length(e.chromosome), length(c.chromosome))
	lmin = Int64(round(p*lmin))
	res = 0.0
	for i in 1:lmin
	    if abs(e.chromosome[i] - c.chromosome[i]) < 0.00001
	        res += 1.0
	    end
	end
	return res
end

Best	Best	Mean

Suggestions

High priority to low priority:

• Save state_ref
• Speed-up centroid reducer's reducing function
• Allow bootstrap of both encoder and controller
• An R^2 controller reasonning about centroids directly?

Reducing experiments length

Speed-up evolution

• Less frames for evaluation (curently 10000)
• Sticky actions, less forward passes in evaluation
• Add preliminary tests to Atari evaluation, e.g.:
  • forward-pass less than 1ms
  • Assert that the encoder only applies a limited number of filters before one output
  • evaluate the quality of the produced feature vector:
    • measure the amount of information and compare to input image (proxy to entropy measure?)
    • reconstruction error (without backward pass?)

Speed-up forward-pass

• Lower the number of input (e.g. centroid reduction may be as usefull applied on 1 grayscale image as on 3 RGB images)
• Encoder: max-out number of filters applied sequentially (if we assume our filters are well chosen, few would be enough)
• Downscale all input images (max-pool dividing the image size by 2)
• Low-cost image filters (threshold, binary, subtract, erode, dilate, not, and, or, xor)

Results

2021-08-11 (5 days):

• Reducer + controller (no encoder)
• 3 RGB images for pooling
• 1 Grayscale image for centroid

Best score	Mean score

Forward pass timing (ms)

	Pooling	Centroid
assault	0.27	0.7
boxing	0.33	0.7
freeway	0.3	0.8

2021-08-24 (5 days):

• Encoder + Reducer + controller
• Downscale (1/2 image)
• Grayscale

Best score	Mean score

Forward pass timing (ms)

	Pooling	Centroid
assault	0.18	0.46
boxing	0.2	1.22

Julia dev env

• rm Cambrian
• rm CartesianGeneticProgramming
• add https://github.com/erwanlecarpentier/Cambrian.jl.git#ipcgp
• add https://github.com/erwanlecarpentier/CartesianGeneticProgramming.jl.git#ipcgp

Python Install

source /path/to/my/virtual/environment/bin/activate
pip install -r py-requirements.txt