def load(fname, gen_size, new_random, data, mutation_r=.1):
load_brain = torch.load(fname)['model']
p1 = car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle)
p1.brain.load_state_dict(load_brain.state_dict().copy())
agents = [car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle) for _ in range(gen_size)]
agents[0] = p1
for i in range(1, gen_size-new_random):
agents[i].brain.load_state_dict(p1.brain.state_dict().copy())
mutate(agents[i].brain, mutation_r)
return agents
def new_gen(agents, data, gen_size = 1, mutation_r=.05, new_random=0):
agents.sort(key=lambda x: x.score, reverse=True)
best_brain = [agents[0].brain, agents[0].score]
new_agents = []
made = 0
for i in range(int((gen_size-new_random)/10)):
made += 1
new_agent = car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle)
new_agent.brain = agents[i].brain
new_agents.append(new_agent)
for j in range(int((gen_size-new_random-int((gen_size-new_random)/10))/10)):
made += 1
new_agent = car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle)
new_agent.brain.load_state_dict(agents[i].brain.state_dict().copy())
mutate(new_agent.brain, mutation_r)
new_agents.append(new_agent)
for i in range(gen_size - made):
new_agents.append(car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle))
return new_agents, best_brain
def reproduce(p1, p2, data, mutation_r=0):
new_dict = {}
w1 = p1.brain.state_dict()
w2 = p2.brain.state_dict()
for layer in w1:
l1 = w1[layer].view(-1)
l2 = w2[layer].view(-1)
new_dict[layer] = torch.zeros(len(l1))
for i in range(len(l1)):
if np.random.rand() < mutation_r:
new_dict[layer][i] = random.randrange(1000)/1000
else:
#new_dict[layer][i] = random.choice([l1[i],l2[i]]) * (1 + random.randrange(2/mutation_r)*mutation_r - mutation_r)
new_dict[layer][i] = (l1[i] + l2[i])/2
new_dict[layer] = new_dict[layer].view(w1[layer].shape)
ret = car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle)
ret.brain.load_state_dict(new_dict)
return ret
'_') + "/"
if not os.path.exists(brain_dir):
os.mkdir(brain_dir)
gameDisplay = pygame.display.set_mode((display_width, display_height))
clock = pygame.time.Clock()
pygame.display.set_caption('Car')
font = pygame.font.SysFont('arialblack', 18)
show_queue = [] # Queue of car state_dicts to show
crashed = False
running_time = 0.0
batch_floor = 0
batch_roof = batch_size
gen = 0
best_brain = (car_ai_instance().brain, 0.0)
data = dataloader([roads.road2, roads.road3], 50)
#agents = [car_ai_instance(pos=data.get_data().pos, angle=data.get_data().angle) for i in range(gen_size)]
agents = load('brains/gen45s-1238393',
gen_size,
new_random,
data,
mutation_r=.1)
key_pressed = {
'right': False,
'left': False,
'up': False,
'down': False,
'space': False
font = pygame.font.SysFont('arialblack', 18)
crashed = False
running_time = 0.0
gen = 0
data = dataloader([roads.road2, roads.road3], 50)
files = os.listdir(dir)
def get_gen(x):
s = x.split('s')[0]
return int(s[3:])
files.sort(key=get_gen)
agent = car_ai_instance(pos=data.get_data().pos,
angle=data.get_data().angle)
agent.brain = torch.load(dir + '/' + files[0])['model']
print('Showing:', files[0])
key_pressed = {
'right': False,
'left': False,
'up': False,
'down': False,
'space': False
}
while not crashed:
gameDisplay.fill(gray)
agent.tick(clock.get_time() / 1000,