return False
if np.sum(state == 3) == 0: # make sure has at least one muscle voxel for movement
return False
return True
if not os.path.isfile("./" + RUN_DIR + "/pickledPops/Gen_0.pickle"):
random.seed(SEED)
np.random.seed(SEED)
my_sim = Sim(dt_frac=DT_FRAC, simulation_time=SIM_TIME, fitness_eval_init_time=INIT_TIME)
my_env = Env(temp_amp=TEMP_AMP, fluid_environment=FLUID_ENV, aggregate_drag_coefficient=AGGREGATE_DRAG_COEF,
lattice_dimension=VOXEL_SIZE, grav_acc=GRAV_ACC, frequency=FREQ, muscle_stiffness=STIFFNESS,
block_position=BLOCK_POS, block_material=BLOCK_MAT, external_block=True)
my_objective_dict = ObjectiveDict()
my_objective_dict.add_objective(name="fitness", maximize=True, tag=FITNESS_TAG,
# meta_func=min_energy
)
my_objective_dict.add_objective(name="age", maximize=False, tag=None)
my_objective_dict.add_objective(name="n_muscle", maximize=False, tag=None,
node_func=partial(count_occurrences, keys=[3]),
output_node_name="material")
# logging only:
my_objective_dict.add_objective(name="n_vox", maximize=False, tag=None, logging_only=True,
node_func=partial(count_occurrences, keys=[1, 3]),
if not os.path.isfile("./" + RUN_DIR + "/pickledPops/Gen_0.pickle"):
random.seed(SEED)
np.random.seed(SEED)
my_sim = Sim(dt_frac=DT_FRAC,
simulation_time=SIM_TIME,
fitness_eval_init_time=INIT_TIME)
my_env = Env(
temp_amp=TEMP_AMP,
fluid_environment=FLUID_ENV,
aggregate_drag_coefficient=AGGREGATE_DRAG_COEF,
lattice_dimension=VOXEL_SIZE,
grav_acc=GRAV_ACC,
frequency=FREQ,
muscle_stiffness=STIFFNESS,
fat_stiffness=STIFFNESS,
# block_position=2, block_material=8, external_block=False,
# falling_prohibited=STOP_IF_BLOCK_TOUCHES_GROUND
)
my_objective_dict = ObjectiveDict()
my_objective_dict.add_objective(name="fitness",
maximize=True,
tag=FITNESS_TAG)
my_objective_dict.add_objective(name="age", maximize=False, tag=None)
my_objective_dict.add_objective(name="n_muscle",
maximize=False,
tag=None,
return True
if not os.path.isfile("./" + RUN_DIR + "/pickledPops/Gen_0.pickle"):
random.seed(SEED)
np.random.seed(SEED)
my_sim = Sim(dt_frac=DT_FRAC,
simulation_time=SIM_TIME,
fitness_eval_init_time=INIT_TIME)
my_env = Env(temp_amp=TEMP_AMP,
fluid_environment=FLUID_ENV,
aggregate_drag_coefficient=AGGREGATE_DRAG_COEF,
lattice_dimension=VOXEL_SIZE,
grav_acc=GRAV_ACC,
frequency=FREQ,
muscle_stiffness=STIFFNESS)
my_objective_dict = ObjectiveDict()
my_objective_dict.add_objective(name="fitness",
maximize=True,
tag=FITNESS_TAG,
meta_func=favor_appendages)
my_objective_dict.add_objective(name="age", maximize=False, tag=None)
my_pop = Population(my_objective_dict,
MyGenotype,
MyPhenotype,
pop_size=POP_SIZE)
def regeneration(params, details=False, save_name=""):
#Caliculation initial cell states
individual_id = params[0][0]
w = params[0][1]
generations = params[1]
settings = params[2]
run_directory = settings["run_directory"]
im_size = settings["im_size"]
num_classes = settings["voxel_types"]
input_dim = settings["number_neighbors"]
if settings['growth_facter']:
input_dim = 3 * 9 * 2
num_classes = settings['voxel_types'] + 1
# Setting up the simulation object
sim = Sim(dt_frac=settings['fraction'],
simulation_time=settings['simtime'],
fitness_eval_init_time=settings['initime'])
# Setting up the environment object
env = Env(sticky_floor=0, time_between_traces=0)
p = MorphNet(input_dim=input_dim,
number_state=num_classes,
recurrent=settings['recurrent'])
vector_to_parameters(torch.tensor(w, dtype=torch.float32), p.parameters())
if settings['growth_facter']:
morphogens = np.zeros(shape=(2, im_size, im_size, im_size))
original_morphogens = np.zeros(shape=(2, im_size, im_size, im_size))
cutedgemorphogens = np.zeros(shape=(2, im_size - 2, im_size - 2,
im_size - 2))
originalcutedgemorphogens = np.zeros(shape=(2, im_size - 2,
im_size - 2, im_size - 2))
else:
cutedgemorphogens = np.zeros(shape=(2, im_size - 2, im_size - 2,
im_size - 2))
morphogens = np.zeros(shape=(1, im_size, im_size, im_size))
out = torch.zeros(1, im_size * im_size * im_size, num_classes)
dev_states = []
alphalist = []
sleep_id = []
iterations = 10
hidden_dim = 64
if settings['data_read']:
hidden_states = torch.ones(
size=(im_size, im_size, 2, 1, im_size * im_size * im_size,
hidden_dim)) #batch_size = im_size*im_size
hidden_states_batched_A = torch.ones(
size=(1, im_size * im_size * im_size,
hidden_dim)) #Hidden layers, batch_size, number units
hidden_states_batched_B = torch.ones(size=(1,
im_size * im_size * im_size,
hidden_dim))
max_fitness = params[6]
else:
hidden_states = torch.ones(
size=(im_size, im_size, 2, 1, im_size * im_size * im_size,
hidden_dim)) #batch_size = im_size*im_size
hidden_states_batched_A = torch.ones(
size=(1, im_size * im_size * im_size,
hidden_dim)) #Hidden layers, batch_size, number units
hidden_states_batched_B = torch.ones(size=(1,
im_size * im_size * im_size,
hidden_dim))
max_fitness = 0
# Import original morphology
original_morphogens = np.load('multiped_morphology.npy')
# Cutting
morphogens = copy.deepcopy(original_morphogens)
morphogens[:, :, :, 0:(im_size - 1) / 2] = 0
if settings['recurrent']:
hidden_states_batched_A = hidden_states_batched_A.reshape(
1, im_size, im_size, im_size, hidden_dim)
hidden_states_batched_B = hidden_states_batched_B.reshape(
1, im_size, im_size, im_size, hidden_dim)
hidden_states_batched_A[0, :, :, 0:(im_size - 1) / 2, :] = 1
hidden_states_batched_B[0, :, :, 0:(im_size - 1) / 2, :] = 1
hidden_states_batched_A = hidden_states_batched_A.reshape(
1, im_size**3, hidden_dim)
hidden_states_batched_B = hidden_states_batched_B.reshape(
1, im_size**3, hidden_dim)
morph_tmp = copy.deepcopy(morphogens)
#Insialize a fitness
#Start main loop
for it in range(iterations):
fitness = 0
batch_inputs = torch.zeros(im_size * im_size * im_size, input_dim)
counter = 0
sleep_id = []
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
if settings['growth_facter']:
cell_input = torch.flatten(
torch.tensor(morphogens[:, i - 1:i + 2,
j - 1:j + 2, k - 1:k + 2]))
else:
cell_input = torch.tensor([
morphogens[0, i - 1, j, k], morphogens[0, i + 1, j,
k],
morphogens[0, i, j + 1,
k], morphogens[0, i, j - 1,
k], morphogens[0, i, j,
k - 1],
morphogens[0, i, j, k + 1], morphogens[0, i, j, k]
])
batch_inputs[counter] = cell_input
counter += 1
output, hs = p(batch_inputs,
(hidden_states_batched_A, hidden_states_batched_B))
past_hidden_states_batched_A = hidden_states_batched_A
past_hidden_states_batched_B = hidden_states_batched_B
if not settings['recurrent']: #ALPHA
output = output.unsqueeze(0) #
# Map out to cell state(0,1,2,3,4)
counter = 0
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
cell_alive = (np.amax(
morphogens[1, i - 1:i + 2, j - 1:j + 2,
k - 1:k + 2])) > 0.1 #ALPHA
if cell_alive: #ALPHA
_, idx = output[0, counter].data[:5].max(0) #ALPHA
alpha = output[0, counter].data[5] #ALPHA
morph_tmp[0, i, j, k] = int(idx.data.numpy())
morph_tmp[1, i, j, k] = F.sigmoid(
alpha
) #ALPHA TODO apply function earlier, should be faster
if settings['recurrent']:
hidden_states_batched_A[0,
counter] = hs[0][0,
counter]
hidden_states_batched_B[0,
counter] = hs[1][0,
counter]
out = output
out[0] = output
else: #ALPHA
morph_tmp[:, i, j,
k] = 0 #If cell is not alive, set state to 0
if settings['recurrent']:
hidden_states_batched_A[
0, counter] = hs[0][0, counter] * 0.0
hidden_states_batched_B[
0, counter] = hs[1][0, counter] * 0.0
counter += 1
#Caliculation each iterartion cell state
morpho = morphogens[0][np.newaxis, :, :]
alpha = morphogens[1][np.newaxis, :, :]
dev_states.append(morpho)
alphalist.append(alpha)
morphogens = copy.deepcopy(morph_tmp)
#Check connection
counter = -1
while counter < 0:
counter = 0
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
if morphogens[0, i, j, k] != 0:
neighbors = np.array([
morphogens[0, i - 1, j, k], morphogens[0, i + 1, j,
k],
morphogens[0, i, j - 1, k], morphogens[0, i, j + 1,
k],
morphogens[0, i, j, k - 1], morphogens[0, i, j,
k + 1]
])
diagonal = np.array([morphogens[0, i-1, j+1, k-1], morphogens[0, i-1, j+1, k+1],morphogens[0, i-1, j+1, k], morphogens[0, i-1, j, k-1], morphogens[0, i-1, j, k+1], morphogens[0, i-1, j-1, k-1],morphogens[0, i-1, j-1, k+1],morphogens[0, i-1, j-1, k],\
morphogens[0, i, j+1, k-1], morphogens[0, i, j+1, k+1], morphogens[0, i, j-1, k-1],morphogens[0, i, j-1, k+1],\
morphogens[0, i+1, j+1, k-1], morphogens[0, i+1, j+1, k+1],morphogens[0, i+1, j+1, k], morphogens[0, i+1, j, k-1], morphogens[0, i+1, j, k+1], morphogens[0, i+1, j-1, k-1],morphogens[0, i+1, j-1, k+1],morphogens[0, i+1, j-1, k]])
if np.count_nonzero(
neighbors
) == 0 and np.count_nonzero(diagonal) != 0:
morphogens[0, i, j, k] = 0
morphogens[1, i, j, k] = 0
counter -= 1
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
if morphogens[0, i, j, k] != 0:
neighbors = np.array([
morphogens[0, i - 1, j, k], morphogens[0, i + 1, j,
k],
morphogens[0, i, j - 1, k], morphogens[0, i, j + 1,
k],
morphogens[0, i, j, k - 1], morphogens[0, i, j,
k + 1]
])
diagonal = np.array([morphogens[0, i-1, j+1, k-1], morphogens[0, i-1, j+1, k+1],morphogens[0, i-1, j+1, k], morphogens[0, i-1, j, k-1], morphogens[0, i-1, j, k+1], morphogens[0, i-1, j-1, k-1],morphogens[0, i-1, j-1, k+1],morphogens[0, i-1, j-1, k],\
morphogens[0, i, j+1, k-1], morphogens[0, i, j+1, k+1], morphogens[0, i, j-1, k-1],morphogens[0, i, j-1, k+1],\
morphogens[0, i+1, j+1, k-1], morphogens[0, i+1, j+1, k+1],morphogens[0, i+1, j+1, k], morphogens[0, i+1, j, k-1], morphogens[0, i+1, j, k+1], morphogens[0, i+1, j-1, k-1],morphogens[0, i+1, j-1, k+1],morphogens[0, i+1, j-1, k]])
all_neighbors = np.count_nonzero(
neighbors) + np.count_nonzero(diagonal)
if all_neighbors == 0:
morphogens[0, i, j, k] = 0
morphogens[1, i, j, k] = 0
cutedgemorphogens[0] = morphogens[0][1:im_size - 1, 1:im_size - 1,
1:im_size - 1]
cutedgemorphogens[1] = morphogens[1][1:im_size - 1, 1:im_size - 1,
1:im_size - 1]
# Culiculate fitness
# reshape
original_morphogens_list = np.reshape(original_morphogens, (2, im_size**3))
morphogens_list = np.reshape(morphogens, (2, im_size**3))
for i in range(im_size**3):
if original_morphogens_list[0, i] == morphogens_list[0, i]:
fitness += 1
print(original_morphogens_list[0, i], morphogens_list[0, i])
print(i, fitness)
#Return fitness
return fitness, sim, env, generations, individual_id, morphogens, hidden_states_batched_A, hidden_states_batched_B, dev_states, alphalist, cutedgemorphogens, out, past_hidden_states_batched_A, past_hidden_states_batched_B
def run_robot_simulation(params, details=False, save_name=""):
#Caliculation initial cell states
individual_id = params[0][0]
w = params[0][1]
generations = params[1] #4
settings = params[2] #5
run_directory = settings["run_directory"]
im_size = settings["im_size"]
num_classes = settings["voxel_types"]
input_dim = settings["number_neighbors"]
if settings['growth_facter']:
input_dim = 3 * 9 * 2
num_classes = settings['voxel_types'] + 1
#Make .vxa file for voxelyze
# Setting up the simulation object
sim = Sim(dt_frac=settings['fraction'],
simulation_time=settings['simtime'],
fitness_eval_init_time=settings['initime'])
# Setting up the environment object
env = Env(sticky_floor=0, time_between_traces=0)
p = MorphNet(input_dim=input_dim,
number_state=num_classes,
recurrent=settings['recurrent'])
vector_to_parameters(torch.tensor(w, dtype=torch.float32), p.parameters())
if settings['growth_facter']:
morphogens = np.zeros(shape=(2, im_size, im_size, im_size))
morphogens[:, (im_size - 1) / 2, (im_size - 1) / 2,
(im_size - 1) / 2] = 1
cutedgemorphogens = np.zeros(shape=(2, im_size - 2, im_size - 2,
im_size - 2))
else:
cutedgemorphogens = np.zeros(shape=(2, im_size - 2, im_size - 2,
im_size - 2))
morphogens = np.zeros(shape=(1, im_size, im_size, im_size))
out = torch.zeros(1, im_size * im_size * im_size, num_classes)
dev_states = []
alphalist = []
sleep_id = []
iterations = 10
hidden_dim = 64
#Store seperate LSTM hidden states for each cell
if settings['data_read']:
#morphogens = params[3]
#hidden_states_batched_A = params[4]
#hidden_states_batched_B = params[5]
hidden_states = torch.ones(
size=(im_size, im_size, 2, 1, im_size * im_size * im_size,
hidden_dim)) #batch_size = im_size*im_size
hidden_states_batched_A = torch.ones(
size=(1, im_size * im_size * im_size,
hidden_dim)) #Hidden layers, batch_size, number units
hidden_states_batched_B = torch.ones(size=(1,
im_size * im_size * im_size,
hidden_dim))
max_fitness = params[6]
else:
hidden_states = torch.ones(
size=(im_size, im_size, 2, 1, im_size * im_size * im_size,
hidden_dim)) #batch_size = im_size*im_size
hidden_states_batched_A = torch.ones(
size=(1, im_size * im_size * im_size,
hidden_dim)) #Hidden layers, batch_size, number units
hidden_states_batched_B = torch.ones(size=(1,
im_size * im_size * im_size,
hidden_dim))
max_fitness = 0
#hidden_states = torch.ones(size=(im_size, im_size, 2, 1, 1, hidden_dim)) #batch_size = im_size*im_size
morph_tmp = copy.deepcopy(morphogens)
#Insialize a fitness
#Start main loop
for it in range(iterations):
fitness = 0
batch_inputs = torch.zeros(im_size * im_size * im_size, input_dim)
counter = 0
sleep_id = []
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
if settings['growth_facter']:
cell_input = torch.flatten(
torch.tensor(morphogens[:, i - 1:i + 2,
j - 1:j + 2, k - 1:k + 2]))
else:
cell_input = torch.tensor([
morphogens[0, i - 1, j, k], morphogens[0, i + 1, j,
k],
morphogens[0, i, j + 1,
k], morphogens[0, i, j - 1,
k], morphogens[0, i, j,
k - 1],
morphogens[0, i, j, k + 1], morphogens[0, i, j, k]
])
batch_inputs[counter] = cell_input
counter += 1
output, hs = p(batch_inputs,
(hidden_states_batched_A, hidden_states_batched_B))
past_hidden_states_batched_A = hidden_states_batched_A
past_hidden_states_batched_B = hidden_states_batched_B
if not settings['recurrent']: #ALPHA
output = output.unsqueeze(0) #
# Map out to cell state(0,1,2,3,4)
counter = 0
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
cell_alive = (np.amax(
morphogens[1, i - 1:i + 2, j - 1:j + 2,
k - 1:k + 2])) > 0.1 #ALPHA
if cell_alive: #ALPHA
_, idx = output[0, counter].data[:5].max(0) #ALPHA
alpha = output[0, counter].data[5] #ALPHA
morph_tmp[0, i, j, k] = int(idx.data.numpy())
morph_tmp[1, i, j, k] = F.sigmoid(
alpha
) #ALPHA TODO apply function earlier, should be faster
if settings['recurrent']:
hidden_states_batched_A[0,
counter] = hs[0][0,
counter]
hidden_states_batched_B[0,
counter] = hs[1][0,
counter]
out = output
out[0] = output
else: #ALPHA
morph_tmp[:, i, j,
k] = 0 #If cell is not alive, set state to 0
if settings['recurrent']:
hidden_states_batched_A[
0, counter] = hs[0][0, counter] * 0.0
hidden_states_batched_B[
0, counter] = hs[1][0, counter] * 0.0
counter += 1
#Caliculation each iterartion cell state
morpho = morphogens[0][np.newaxis, :, :]
alpha = morphogens[1][np.newaxis, :, :]
dev_states.append(morpho)
alphalist.append(alpha)
morphogens = copy.deepcopy(morph_tmp)
#Check connection
counter = -1
while counter < 0:
counter = 0
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
if morphogens[0, i, j, k] != 0:
neighbors = np.array([
morphogens[0, i - 1, j, k], morphogens[0, i + 1, j,
k],
morphogens[0, i, j - 1, k], morphogens[0, i, j + 1,
k],
morphogens[0, i, j, k - 1], morphogens[0, i, j,
k + 1]
])
diagonal = np.array([morphogens[0, i-1, j+1, k-1], morphogens[0, i-1, j+1, k+1],morphogens[0, i-1, j+1, k], morphogens[0, i-1, j, k-1], morphogens[0, i-1, j, k+1], morphogens[0, i-1, j-1, k-1],morphogens[0, i-1, j-1, k+1],morphogens[0, i-1, j-1, k],\
morphogens[0, i, j+1, k-1], morphogens[0, i, j+1, k+1], morphogens[0, i, j-1, k-1],morphogens[0, i, j-1, k+1],\
morphogens[0, i+1, j+1, k-1], morphogens[0, i+1, j+1, k+1],morphogens[0, i+1, j+1, k], morphogens[0, i+1, j, k-1], morphogens[0, i+1, j, k+1], morphogens[0, i+1, j-1, k-1],morphogens[0, i+1, j-1, k+1],morphogens[0, i+1, j-1, k]])
if np.count_nonzero(
neighbors
) == 0 and np.count_nonzero(diagonal) != 0:
morphogens[0, i, j, k] = 0
morphogens[1, i, j, k] = 0
counter -= 1
for i in range(1, im_size - 1):
for j in range(1, im_size - 1):
for k in range(1, im_size - 1):
if morphogens[0, i, j, k] != 0:
neighbors = np.array([
morphogens[0, i - 1, j, k], morphogens[0, i + 1, j,
k],
morphogens[0, i, j - 1, k], morphogens[0, i, j + 1,
k],
morphogens[0, i, j, k - 1], morphogens[0, i, j,
k + 1]
])
diagonal = np.array([morphogens[0, i-1, j+1, k-1], morphogens[0, i-1, j+1, k+1],morphogens[0, i-1, j+1, k], morphogens[0, i-1, j, k-1], morphogens[0, i-1, j, k+1], morphogens[0, i-1, j-1, k-1],morphogens[0, i-1, j-1, k+1],morphogens[0, i-1, j-1, k],\
morphogens[0, i, j+1, k-1], morphogens[0, i, j+1, k+1], morphogens[0, i, j-1, k-1],morphogens[0, i, j-1, k+1],\
morphogens[0, i+1, j+1, k-1], morphogens[0, i+1, j+1, k+1],morphogens[0, i+1, j+1, k], morphogens[0, i+1, j, k-1], morphogens[0, i+1, j, k+1], morphogens[0, i+1, j-1, k-1],morphogens[0, i+1, j-1, k+1],morphogens[0, i+1, j-1, k]])
all_neighbors = np.count_nonzero(
neighbors) + np.count_nonzero(diagonal)
if all_neighbors == 0:
morphogens[0, i, j, k] = 0
morphogens[1, i, j, k] = 0
cutedgemorphogens[0] = morphogens[0][1:im_size - 1, 1:im_size - 1,
1:im_size - 1]
cutedgemorphogens[1] = morphogens[1][1:im_size - 1, 1:im_size - 1,
1:im_size - 1]
#Save vxa file and caliculate distance
write_voxelyze_file(sim, env, generations, individual_id,
cutedgemorphogens, fitness, settings['im_size'],
settings['run_directory'], settings['run_name'])
#Evaluate with voxelyze
#start_time = time.time()
filepath = settings[
'run_directory'] + "/tempFiles/softbotsOutput--gene_{0}_id_{1}.xml".format(
generations, individual_id)
if np.any(morphogens[0] != 0):
proc = sub.Popen("./voxelyze -f " + settings['run_directory'] +
"/voxelyzeFiles/" + settings['run_name'] +
"--gene_{0}_id_{1}_fitness_0.vxa".format(
generations, individual_id),
stdout=sub.PIPE,
shell=True)
proc.wait()
while os.path.exists(filepath) == False:
sleep(0.00000000001)
else:
print("gene_{0}_id_{1} is dead".format(generations, individual_id))
#Read results
#All subprocess done
if os.path.exists(filepath):
# voxels cost
voxels_cost = np.count_nonzero(cutedgemorphogens[0]) * 0.0583
fitness = read_voxlyze_results(generations, individual_id,
filepath) - voxels_cost
sub.Popen("rm " + settings['run_directory'] +
"/tempFiles/softbotsOutput--gene_{0}_id_{1}.xml".format(
generations, individual_id),
stdout=sub.PIPE,
shell=True)
else:
fitness = -20
sub.Popen(
"rm " + settings['run_directory'] + "/voxelyzeFiles/" +
settings['run_name'] +
"--gene_{0}_id_{1}_fitness_0.vxa".format(generations, individual_id),
stdout=sub.PIPE,
shell=True)
#Return fitness
return fitness, sim, env, generations, individual_id, morphogens, hidden_states_batched_A, hidden_states_batched_B, dev_states, alphalist, cutedgemorphogens, out, past_hidden_states_batched_A, past_hidden_states_batched_B
def run_robot_simulation(params, details=False, save_name=""):
#Caliculation initial cell states
individual_id = params[0][0]
w = params[0][1]
generations = params[1]
settings = params[2]
run_directory = settings["run_directory"]
im_size = settings["im_size"]
num_classes = settings["voxel_types"]
input_dim = settings["number_neighbors"]
if settings['growth_facter']:
input_dim = 9 * 2
voxel_types = settings['voxel_types'] + 1
if settings['cell_sleep']:
input_dim = input_dim + 1
num_classes = num_classes + 1
#print(input_dim, num_classes)
#Make .vxa file for voxelyze
# Setting up the simulation object
sim = Sim(dt_frac=settings['fraction'],
simulation_time=settings['simtime'],
fitness_eval_init_time=settings['initime'])
# Setting up the environment object
env = Env(sticky_floor=0, time_between_traces=0)
p = MorphNet(input_dim=input_dim,
number_state=num_classes,
recurrent=settings['recurrent'])
vector_to_parameters(torch.tensor(w, dtype=torch.float32), p.parameters())
morphogens = np.zeros(shape=(1, im_size, im_size))
morphogens[0, (im_size - 1) / 2, (im_size - 1) / 2] = 1
if settings['growth_facter']:
morphogens = np.zeros(shape=(2, im_size, im_size))
morphogens[:, (im_size - 1) / 2, (im_size - 1) / 2] = 1
#print(morphogens.shape)
morphogens_sleep = np.zeros(shape=(1, im_size, im_size))
out_sleep = torch.zeros(1, im_size * im_size, num_classes - 1)
out = torch.zeros(1, im_size * im_size, num_classes)
#print(out.shape, out[0].shape, out[0][0], out[0][counter][input_dim-1])
dev_states = []
sleep_id = []
iterations = 10
hidden_dim = 64
#Store seperate LSTM hidden states for each cell
hidden_states = torch.ones(
size=(im_size, im_size, 2, 1, im_size * im_size,
hidden_dim)) #batch_size = im_size*im_size
hidden_states_batched_A = torch.ones(
size=(1, im_size * im_size,
hidden_dim)) #Hidden layers, batch_size, number units
hidden_states_batched_B = torch.ones(
size=(1, im_size * im_size,
hidden_dim)) #Hidden layers, batch_size, number units
#hidden_states = torch.ones(size=(im_size, im_size, 2, 1, 1, hidden_dim)) #batch_size = im_size*im_size
morph_tmp = copy.deepcopy(morphogens)
#Insialize a fitness
max_fitness = 0
#Start main loop
for it in range(iterations):
fitness = 0
batch_inputs = torch.zeros(im_size * im_size, input_dim)
counter = 0
sleep_id = []
if settings['cell_sleep']:
for i in range(1, im_size - 1):
for e in range(1, im_size - 1):
if out[0][counter][num_classes -
1] <= 0 and counter not in sleep_id:
morphogens_sleep[0, i, e] = morphogens[0, i, e]
else:
#print(out[0][counter][input_dim-1])
morphogens_sleep[0, i, e] = 0
#print(batch_inputs[counter].shape,batch_inputs[counter][4],input_dim-1)
for i in range(1, im_size - 1):
for e in range(1, im_size - 1):
#batch_inputs[counter] = torch.tensor( [morphogens[0, i-1, e],morphogens[0, i+1, e],morphogens[0, i, e+1],morphogens[0, i, e-1],morphogens[0,i,e] ] )
if settings['cell_sleep']:
if out[0][counter][num_classes -
1] <= 0 and counter not in sleep_id:
cell_input = torch.tensor([
morphogens_sleep[0, i - 1,
e], morphogens_sleep[0, i + 1, e],
morphogens_sleep[0, i, e + 1],
morphogens_sleep[0, i,
e - 1], morphogens_sleep[0, i, e],
out[0][counter][num_classes - 1]
])
else:
cell_input = torch.zeros(input_dim)
elif settings['growth_facter']:
cell_input = torch.flatten(
torch.tensor(morphogens[:, i - 1:i + 2, e - 1:e + 2]))
else:
cell_input = torch.tensor([
morphogens[0, i - 1, e], morphogens[0, i + 1, e],
morphogens[0, i, e + 1], morphogens[0, i, e - 1],
morphogens[0, i, e]
])
batch_inputs[counter] = cell_input
#print(input_dim)
#print(out[0][counter])
#print(counter)
counter += 1
output, hs = p(batch_inputs,
(hidden_states_batched_A, hidden_states_batched_B))
#print(output.shape)
if not settings['recurrent']: #ALPHA
output = output.unsqueeze(
0) #So we have recurrent one and non-recurrent in same format
#print(output.shape)
#print(output[0, counter].shape, output[0, counter])
#print(out.shape, out[0].shape, output.shape, out_sleep.shape)
#print(morphogens[0])
# Map out to cell state(0,1,2,3,4)
counter = 0
for i in range(1, im_size - 1):
for e in range(1, im_size - 1):
#print(i,e)
#print(i, e, counter, int(counter/(im_size-2))+1, counter%(im_size-2)+1)
#print (counter, i,e, morphogens[0, i, e])
if settings['growth_facter']:
cell_alive = (np.amax(
morphogens[1, i - 1:i + 2, e - 1:e + 2])) > 0.1 #ALPHA
if cell_alive: #ALPHA
_, idx = output[0, counter].data[:4].max(0) #ALPHA
#print(out[0].shape, out[0, counter].shape,out[0, counter].data.shape)
#print(out[0, counter].data[:4])
alpha = output[0, counter].data[4] #ALPHA
morph_tmp[0, i, e] = int(idx.data.numpy())
morph_tmp[1, i, e] = F.relu(
alpha
) #ALPHA TODO apply function earlier, should be faster
if settings['recurrent']:
hidden_states_batched_A[0,
counter] = hs[0][0,
counter]
hidden_states_batched_B[0,
counter] = hs[1][0,
counter]
out = output
out[0] = output
else: #ALPHA
morph_tmp[:, i,
e] = 0 #If cell is not alive, set state to 0
if settings['recurrent']:
hidden_states_batched_A[
0, counter] = hs[0][0, counter] * 0.0
hidden_states_batched_B[
0, counter] = hs[1][0, counter] * 0.0
#print(counter)
counter += 1
else:
if settings['cell_sleep']:
#Make cell sleep version morphogen
if out[0][counter][num_classes -
1] <= 0 and counter not in sleep_id:
sum_cells = morphogens_sleep[
0, i - 1, e] + morphogens_sleep[
0, i + 1, e] + morphogens_sleep[
0, i, e + 1] + morphogens_sleep[
0, i,
e - 1] + morphogens_sleep[0, i, e]
else:
sum_cells = 0
sleep_id.append(counter)
sleep_id = list(set(sleep_id))
#print(sleep_id)
else:
sum_cells = morphogens[0, i - 1, e] + morphogens[
0, i + 1,
e] + morphogens[0, i, e + 1] + morphogens[
0, i, e - 1] + morphogens[0, i, e]
#print(counter, output[counter], out_sleep[0,counter])
#print(out.shape,out[0,7],output[0, counter][0:1])
if sum_cells > 0: #If any of the surrounded cells or the cell itself is activate, modify the that position
#print(out[0, counter].data)
#print(out_sleep[0,1].shape, output[0,1].shape)
#print(counter, out_sleep[counter])
# Chose the highest value state
if settings['cell_sleep']:
#print(counter, out_sleep[0, counter] )
if settings['recurrent']:
out_sleep[0, counter] = output[
0, counter][0:num_classes - 1]
#print(out_sleep[0, counter])
_, idx = out_sleep[0, counter].data.max(0)
else:
#print(out_sleep.shape, output.shape)
out_sleep[
0,
counter] = output[counter][0:num_classes -
1]
_, idx = out_sleep[0, counter].data.max(0)
else:
if settings['recurrent']:
_, idx = output[0, counter].data.max(0)
#print(_, idx)
else:
_, idx = output[counter].data.max(0)
# Next cell states
morph_tmp[0, i, e] = int(idx.data.numpy())
#print(out.shape, out_sleep.shape, output.shape)
#TODO if it's mot recurrent, "expression" would also not work right now?
#A is hidden state, B is cell state
# Update cell state and hideen states
if settings['recurrent']:
hidden_states_batched_A[0,
counter] = hs[0][0,
counter]
hidden_states_batched_B[0,
counter] = hs[1][0,
counter]
out = output
out[0] = output
#print(out[0].shape, output.shape)
#print(counter)
counter += 1
#Caliculation each iterartion cell state
morpho = morphogens[0][np.newaxis, :, :]
alpha = morphogens[1][np.newaxis, :, :]
dev_states.append(morpho)
#print(type(dev_states),len(dev_states))
morphogens = copy.deepcopy(morph_tmp)
#print(morphogens.shape)
#print(morpho.shape)
#print(type(morphogens),morphogens.shape)# morphogen[batchsize,ind_id,im_size]
#print(individual_id, morphogens,morphogens[0][4], morphogens[0][4][5])
#Save vxa file and caliculate distance
#print(fitness)
write_voxelyze_file(sim, env, generations, individual_id, morpho, fitness,
settings['im_size'], settings['run_directory'],
settings['run_name'])
#Evaluate with voxelyze
#start_time = time.time()
filepath = settings[
'run_directory'] + "/tempFiles/softbotsOutput--gene_{0}_id_{1}.xml".format(
generations, individual_id)
#print(filepath)
if np.any(morpho != 0):
proc = sub.Popen("./voxelyze -f " + settings['run_directory'] +
"/voxelyzeFiles/" + settings['run_name'] +
"--gene_{0}_id_{1}_fitness_0.vxa".format(
generations, individual_id),
stdout=sub.PIPE,
shell=True)
proc.wait()
while os.path.exists(filepath) == False:
sleep(0.0000000001)
else:
print("gene_{0}_id_{1} is dead".format(generations, individual_id))
#Read results
#All subprocess done
#print(filepath)
if os.path.exists(filepath):
fitness = read_voxlyze_results(generations, individual_id, filepath)
#print(fitness)
sub.Popen("rm " + settings['run_directory'] +
"/tempFiles/softbotsOutput--gene_{0}_id_{1}.xml".format(
generations, individual_id),
stdout=sub.PIPE,
shell=True)
else:
fitness = 0
sub.Popen(
"rm " + settings['run_directory'] + "/voxelyzeFiles/" +
settings['run_name'] +
"--gene_{0}_id_{1}_fitness_0.vxa".format(generations, individual_id),
stdout=sub.PIPE,
shell=True)
print("gene_{0}_id_{1}_fitness{2}".format(generations, individual_id,
fitness))
'''
total_log = [generations, individual_id,fitness]
str_ = str(total_log)
with open("{0}/fitnessFiles/gene_{1}_id_{2}_fitness{3}.txt".format(settings['run_directory'],generations, individual_id,fitness), 'wt') as f:
f.write(str_)
'''
#Return fitness
return fitness, sim, env, generations, individual_id, morpho, hidden_states_batched_A, hidden_states_batched_B, dev_states