def optim_result(param):
pop_size = int(param[0, 0])
select_rate = param[0, 1]
select_size = int(pop_size * select_rate)
lr = param[0, 2]
mu_init = param[0, 3]
mu_rate = param[0, 4]
mode = int(param[0, 5])
sphere_norm = int(param[0, 6])
n_gen = 3000 // pop_size # 4000 is the total evaluation budget
optimizer = ZOHA_Sphere_lr_euclid(120,
population_size=pop_size,
select_size=select_size,
lr=lr,
maximize=True,
rankweight=True,
rankbasis=True,
sphere_norm=sphere_norm)
optimizer.lr_schedule(
n_gen, lim=(mu_init, mu_rate * mu_init), mode=mode_dict[mode]
) # the space is much tighter than the fc6 space, so step size
# should be tuned for that
Exp = ExperimentEvolve(("alexnet", "fc8", 1),
max_step=n_gen,
backend="torch",
optimizer=optimizer,
GAN="BigBiGAN",
verbose=False)
# note using torch and batch processing evolution is much faster ~ 27sec 100 gens
Exp.run()
return np.percentile(
Exp.scores_all[Exp.generations > Exp.generations.max() - 5], 99.5)
def optim_result(param):
pop_size = int(param[0,0])
select_rate = param[0,1]
select_size = int(pop_size * select_rate)
lr = param[0,2]
mu_init = param[0,3]
mu_rate = param[0,4]
mode = int(param[0,5])
n_gen = 4000 // pop_size # 4000 is the total evaluation budget
optimizer = ZOHA_Sphere_lr_euclid(4096, population_size=pop_size, select_size=select_size, lr=lr)
optimizer.lr_schedule(n_gen=n_gen, lim=(mu_init, mu_rate*mu_init), mode=mode_dict[mode])
Exp = ExperimentEvolve(("alexnet", "fc8", 1), max_step=n_gen, backend="torch", optimizer=optimizer, GAN="fc6")
# note using torch and batch processing evolution is much faster ~ 27sec 100 gens
# Exp = ExperimentEvolve(("caffe-net", "fc8", 1), max_step=50, backend="caffe", optimizer=optimizer, GAN="fc6")
Exp.run()
return np.percentile(Exp.scores_all, 99.5)
from time import time
import matplotlib.pylab as plt
for unit in unit_arr[:]:
savedir = join(recorddir, "optim_cmp", "%s_%s_%d" % (unit[0], unit[1], unit[2]))
os.makedirs(savedir, exist_ok=True)
for Optim_str in Optim_arr:
for triali in range(10):
t0 = time()
if Optim_str == "Genetic":
optim = Genetic(population_size, mutation_rate, mutation_size, kT_multiplier, recorddir,
parental_skew=0.5, n_conserve=0)
elif Optim_str == "CholCMA":
optim = CholeskyCMAES(recorddir=recorddir, space_dimen=code_length, init_sigma=init_sigma,
Aupdate_freq=Aupdate_freq, init_code=initcode) # np.zeros([1, code_length])
experiment = ExperimentEvolve(unit, max_step=100, optimizer=optim)
experiment.run()
fig0 = experiment.visualize_best(show=False)
fig0.savefig(join(savedir, "%s_BestImgTrial%03d.png" % (Optim_str, triali,)))
fig = experiment.visualize_trajectory(show=False)
fig.savefig(join(savedir, "%s_ScoreTrajTrial%03d.png" % (Optim_str, triali,)))
# fig2 = experiment.visualize_exp(show=False)
# fig2.savefig(join(savedir, "EvolveTrial%03d.png" % (triali)))
plt.close('all')
np.savez(join(savedir, "%s_scores_trial%03d.npz" % (Optim_str, triali)),
generations=experiment.generations,
scores_all=experiment.scores_all)
print("Optimization with %s took %.1f s" % (Optim_str, time() - t0))
# lastgen_max = [experiment.scores_all[experiment.generations == geni].max() for geni in
# range(experiment.generations.max() - 10, experiment.generations.max() + 1)]
# best_scores_col.append(lastgen_max)
eigvect_avg = data["eigvect_avg"]
savedir = os.path.join(recorddir, "%s_%s_eig_subspac%d" % (netname, layer, subspace_d))
os.makedirs(savedir, exist_ok=True)
pos_dict = {"conv5": (7, 7), "conv4": (7, 7), "conv3": (7, 7), "conv2": (14, 14), "conv1": (28, 28)}
best_scores_col = []
for chi in range(100):
if "fc" in layer:
unit = (netname, layer, chi)
else:
unit = (netname, layer, chi, *pos_dict[layer])
for triali in range(10):
optimizer = ZOHA_Sphere_lr_euclid_ReducDim(4096, subspace_d, population_size=40, select_size=20)
optimizer.lr_schedule(n_gen=n_gen, mode="inv")
optimizer.get_basis("rand")
experiment = ExperimentEvolve(unit, max_step=n_gen, backend="torch", optimizer=optimizer, GAN="fc6")
experiment.run()
fig0 = experiment.visualize_best(show=False)
fig0.savefig(join(savedir, "Subspc%dBestImgChan%02dtr%01d.png" % (subspace_d, chi, triali)))
fig = experiment.visualize_trajectory(show=False)
fig.savefig(join(savedir, "Subspc%dScoreTrajChan%02dtr%01d.png" % (subspace_d, chi, triali)))
fig2 = experiment.visualize_exp(show=False)
fig2.savefig(join(savedir, "Subspc%dEvolveChan%02dtr%01d.png" % (subspace_d, chi, triali)))
plt.close("all")
np.savez(join(savedir, "scores_subspc%dChan%02dtr%01d.npz" % (subspace_d, chi, triali)),
generations=experiment.generations,
scores_all=experiment.scores_all, codes_fin=experiment.codes_all[experiment.generations==experiment.max_steps-1,:])
lastgen_max = [experiment.scores_all[experiment.generations == geni].max() for geni in
range(experiment.generations.max() - 10, experiment.generations.max() + 1)]
best_scores_col.append(lastgen_max)
# break
imgs = BigBiGAN_render(SLERP(Vec1, Vec2, ticks, (0, 1)), 255.0)
mtg = build_montages(imgs, (128, 128), (ticks, 1))[0]
Img = Image.fromarray(np.uint8(mtg))
Img.show()
#%%
optimizer = CholeskyCMAES(120,
population_size=None,
init_sigma=0.25,
init_code=0.8 * np.random.randn(1, 120),
Aupdate_freq=10,
maximize=True,
random_seed=None,
optim_params={})
experiment = ExperimentEvolve(("alexnet", "fc8", 1),
max_step=100,
backend="torch",
optimizer=optimizer,
GAN="BigBiGAN",
verbose=True)
experiment.run()
#%%
optimizerZH = ZOHA_Sphere_lr_euclid(120,
population_size=40,
select_size=15,
lr=1.0,
maximize=True,
rankweight=True,
rankbasis=True,
sphere_norm=9)
optimizerZH.lr_schedule(
100, lim=(10, 6),
mode="lin") # the space is much tighter than the fc6 space, so step size
#%%
# optimizer = ZOHA_Sphere_lr_euclid(4096, population_size=40, select_size=20)
# optimizer.lr_schedule(n_gen=100)
# codes = np.random.randn(40, 4096)
# scores = np.random.randn(40)
# optimizer.step_simple(scores, codes)
#%%
if __name__ == "__main__":
from insilico_Exp import ExperimentEvolve
optimizer = ZOHA_Sphere_lr_euclid(4096, population_size=40, select_size=20)
optimizer.lr_schedule(n_gen=100, mode="exp")
Exp = ExperimentEvolve(("alexnet", "fc8", 1),
max_step=100,
backend="torch",
optimizer=optimizer,
GAN="fc6")
# note using torch and batch processing evolution is much faster ~ 27sec 100 gens
# Exp = ExperimentEvolve(("caffe-net", "fc8", 1), max_step=50, backend="caffe", optimizer=optimizer, GAN="fc6")
Exp.run()
#%%
optimizer = ZOHA_Sphere_lr_euclid_ReducDim(4096,
50,
population_size=40,
select_size=20)
optimizer.lr_schedule(n_gen=50, mode="exp")
optimizer.get_basis("rand")
Exp = ExperimentEvolve(("alexnet", "fc8", 1),
max_step=100,
backend="torch",