def __init__(self,
in_dim,
emb_dim,
pb_limits=None,
batch_sz=128,
log_dir="/tmp/"):
self.frozen = MiscUtils.SmallEncoder1d(in_dim,
emb_dim,
num_hidden=3,
non_lin="leaky_relu",
use_bn=False)
self.frozen.eval()
self.learnt = MiscUtils.SmallEncoder1d(in_dim,
emb_dim,
num_hidden=5,
non_lin="leaky_relu",
use_bn=False)
self.optimizer = torch.optim.Adam(self.learnt.parameters(), lr=1e-2)
self.archive = None
self.pop = None
self.batch_sz = batch_sz
self.epoch = 0
self.log_dir = log_dir
if pb_limits is not None:
MiscUtils.make_networks_divergent(self.frozen,
self.learnt,
pb_limits,
iters=50)
def compare_execution_time(min_d_e, max_d_e, pop_sz=50, offspring_sz=100, archive_size=8000,num_tests=50):
global_time_networks = []
global_time_archive = []
bd_size_range = [2**x for x in range(min_d_e,max_d_e+1,1)]
print(bd_size_range)
for descr_dim in bd_size_range:
num_gens = num_tests
emb_dim = descr_dim * 2
pop_bds = torch.rand(offspring_sz + pop_sz, descr_dim)
frozen = MiscUtils.SmallEncoder1d(descr_dim,
emb_dim,
num_hidden=3,
non_lin="leaky_relu",
use_bn=False)
frozen.eval()
learnt = MiscUtils.SmallEncoder1d(descr_dim,
emb_dim,
num_hidden=5,
non_lin="leaky_relu",
use_bn=False)
time_hist = []
frozen.eval()
optimizer = torch.optim.Adam(
[x for x in learnt.parameters() if x.requires_grad], lr=1e-2)
batch_sz = 256
for i in range(num_gens):
t1_n = time.time()
learnt.eval()
for batch_i in range(0, pop_bds.shape[0], batch_sz):
batch = pop_bds[batch_i:batch_i + batch_sz]
with torch.no_grad():
e_frozen = frozen(batch)
e_pred = learnt(batch)
nov = (e_pred - e_frozen).norm(dim=1)
learnt.train()
#this is how training is done, note that we can further reduce runtime by removing the extra frozen forward passes that we've made before when computing novelty
for _ in range(5):#In the experiments presented in the paper, usually this was set to 3 so BR-NS in the paper should be slightly faster
for batch_i in range(0, pop_bds.shape[0], batch_sz):
batch = pop_bds[batch_i:batch_i + batch_sz]
with torch.no_grad():
e_frozen = frozen(batch)
optimizer.zero_grad()
e_l = learnt(batch)
loss = (e_l - e_frozen).norm()**2
loss /= batch_sz
loss.backward()
optimizer.step()
t2_n = time.time()
time_hist.append(t2_n - t1_n)
mean_t_nets = np.array(time_hist).mean()
print(descr_dim, mean_t_nets)
global_time_networks.append(mean_t_nets)
if 1:
knn_k = 15
kdt_bds = np.random.rand(archive_size, descr_dim)
times = []
for i in range(num_gens):
t1_a = time.time()
#note that the kdtree has to be created everytime as after adding elements, you can't just reuse the same kdtree (some cells might have become much more dense)
kdt = KDTree(kdt_bds, leaf_size=20, metric='euclidean')
dists, ids = kdt.query(pop_bds, knn_k, return_distance=True)
t2_a = time.time()
times.append(t2_a - t1_a)
mean_t_arch = np.array(times).mean()
global_time_archive.append(mean_t_arch)
print(descr_dim, mean_t_arch)
if 1:
gt_arc_ms = [x * 1000 for x in global_time_archive]
gt_net_ms = [x * 1000 for x in global_time_networks]
plt.plot(bd_size_range,
gt_arc_ms,
"r",
label=f"Archive-based NS (size=={archive_size})",
linewidth=5)
plt.plot(bd_size_range, gt_net_ms, "b", label="BR-NS", linewidth=5)
plt.grid("on")
plt.legend(fontsize=28)
plt.xlabel("behavior descriptor dimensionality", fontsize=28)
plt.ylabel("time (ms)", fontsize=28)
plt.xticks(fontsize=28)
plt.yticks(fontsize=28)
plt.xlim(0, 2**max_d_e)
plt.show()
