def test_agent_query(self):
ag1 = model.Agent('1', 'agent1', '127.0.0.1', datetime.now())
ag1.save()
ag2 = model.Agent('2', 'agent2', '127.0.0.2', datetime.now())
ag2.save()
ag3 = model.Agent('3', 'agent3', '127.0.0.3', datetime.now())
ag3.save()
agents = model.Agent.query()
self.assertEqual(3, len(agents))
agents = model.Agent.query(where='aid=?', params=['1'])
self.assertEqual(1, len(agents))
self.assertEqual(ag1, agents[0])
agents = model.Agent.query(orderby='aid ASC')
self.assertEqual(3, len(agents))
self.assertEqual(ag1, agents[0])
self.assertEqual(ag2, agents[1])
self.assertEqual(ag3, agents[2])
agents = model.Agent.query(orderby='aid ASC', offset=1, limit=2)
self.assertEqual(2, len(agents))
self.assertEqual(ag2, agents[0])
self.assertEqual(ag3, agents[1])
def test_agent_count(self):
ag1 = model.Agent('1', 'agent1', '127.0.0.1', datetime.now())
ag2 = model.Agent('2', 'agent2', '127.0.0.1', datetime.now())
ag3 = model.Agent('3', 'agent3', '127.0.0.1', datetime.now())
model.Agent.save_all([ag1, ag2, ag3])
self.assertEqual(3, model.Agent.count())
self.assertEqual(1, model.Agent.count(where='aid=?', params=['1']))
self.assertEqual(2, model.Agent.count(where='aid>?', params=['1']))
self.assertEqual(0, model.Agent.count(where='name=?', params=['1']))
def test_remove_agent(self):
ag = model.Agent('12345678', 'agent1', '127.0.0.1', datetime.now())
ag.save()
agents = model.Agent.query()
self.assertEqual(1, len(agents))
self.assertEqual(ag, agents[0])
ag.remove()
agents = model.Agent.query()
self.assertEqual(0, len(agents))
def test_add_agent(self):
aglist = model.Agent.query()
self.assertEqual(0, len(aglist))
ag = model.Agent('12345678', 'agent1', '127.0.0.1', datetime.now())
ag.save()
aglist = model.Agent.query()
self.assertEqual(1, len(aglist))
self.assertEqual(ag, aglist[0])
def test_update_agent(self):
ag = model.Agent('12345678', 'agent1', '127.0.0.1', datetime.now())
ag.save()
ag.set(last_cpu_util=99,
last_mem_util=55.5,
last_sys_load1=1.1,
last_sys_cs=123)
aglist = model.Agent.query()
self.assertEqual(1, len(aglist))
nag = aglist[0]
self.assertEqual(99, nag.last_cpu_util)
self.assertEqual(55.5, nag.last_mem_util)
self.assertEqual(1.1, nag.last_sys_load1)
self.assertEqual(123, nag.last_sys_cs)
def _agent_reg(self, msg):
agent = self.find_agent(msg.agentid)
ahostname = msg.body['hostname']
aname = ahostname + '@' + msg.body['os']
if agent:
agent.set(name=aname)
logging.info('activate existing agent %s', agent)
# TODO activation
else:
agent = model.Agent(msg.agentid,
aname,
ahostname,
create_at=datetime.utcnow())
agent.save()
logging.info('new agent %s registered', agent)
self._agents[agent.aid] = agent
return True
is_renforcement = True
is_heuristique = False
is_ai = True
train = True
agent = None
if is_renforcement:
agent = model.reload_agent_renforcement(path, reload_model,
reload_nb_tentative)
if reload_model and is_renforcement:
nb_tentative = model.get_nb_tentative(path)
agent = model.Agent(INPUT_LEN, 4, nb_tentative)
agent.load_model(path + "DDQN_model_cp.h5")
else:
agent = model.Agent(INPUT_LEN, 4)
max_frame = 1200
t = 0
while train:
t = t + 1
pygame.event.get()
running, score_update, checkpoint = game_loop(screen,
clock,
car,
print('Initializing OpenAI environment...')
if args.fskip % 2 == 0 and args.env == 'SpaceInvaders-v0':
print("SpaceInvaders needs odd frameskip due to bullet alternations")
args.fskip = args.fskip -1
envs = MultiEnvironment(args.env, args.batch_size, args.fskip)
action_size = envs.get_action_size()
print('Building models...')
torch.cuda.set_device(args.gpu)
if not (os.path.isfile(args.agent_file) and os.path.isfile(args.agent_file) and os.path.isfile(args.agent_file)):
print("need an agent file")
exit()
args.agent_file = args.env + ".model.80.tar"
agent = model.Agent(action_size).cuda()
agent.load_state_dict(torch.load(args.agent_file, map_location=map_loc))
Z_dim = args.latent
encoder = model.Encoder(Z_dim).cuda()
generator = model.Generator(Z_dim).cuda()
encoder.train()
generator.train()
optim_gen = optim.Adam(generator.parameters(), lr=args.lr, betas=(0.0,0.9))
optim_enc = optim.Adam(filter(lambda p: p.requires_grad, encoder.parameters()), lr=args.lr, betas=(0.0,0.9))
print('finished building model')
import environment
import model
import torch
from itertools import count
from settings import DEVICE, SCREEN_WIDTH, TARGET_UPDATE, EPOCHS
# Environment Setup
world = environment.init()
world.reset()
# Log how long our agent lasts for each iteration
durations = []
# Training
agent = model.Agent(DEVICE)
for i in range(EPOCHS):
# Initialize environment
world.reset()
# Get current state
last_screen = environment.get_screen(world, SCREEN_WIDTH, DEVICE)
current_screen = environment.get_screen(world, SCREEN_WIDTH, DEVICE)
state = current_screen - last_screen
for t in count():
# Select and perform an action
action = agent.select_action(state)
_, reward, done, _ = world.step(action.item())
reward = torch.tensor([reward], device=DEVICE)
print('Initializing OpenAI environment...')
if args.fskip % 2 == 0 and args.env == 'SpaceInvaders-v0':
print("SpaceInvaders needs odd frameskip due to bullet alternations")
args.fskip = args.fskip -1
envs = MultiEnvironment(args.env, args.batch_size, args.fskip)
action_size = envs.get_action_size()
print('Building models...')
torch.cuda.set_device(args.gpu)
if not (os.path.isfile(args.agent_file) and os.path.isfile(args.agent_file) and os.path.isfile(args.agent_file)):
print("need an agent file")
exit()
args.agent_file = args.env + ".model.80.tar"
agent = model.Agent(action_size, args.agent_latent).cuda()
agent.load_state_dict(torch.load(args.agent_file, map_location=map_loc))
Z_dim = args.latent
wae_z_dim = args.wae_latent
encoder = model.Encoder(Z_dim).cuda()
generator = model.Generator(Z_dim, action_size).cuda()
discriminator = model.Discriminator(Z_dim, action_size).cuda()
Q = model.Q_net(args.wae_latent).cuda()
P = model.P_net(args.wae_latent).cuda()
Q.load_state_dict(torch.load(args.Q, map_location=map_loc))
P.load_state_dict(torch.load(args.P, map_location=map_loc))
encoder.train()
from tqdm import tqdm
from multiprocessing import Process, Lock, Manager
import model
import helper
#Remove tensorflow initialization messages
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)
with tf.Session(config=tf.ConfigProto(device_count={'GPU': 0})) as sess:
pass
#Create Agent
agent = model.Agent(
6, 64, 0.001, "epAgent"
) #Model Architecture set here-- (blocks, filters/block, l2reg)
#Load in from previous run
constant = 0 #Set this constant to the iteration you want to start on (default for brand new run)
"""
agent.model = helper.load_model("bin/model.json", "bin/modelCycle30.h5")
agent.running_model = helper.load_model("bin/model.json", "bin/modelRunAv30.h5")
agent.elo = 720
agent.train_count = constant
"""
#Create chess board and initialize stockfish
board = chess.Board()
stockfish = chess.uci.popen_engine("stockfish")
info_handler = chess.uci.InfoHandler()
for participant_i in range(args.participant_i, args.participant_i + args.participants_per_job):
# Get participant data_in and data_out
with open(os.path.join(args.path, 'data', str(participant_i)), 'r') as file:
data_in, data_out, participant_id, initial_hypothesis = file.read().strip().split('\n')
# If an initial hypothesis has been set and you're using the
# simplicity prior, construct the initial hypothesis array.
# This is a hack to get MH to converge faster in cases where
# we expect the agent will produce large categories (which can
# be very time consuming to rectangularize).
if args.prior == 'simplicity':
initial_hypothesis = eval(initial_hypothesis)
else:
initial_hypothesis = None
if initial_hypothesis is not None:
initial_hypothesis = np.full(shape, initial_hypothesis, dtype=int)
# Get model agent to infer a language based on participant's
# data_in and then measure likelihood of participant's data_out
# given agent's inferred language.
agent = model.Agent(shape, maxcats, args.prior, args.weight, args.noise, exposures, mcmc_iterations)
agent.learn(eval(data_in), initial_hypothesis)
lhood = agent._likelihood(eval(data_out), agent.language)
# Write out weight, noise, measured likelihood, and inferred
# language to file for collection by collect.py
with open(os.path.join(args.path, 'cand', str(args.iteration), str(participant_i)), 'w') as file:
file.write('\n'.join(map(str, ([args.weight, args.noise, lhood, agent.language.flatten().tolist()]))))
def step(self):
# Iterate over all the agents
for pos, agent in self.agents.items():
# Steps are not used yet
self.steps += 1
#if agent exists
if self.agents[pos]:
agent.step()
# Get the agents that need to be updated in the dictionary
agents_to_be_updated = []
for pos, agent in self.agents.items():
if agent.update is True:
agents_to_be_updated.append(agent)
# Update the dictionary
for agent in agents_to_be_updated:
self.remove_by_key(agent.old_pos)
self.add(agent)
agent.update = False
# Prevent 0-agent crash
if len(self.agents) == 0:
print("Number of agents reached zero, exiting..")
sys.exit()
# Add plotting information:
self.model.happy_plot.append(float((self.model.happy)/len(self.agents)))
self.model.moves_plot.append(self.model.moves)
self.model.deaths_plot.append(self.model.deaths)
self.model.total_agents.append(len(self.agents))
self.model.adult_agents.append((sum(agent.type == 1 for agent in self.agents.values())))
self.model.young_agents.append((sum(agent.type == 0 for agent in self.agents.values())))
self.model.elderly_agents.append((sum(agent.type == 2 for agent in self.agents.values())))
self.model.similar_neighbors.append(sum(agent.similar for agent in self.agents.values())/len(self.agents))
# Remove agents that are too old from grid
agents_removed = {pos: agent for pos, agent in self.agents.items() if agent.destroy is True}
for pos, agent in agents_removed.items():
self.model.grid.remove_agent(agent)
# Remove agents that are too old from the agents dictionary
#self.agents = {pos: agent for pos, agent in self.agents.items() if agent.destroy is False}
# Get the agents that need to be deleted
to_delete = []
for pos, agent in self.agents.items():
if agent.destroy is True:
to_delete.append(agent.pos)
# Delete the agents from the dictionary
for pos in to_delete:
agent = self.agents[pos]
self.remove(agent)
# Let adults reproduce with a certain percentage
# NOTE: Rather not have this in this step function, but cannot mutate
# dictionary while iterating over it (if we put this function in Agent class)
counter = 0
for i in range(sum(agent.type == 1 for agent in self.agents.values())):
if random.random() <= self.model.reproduction:
if float(len(self.model.grid.empty_spots)) >= 0.05*self.model.width*self.model.height:
counter += 1
# Create agents, place them on the grid and add them to the scheduler
agent = model.Agent((None, None), self.model, agent_type=0, age=0)
self.model.grid.place_agent_on_empty(agent)
self.add(agent)
# Add the number of births to the model for this epoch
self.model.births_plot.append(counter)
if not self.agents:
self.model.running = False
def main():
#load models
#load up an atari game
#run (and save) every frame of the game
#args = parse_args()
args = parse_args()
if args.missing == "none":
args.seed = 45
if args.missing == "agent":
args.seed = 13
MAX_ITERS = args.iters
speed = args.speed
frames_to_cf = args.frames_to_cf
seed = args.seed
img_dir = args.img_dir
img_dir = os.path.join(img_dir, "imgs_{}_{}".format(args.enc_file[:-7]+args.enc_file[-2:], args.agent_file[-7:]))
if args.enc_file == None or args.gen_file == None:
print("Need to load models for the gen and enc")
exit()
if not os.path.isfile(args.agent_file):
args.agent_file = args.env + ".model.80.tar"
if not os.path.isfile(args.agent_file):
print("bad agent_file")
exit()
map_loc = {
'cuda:0': 'cuda:'+str(args.gpu),
'cuda:1': 'cuda:'+str(args.gpu),
'cuda:2': 'cuda:'+str(args.gpu),
'cuda:3': 'cuda:'+str(args.gpu),
'cuda:4': 'cuda:'+str(args.gpu),
'cuda:5': 'cuda:'+str(args.gpu),
'cuda:7': 'cuda:'+str(args.gpu),
'cuda:6': 'cuda:'+str(args.gpu),
'cuda:8': 'cuda:'+str(args.gpu),
'cuda:9': 'cuda:'+str(args.gpu),
'cuda:10': 'cuda:'+str(args.gpu),
'cuda:11': 'cuda:'+str(args.gpu),
'cuda:12': 'cuda:'+str(args.gpu),
'cuda:13': 'cuda:'+str(args.gpu),
'cuda:14': 'cuda:'+str(args.gpu),
'cuda:15': 'cuda:'+str(args.gpu),
'cpu': 'cpu',
}
if args.frame_skip % 2 ==0 and args.env == 'SpaceInvaders-v0':
print("SpaceInvaders needs odd frameskip due to bullet alternations")
args.frame_skip = args.frame_skip - 1
#run every model on all frames (4*n frames))
print('Loading model...')
torch.cuda.set_device(args.gpu)
torch.manual_seed(args.seed)
#number of updates to discriminator for every update to generator
envs = MultiEnvironment(args.env, 1, args.frame_skip)
agent = model.Agent(envs.get_action_size(), args.agent_latent).cuda() #cuda is fine here cause we are just using it for perceptual loss and copying to discrim
agent.load_state_dict(torch.load(args.agent_file))
encoder = model.Encoder(args.latent).cuda()
generator = model.Generator(args.latent, envs.get_action_size()).cuda()
Q = model.Q_net(args.wae_latent).cuda()
P = model.P_net(args.wae_latent).cuda()
Q.load_state_dict(torch.load(args.Q, map_location=map_loc))
P.load_state_dict(torch.load(args.P, map_location=map_loc))
encoder.load_state_dict(torch.load(args.enc_file, map_location=map_loc))
generator.load_state_dict(torch.load(args.gen_file, map_location=map_loc))
encoder.eval()
generator.eval()
Q.eval()
P.eval()
os.makedirs(img_dir, exist_ok=True)
'''def rm_file(filename):
if os.path.exists(filename): os.remove(os.path.join(img_dir,filename))
rm_file('log.txt')
rm_file('probabilities.csv')
rm_file('every_action.csv')
#printlog('frame, entropy, max_less_avg, advantage_max_less_avg, l2_z, first_step, action, distance, iterations, Q[a], avg_step_size, l2_mag_cf, p_delta_start_action, p_delta_target_action, gold_delta_start_action, gold_delta_target_action', 'every_action.csv')
printlog("frame, probabilty_array, entropy, Qs, average(Qs), advantage_max_less_avg", img_dir, 'probabilities.csv')
'''
print('finished loading models')
run_game(encoder, generator, agent, Q, P, envs, seed, img_dir, args.missing, frames_to_cf, speed, MAX_ITERS, args.cf_all_actions, args.salient_intensity, args.last_frame_diff)
