def compare_champs(self):
self.epoch_len = self.hs.hist_full_size - (self.hd + 1)
r_start = self.epoch_len
champ_current = open("./champ_data/binance/latest_greatest.pkl", 'rb')
g = pickle.load(champ_current)
champ_current.close()
[cppn] = create_cppn(g, self.config, self.leaf_names, ["cppn_out"])
net_builder = ESNetwork(self.substrate, cppn, self.params)
champ_fit = self.evaluate_champ(net_builder, r_start, g, 0)
for ix, f in enumerate(os.listdir("./champ_data/binance")):
if (f != "lastest_greatest.pkl"):
champ_file = open("./champ_data/binance/" + f, 'rb')
g = pickle.load(champ_file)
champ_file.close()
[cppn] = create_cppn(g, self.config, self.leaf_names,
["cppn_out"])
net_builder = ESNetwork(self.substrate, cppn, self.params)
g.fitness = self.evaluate_champ(net_builder,
r_start,
g,
champ_num=ix)
if (g.fitness > champ_fit):
with open("./champ_data/binance/latest_greatest.pkl",
'wb') as output:
pickle.dump(g, output)
return
def make_net(genome, config, bs):
#start by setting up a substrate for this bad cartpole boi
params = {
"initial_depth": 2,
"max_depth": 4,
"variance_threshold": 0.8,
"band_threshold": 0.05,
"iteration_level": 3,
"division_threshold": 0.3,
"max_weight": 34.0,
"activation": "sigmoid"
}
input_cords = []
output_cords = [(0.0, -1.0, -1.0)]
sign = 1
# we will use a 3 dimensional substrate, coords laid out here
for i in range(4):
input_cords.append((0.0 - i / 10 * sign, 0.0, 0.0))
sign *= -1
leaf_names = []
for i in range(len(output_cords[0])):
leaf_names.append(str(i) + "_in")
leaf_names.append(str(i) + "_out")
[cppn] = create_cppn(genome, config, leaf_names, ['cppn_out'])
net_builder = ESNetwork(Substrate(input_cords, output_cords), cppn, params)
net = net_builder.create_phenotype_network_nd()
return net
def make_net(genome, config, bs):
#start by setting up a substrate for this bad cartpole boi
params = {"initial_depth": 2,
"max_depth": 4,
"variance_threshold": 0.00013,
"band_threshold": 0.00013,
"iteration_level": 3,
"division_threshold": 0.00013,
"max_weight": 3.0,
"activation": "tanh"}
input_cords = []
output_cords = [(0.0, -1.0, 0.0)]
sign = 1
for i in range(4):
input_cords.append((0.0 - i/10*sign, 1.0, 0.0))
sign *= -1
leaf_names = []
for i in range(3):
leaf_names.append('leaf_one_'+str(i))
leaf_names.append('leaf_two_'+str(i))
[cppn] = create_cppn(genome, config, leaf_names, ['cppn_out'])
# print(cppn)
print(leaf_names)
print(config.genome_config.input_keys)
net_builder = ESNetwork(Substrate(input_cords, output_cords), cppn, params)
net = net_builder.create_phenotype_network_nd()
return net
def trial_run(self):
r_start = 0
file = open("es_trade_god_cppn_3d.pkl", 'rb')
[cppn] = pickle.load(file)
network = ESNetwork(self.subStrate, cppn, self.params)
net = network.create_phenotype_network_nd()
fitness = self.evaluate(net, network, r_start)
return fitness
def eval_fitness(self, genomes, config):
self.epoch_len = 16
r_start = randint(0 + self.epoch_len, self.hs.hist_full_size - self.hd)
best_g_fit = 0.0
champ_counter = self.gen_count % 10
#img_count = 0
for idx, g in genomes:
[cppn] = create_cppn(g, config, self.leaf_names, ["cppn_out"])
net_builder = ESNetwork(self.substrate, cppn, self.params)
#cppn = neat.nn.FeedForwardNetwork.create(g, config)
#network = ESNetwork(self.subStrate, cppn, self.params, self.hd)
#net = network.create_phenotype_network_nd()
train_ft = self.evaluate(net_builder, r_start, g)
g.fitness = train_ft
if (g.fitness > best_g_fit):
best_g_fit = g.fitness
with open(
"./champ_data/kraken/latest_greatest" +
str(champ_counter) + ".pkl", 'wb') as output:
pickle.dump(g, output)
#img_count += 1
'''
if(champ_counter == 0):
self.refresh()
self.compare_champs()
'''
self.gen_count += 1
return
def validate_fitness(self):
config = self.config
genomes = neat.Checkpointer.restore_checkpoint("./pkl_pops/pop-checkpoint-27").population
self.epoch_len = 233
r_start = self.hs.hist_full_size - self.epoch_len-1
best_g_fit = 1.0
for idx in genomes:
g = genomes[idx]
cppn = neat.nn.FeedForwardNetwork.create(g, config)
network = ESNetwork(self.subStrate, cppn, self.params, self.hd)
net = network.create_phenotype_network_nd()
g.fitness = self.evaluate(net, network, r_start, g)
if(g.fitness > best_g_fit):
best_g_fit = g.fitness
with open('./champ_data/kraken/latest_greatest.pkl', 'wb') as output:
pickle.dump(g, output)
return
def make_net(genome,
config,
batch_size,
state_space_dim=111,
action_space_dim=8) -> RecurrentNet:
#start by setting up a substrate for this bad ant boi
params = {
"initial_depth": 1,
"max_depth": 5,
"variance_threshold": 0.8,
"band_threshold": 0.05,
"iteration_level": 3,
"division_threshold": 0.3,
"max_weight": 10.0,
"activation": "relu"
}
assert params[
"activation"] in piecewise_linear_activations, f"for NAS without search, network activation needs to be piecewise linear"
# FIXME this can just be list
joint_name_dict = {
0: "hip_4",
1: "ankle_4",
2: "hip_1",
3: "ankle_1",
4: "hip_2",
5: "ankle_2",
6: "hip_3",
7: "ankle_3",
}
assert len(joint_name_dict) == len(output_coords)
leaf_names = []
for i in range(len(output_coords[0])):
leaf_names.append(str(i) + "_in")
leaf_names.append(str(i) + "_out")
input_coords = get_in_coords()
[cppn] = create_cppn(genome, config, leaf_names, ['cppn_out'])
net_builder = ESNetwork(Substrate(input_coords, output_coords), cppn,
params)
net = net_builder.create_phenotype_network_nd(batch_size=batch_size)
return net # RecurrentNet ((not necessarily actually recurrent))
def make_net(genome, config, bs):
#start by setting up a substrate for this bad cartpole boi
params = {
"initial_depth": 1,
"max_depth": 2,
"variance_threshold": 0.8,
"band_threshold": 0.13,
"iteration_level": 3,
"division_threshold": 0.3,
"max_weight": 3.0,
"activation": "relu"
}
input_cords = [(0.1, 0.1, 0.1), (-0.1, -0.1, -0.1)]
output_cords = [(0.0, 1.0, -1.0), (0.0, 0.0, -1.0), (0.0, -1.0, -1.0)]
leaf_names = []
for i in range(len(output_cords[0])):
leaf_names.append(str(i) + "_in")
leaf_names.append(str(i) + "_out")
[cppn] = create_cppn(genome, config, leaf_names, ['cppn_out'])
net_builder = ESNetwork(Substrate(input_cords, output_cords), cppn, params)
net = net_builder.create_phenotype_network_nd('./genome_vis')
return net
def make_net(genome, config, _batch_size):
params = {
"initial_depth": 1,
"max_depth": 2,
"variance_threshold": 0.8,
"band_threshold": 0.05,
"iteration_level": 3,
"division_threshold": 0.3,
"max_weight": 30.0,
"activation": "tanh"
}
input_cords = [(-1.0, 1.0, 0.0), (-.5, 1.0, 0.0), (.5, 1.0, 0.0),
(1.0, 1.0, 0.0)]
output_cords = [(-1.0, -1.0, -1.0), (0.0, -1.0, -1.0), (1.0, -1.0, -1.0)]
leaf_names = []
for i in range(len(output_cords[0])):
leaf_names.append(str(i) + "_in")
leaf_names.append(str(i) + "_out")
[cppn] = create_cppn(genome, config, leaf_names, ['cppn_out'])
net_builder = ESNetwork(Substrate(input_cords, output_cords), cppn, params)
#net = net_builder.create_phenotype_network_nd('./genome_vis')
return net_builder