def fitnessFunction(self, individual, lambda_):
if self.functionType_ != "gpTSNE":
return super().fitnessFunction(individual, lambda_)
else:
fitness = evalTSNEMO(self.data_t, self.toolbox, individual)
return self.skew_fitness(fitness, lambda_)
def eval_pso(gp_ind, ephemerals_indxs, toolbox, data_t, reference, ind):
update_ercs(ephemerals_indxs, gp_ind, ind, reference)
return evalTSNEMO(data_t, toolbox, gp_ind)
creator.create("FitnessMin", base.Fitness, weights=(-1.0,) * tsnedata.nobj)
creator.create("Individual", list, fitness=creator.FitnessMin, pset=pset)
toolbox = ParallelToolbox()
toolbox.register("compile", gp.compile, pset=pset)
toolbox.register("evaluate", evalTSNEMO, tsnedata.data_t, toolbox)
tree1 = from_string_np_terms(tree_1_str, pset)
tree2 = from_string_np_terms(tree_2_str, pset)
print(str(tree1))
print(str(tree2))
ind = creator.Individual([tree1, tree2])
print(ind)
tsnedata.fitnessCache = cachetools.LRUCache(maxsize=1e6)
tsnedata.outdir = args.outdir
tsnedata.dataset = args.dataset
tsnedata.degrees_of_freedom = max(num_trees - 1, 1)
tsnedata._DOF = (tsnedata.degrees_of_freedom + 1.0) / -2.0
dists = t_sne.pairwise_distances(data, metric="euclidean", squared=True)
tsnedata.P_tsne = t_sne._joint_probabilities(dists, perplexity, verbose=True)
tsnedata.max_P_tsne = np.maximum(tsnedata.P_tsne, MACHINE_EPSILON)
best, reference = do_pso(ind, toolbox, tsnedata.data_t, args.gens)
print(best)
# We still want to output ones that have no constants (even though the vals didn't change!)
if best is not None:
ephemerals_indxs = collect_ephemeral_indices(ind) # [(0,5),....(1,4),...]
update_ercs(ephemerals_indxs, ind, best, reference)
ind.fitness.setValues(evalTSNEMO(tsnedata.data_t, toolbox, ind))
output_ind(ind, toolbox, tsnedata, suffix="-pso", compress=False)
def eval_cmaes(gp_ind, ephemerals_indxs, toolbox, data_t, ind):
# ind is a list here.
update_ercs(ephemerals_indxs, gp_ind, ind)
return evalTSNEMO(data_t, toolbox, gp_ind)
output_ind(res, toolbox, tsnedata, compress=False)
p = Path(tsnedata.outdir, args.logfile + '.gz')
with gz.open(p, 'wt') as file:
file.write(str(logbook))
pop_stats = [str(p.fitness) for p in pop]
pop_stats.sort()
hof_stats = [str(h.fitness) for h in hof]
# hof_stats.sort()
print("POP:")
print("\n".join(pop_stats))
print("PF:")
print("\n".join(hof_stats))
# do it all after so if it's too slow at least the above is outputted already
if post_cmaes:
for res in hof:
cmaes_res = do_pso(res, toolbox, tsnedata.data_t, 1000)
if cmaes_res: # since this is just the vector of consts...
ephemerals_indxs = collect_ephemeral_indices(
res) # [(0,5),....(1,4),...]
for index, val in enumerate(cmaes_res):
ep_indx = ephemerals_indxs[index]
res[ep_indx[0]][ep_indx[1]].value = val
res.fitness.setValues(evalTSNEMO(tsnedata.data_t, toolbox,
res))
output_ind(res, toolbox, tsnedata, "-pso")