GP / gp1.py
woletee
this is the commit for adding the gp interface
ded89f7
raw
history blame
7.52 kB
import numpy as np
import sys, os, json
from deap import base, creator, gp, tools, algorithms
from dsl import *
import glob
from dsl import _objects
# Custom type definition (DEAP compatibility)
class GridList(list):
pass
# DEAP GP Setup
creator.create("FitnessMax", base.Fitness, weights=(1.0,))
creator.create("Individual", gp.PrimitiveTree, fitness=creator.FitnessMax)
pset = gp.PrimitiveSetTyped("MAIN", [Grid], Grid)
# Basic Grid primitives (all your previously defined primitives)
pset.addPrimitive(ic_compress2, [Grid], Grid)
pset.addPrimitive(flipy, [Grid], Grid)
pset.addPrimitive(rot90, [Grid], Grid)
pset.addPrimitive(rot180, [Grid], Grid)
pset.addPrimitive(mirrorX, [Grid], Grid)
pset.addPrimitive(mirrorY, [Grid], Grid)
pset.addPrimitive(overlay, [Grid, Grid], Grid)
pset.addPrimitive(set_bg, [int, Grid], Grid)
pset.addPrimitive(ic_connectY, [Grid], Grid)
pset.addPrimitive(ic_connectX, [Grid], Grid)
pset.addPrimitive(ic_compress3, [Grid], Grid)
pset.addPrimitive(ic_erasecol, [int, Grid], Grid)
pset.addPrimitive(left_half, [Grid], Grid)
pset.addPrimitive(right_half, [Grid], Grid)
pset.addPrimitive(top_half, [Grid], Grid)
pset.addPrimitive(repeatX, [Grid], Grid)
pset.addPrimitive(flipx, [Grid], Grid)
pset.addPrimitive(setcol, [Colour, Grid], Grid)
pset.addPrimitive(ic_embed, [Grid, Grid], Grid)
pset.addPrimitive(rot270, [Grid], Grid)
# GridList-based primitives
pset.addPrimitive(ic_splitall, [Grid], GridList)
pset.addPrimitive(ic_composegrowing, [GridList], Grid)
pset.addPrimitive(lambda x: GridList([x]), [Grid], GridList, name="toGridList")
pset.addTerminal(GridList([]), GridList)
pset.addPrimitive(ic_pickunique, [GridList], Grid)
pset.addPrimitive(gravity_right, [Grid], Grid)
pset.addPrimitive(split8, [Grid], GridList)
pset.addPrimitive(ic_makeborder, [Grid], Grid)
pset.addPrimitive(ic_filtercol, [Colour, Grid], Grid)
pset.addPrimitive(ic_invert, [Grid], Grid)
pset.addPrimitive(logical_and, [Grid, Grid], Grid)
pset.addPrimitive(fillobj, [Colour, Grid], Grid)
pset.addPrimitive(topcol, [Grid], Colour)
pset.addPrimitive(rarestcol, [Grid], Colour)
pset.addPrimitive(gravity_down, [Grid], Grid)
pset.addPrimitive(pickcommon, [GridList], Grid)
pset.addPrimitive(swapxy, [Grid], Grid)
pset.addPrimitive(topcol, [Grid], Colour)
pset.addPrimitive(setcol, [Colour, Grid], Grid)
pset.addPrimitive(get_bg, [Grid], Colour)
pset.addPrimitive(rarestcol, [Grid], Colour)
pset.addPrimitive(ic_fill, [Colour, Grid], Grid)
pset.addPrimitive(ic_center, [Grid], Grid)
pset.addPrimitive(countToY, [Count, Colour], Grid)
pset.addPrimitive(countPixels, [Grid], Count)
pset.addPrimitive(ic_splitcols, [Grid], GridList)
pset.addPrimitive(grid_split, [Grid], GridList)
pset.addPrimitive(_objects, [Grid], GridList)
pset.addPrimitive(overlay, [Grid, Grid], Grid)
pset.addPrimitive(stack_no_crop, [GridList], Grid)
pset.addPrimitive(move_down, [Grid], Grid)
pset.addPrimitive(draw_line, [Grid, int], Grid)
pset.addPrimitive(draw_line_slant_up, [Grid, Grid], Grid)
pset.addPrimitive(draw_line_slant_down, [Grid], Grid)
pset.addPrimitive(rarestcol, [Grid], int)
pset.addPrimitive(lambda: 1, [], int, name="int_one")
# Integer terminals
for i in range(1, 10):
pset.addTerminal(i, int)
import operator # needed for operator.attrgetter
toolbox = base.Toolbox()
toolbox.register("compile", gp.compile, pset=pset)
toolbox.register("select", tools.selTournament, tournsize=3)
# Use leaf-biased crossover to avoid excessive tree growth
toolbox.register("mate", gp.cxOnePointLeafBiased, termpb=0.1)
# Mutation setup remains the same
toolbox.register("expr_mut", gp.genFull, min_=0, max_=2)
toolbox.register("mutate", gp.mutUniform, expr=toolbox.expr_mut, pset=pset)
# Explicitly limit tree height to avoid memory errors
MAX_TREE_HEIGHT = 17 # recommended limit
toolbox.decorate("mate", gp.staticLimit(operator.attrgetter("height"), MAX_TREE_HEIGHT))
toolbox.decorate("mutate", gp.staticLimit(operator.attrgetter("height"), MAX_TREE_HEIGHT))
# Population initialization (unchanged)
toolbox.register("population", tools.initRepeat, list,
lambda: creator.Individual(gp.genHalfAndHalf(pset, min_=1, max_=3)))
def evaluate_task(individual, task):
func = toolbox.compile(expr=individual)
total = 0
for example in task['train']:
inp = Grid(np.array(example["input"]))
tgt = Grid(np.array(example["output"]))
try:
out = func(inp)
if out.grid.shape == tgt.grid.shape:
total += np.sum(out.grid == tgt.grid)
except:
pass
return (total,)
toolbox.register("evaluate", evaluate_task)
# Folder containing your tasks
training_folder = "./training/"
task_files = glob.glob(training_folder + "*.json")
results = []
# Evaluate each task separately
for task_file in task_files:
task_name = os.path.basename(task_file)
print(f"Processing {task_name}")
with open(task_file, 'r') as f:
task = json.load(f)
# GP initialization per task
pop = toolbox.population(n=150)
hof = tools.HallOfFame(1)
for gen in range(250): # Adjust number of generations if needed
offspring = algorithms.varAnd(pop, toolbox, cxpb=0.5, mutpb=0.2)
fits = toolbox.map(lambda ind: toolbox.evaluate(ind, task), offspring)
for fit, ind in zip(fits, offspring):
ind.fitness.values = fit
hof.update(offspring)
pop = toolbox.select(offspring, k=len(pop))
# Evaluate best individual on the test set
best_ind = hof[0]
func = toolbox.compile(expr=best_ind)
correct = False
try:
for test_case in task["test"]:
test_inp = Grid(np.array(test_case["input"]))
expected_out = np.array(test_case["output"])
output_grid = func(test_inp).grid
if np.array_equal(output_grid, expected_out):
correct = True
else:
correct = False
break
except:
correct = False
results.append({
"task_name": task_name,
"best_program": str(best_ind),
"solution_found": correct
})
print(f"{task_name} completed. Solution found: {correct}")
# Save summary results to JSON
with open("tasks_results_summary.json", "w") as f:
json.dump(results, f, indent=2)
print("All tasks processed. Results saved to tasks_results_summary.json.")
# Put everything you already have here...
# And then add this at the bottom:
def run_task(task_path):
with open(task_path, 'r') as f:
task = json.load(f)
pop = toolbox.population(n=150)
hof = tools.HallOfFame(1)
for gen in range(250):
offspring = algorithms.varAnd(pop, toolbox, cxpb=0.5, mutpb=0.2)
fits = toolbox.map(lambda ind: toolbox.evaluate(ind, task), offspring)
for fit, ind in zip(fits, offspring):
ind.fitness.values = fit
hof.update(offspring)
pop = toolbox.select(offspring, k=len(pop))
best_ind = hof[0]
func = toolbox.compile(expr=best_ind)
# We'll just use the first test example for visual output
test_example = task["test"][0]
input_grid = np.array(test_example["input"])
target_grid = np.array(test_example["output"])
try:
output_grid = func(Grid(input_grid)).grid
correct = np.array_equal(output_grid, target_grid)
except:
output_grid = np.zeros_like(input_grid)
correct = False
return str(best_ind), correct, input_grid.tolist(), target_grid.tolist(), output_grid.tolist()