def intermediate_gen(self):
for n in range(self.pop.count):
num = random.random()
for i in range(self.pop.count):
if num <= self.CDF[i]:
keep = i
break
tmp_agent = agents.agent()
tmp_agent.load_gene(self.pop.strings[keep].gene[:])
self.intgen.append(tmp_agent)
def mutation_random(self): # randomly mutates swapping two locations
for i in range(self.pop.count):
temp = agents.agent()
s = 0
num = random.random()
if num <= self.Pm:
first = random.randint(0, self.pop.length - 1)
second = random.randint(0, self.pop.length - 1)
temp.gene[:] = self.newpop[i].gene[:]
s = temp.gene[first]
temp.gene[first] = temp.gene[second]
temp.gene[second] = s
self.newpop[i].gene[:] = temp.gene[:]
def createAgent(self, agentType, strategy = 'TFT'): if agentType == 'Learner': iterations_exploring = 1000 T = pow(.01, 1.0/iterations_exploring) alpha = .5 gamma = .95 numHistoryMoves = random.choice([1,2,3]) numHistoryMoves = 1 agentx = agent(numHistoryMoves, alpha, gamma, T, self.payoffMatrix) agentx.sneakStrategy = strategy elif agentType == 'Meta': agentx = metaAgent(self.learntPolicies, self.payoffMatrix) return agentx
def initialization(self):
for n in range(self.count):
string = agents.agent()
for i in range(self.length):
if self.binary:
string.gene.append(random.randint(0, 1))
else:
while len(string.gene) < self.length:
temp = random.randint(1, self.length)
if string.gene.count(temp) == 0:
string.gene.append(temp)
# just for testing (assign random values simulating a run of the game)
# string.moves = random.randint(4,21)
# string.result = random.randint(2,4) / 2.0
self.strings.append(string)
def mutation_improve(self): # only mutates if the swapping will produce a better fitness
for i in range(self.pop.count):
changed = False
count = 0
highest = 0
temp = agents.agent()
s = 0
num = random.random()
if num <= self.Pm:
while count < 10 and not changed:
first = random.randint(0, self.pop.length - 1)
second = random.randint(0, self.pop.length - 1)
temp.gene[:] = self.newpop[i].gene[:]
s = temp.gene[first]
temp.gene[first] = temp.gene[second]
temp.gene[second] = s
highest = self.newpop[i].fitness(self.pop.length)
if temp.fitness(self.pop.length) > highest:
self.newpop[i].gene[:] = temp.gene[:]
changed = True
count += 1
def runOneToOne():
stra = 'TFT'
opponentx = strategy(stra)
iterations_exploring = 5000
T = pow(.01, 1.0/iterations_exploring)
alpha = .5
gamma = .95
numHistoryMoves = 2
payoffMatrix = {"DD":1, "CC":3, "CD":0, "DC":5}
agentx = agent(numHistoryMoves, alpha, gamma, T, payoffMatrix)
for i in range(5000):
agentsMove = agentx.agentsMove(i)
responseMove = opponentx.next()
agentx.updateAgent(agentsMove, responseMove)
opponentx.addAgentsMove(agentsMove)
agentx.interpretResults()
iterations, diffs = zip(*agentx.agentsConvergence)
import agents test = agents.agent() for i in range(1,10) : test.gene.append(i) test.fitness(10)
def train_fn(X_train_shards,
Y_train_shards,
X_test,
Y_test,
return_dict,
mal_data_X=None,
mal_data_Y=None):
# Start the training process
num_agents_per_time = int(args.C * args.k)
simul_agents = gv.num_gpus * gv.max_agents_per_gpu
simul_num = min(num_agents_per_time, simul_agents)
alpha_i = 1.0 / args.k
agent_indices = np.arange(args.k)
if args.mal:
mal_agent_index = gv.mal_agent_index
guidance = {
'gradient': (np.ones(args.k) * alpha_i, np.ones(args.k)),
'data': (np.ones(args.k) * alpha_i, np.zeros(args.k))
}
unupated_frac = (args.k - num_agents_per_time) / float(args.k)
t = 0
mal_visible = []
eval_loss_list = []
loss_track_list = []
lr = args.eta
loss_count = 0
if args.gar == 'krum':
krum_select_indices = []
while return_dict['eval_success'] < gv.max_acc and t < args.T:
print('Time step %s' % t)
process_list = []
mal_active = 0
# curr_agents = np.random.choice(agent_indices, num_agents_per_time,
# replace=False)
# if t == 0:
# curr_agents = np.random.choice(agent_indices, num_agents_per_time, replace=False)
# else:
curr_agents = np.random.choice(
agent_indices,
num_agents_per_time,
p=np.exp(guidance['gradient'][0] / guidance['gradient'][1]) /
np.sum(np.exp(guidance['gradient'][0] / guidance['gradient'][1]),
axis=0),
replace=False)
print('Set of agents chosen: %s' % curr_agents)
if t == 0:
if 'MNIST' in args.dataset:
global_model = model_mnist(type=args.model_num)
elif args.dataset == 'CIFAR-10':
global_model = cifar_10_model()
elif args.dataset == 'census':
global_model = census_model_1()
global_weights = global_model.get_weights()
np.save(gv.dir_name + 'global_weights_t%s.npy' % t, global_weights)
else:
global_weights = np.load(gv.dir_name +
'global_weights_t%s.npy' % t,
allow_pickle=True)
k = 0
agents_left = 1e4
while k < num_agents_per_time:
# true_simul = min(simul_num,agents_left)
true_simul = 1
print('training %s agents' % true_simul)
for l in range(true_simul):
gpu_index = int(l / gv.max_agents_per_gpu)
gpu_id = gv.gpu_ids[gpu_index]
i = curr_agents[k]
if args.mal is False or i != mal_agent_index:
# p = Process(target=agent, args=(i, X_train_shards[i],
# Y_train_shards[i], t, gpu_id, return_dict, X_test, Y_test,lr))
agent(i, X_train_shards[i], Y_train_shards[i], t, gpu_id,
return_dict, X_test, Y_test, lr)
elif args.mal is True and i == mal_agent_index:
# p = Process(target=mal_agent, args=(X_train_shards[mal_agent_index],
# Y_train_shards[mal_agent_index], mal_data_X, mal_data_Y, t, gpu_id, return_dict, mal_visible, X_test, Y_test))
mal_agent(X_train_shards[mal_agent_index],
Y_train_shards[mal_agent_index], mal_data_X,
mal_data_Y, t, gpu_id, return_dict, mal_visible,
X_test, Y_test)
mal_active = 1
# print (return_dict[str(i)])
guidance['gradient'][0][i] += np.sqrt(
np.array([(x**2).mean()
for x in return_dict[str(i)]]).mean())
guidance['gradient'][1][i] += 1
# p.start()
# process_list.append(p)
k += 1
# for item in process_list:
# item.join()
agents_left = num_agents_per_time - k
print('Agents left:%s' % agents_left)
if mal_active == 1:
mal_visible.append(t)
print('Joined all processes for time step %s' % t)
if 'avg' in args.gar:
if args.mal:
count = 0
for k in range(num_agents_per_time):
if curr_agents[k] != mal_agent_index:
if count == 0:
ben_delta = alpha_i * return_dict[str(
curr_agents[k])]
np.save(gv.dir_name + 'ben_delta_sample%s.npy' % t,
return_dict[str(curr_agents[k])])
count += 1
else:
ben_delta += alpha_i * return_dict[str(
curr_agents[k])]
np.save(gv.dir_name + 'ben_delta_t%s.npy' % t, ben_delta)
global_weights += alpha_i * return_dict[str(mal_agent_index)]
global_weights += ben_delta
else:
for k in range(num_agents_per_time):
global_weights += alpha_i * return_dict[str(
curr_agents[k])]
elif 'krum' in args.gar:
collated_weights = []
collated_bias = []
agg_num = int(num_agents_per_time - 1 - 2)
for k in range(num_agents_per_time):
# weights_curr, bias_curr = collate_weights(return_dict[str(curr_agents[k])])
weights_curr, bias_curr = collate_weights(return_dict[str(k)])
collated_weights.append(weights_curr)
collated_bias.append(collated_bias)
score_array = np.zeros(num_agents_per_time)
for k in range(num_agents_per_time):
dists = []
for i in range(num_agents_per_time):
if i == k:
continue
else:
dists.append(
np.linalg.norm(collated_weights[k] -
collated_weights[i]))
dists = np.sort(np.array(dists))
dists_subset = dists[:agg_num]
score_array[k] = np.sum(dists_subset)
print(score_array)
krum_index = np.argmin(score_array)
print(krum_index)
global_weights += return_dict[str(krum_index)]
if krum_index == mal_agent_index:
krum_select_indices.append(t)
elif 'coomed' in args.gar:
# Fix for mean aggregation first!
weight_tuple_0 = return_dict[str(curr_agents[0])]
weights_0, bias_0 = collate_weights(weight_tuple_0)
weights_array = np.zeros((num_agents_per_time, len(weights_0)))
bias_array = np.zeros((num_agents_per_time, len(bias_0)))
# collated_weights = []
# collated_bias = []
for k in range(num_agents_per_time):
weight_tuple = return_dict[str(curr_agents[k])]
weights_curr, bias_curr = collate_weights(weight_tuple)
weights_array[k, :] = weights_curr
bias_array[k, :] = bias_curr
shape_size = model_shape_size(weight_tuple)
# weights_array = np.reshape(np.array(collated_weights),(len(weights_curr),num_agents_per_time))
# bias_array = np.reshape(np.array(collated_bias),(len(bias_curr),num_agents_per_time))
med_weights = np.median(weights_array, axis=0)
med_bias = np.median(bias_array, axis=0)
num_layers = len(shape_size[0])
update_list = []
w_count = 0
b_count = 0
for i in range(num_layers):
weights_length = shape_size[2][i]
update_list.append(med_weights[w_count:w_count +
weights_length].reshape(
shape_size[0][i]))
w_count += weights_length
bias_length = shape_size[3][i]
update_list.append(med_bias[b_count:b_count +
bias_length].reshape(
shape_size[1][i]))
b_count += bias_length
assert model_shape_size(update_list) == shape_size
global_weights += update_list
# Saving for the next update
np.save(gv.dir_name + 'global_weights_t%s.npy' % (t + 1),
global_weights)
# Evaluate global weight
if args.mal:
# p_eval = Process(target=eval_func, args=(
# X_test, Y_test, t + 1, return_dict, mal_data_X, mal_data_Y), kwargs={'global_weights': global_weights})
eval_func(X_test,
Y_test,
t + 1,
return_dict,
mal_data_X,
mal_data_Y,
global_weights=global_weights)
else:
# p_eval = Process(target=eval_func, args=(
# X_test, Y_test, t + 1, return_dict), kwargs={'global_weights': global_weights})
eval_func(X_test,
Y_test,
t + 1,
return_dict,
global_weights=global_weights)
# p_eval.start()
# p_eval.join()
eval_loss_list.append(return_dict['eval_loss'])
t += 1
return t
validation_steps=valid_steps,
callbacks=[check_point, tensorboard])
'''
prune on vgg16, the layer index can be found by model.summary()
'''
state = 0
base_model = vgg16
base_model.evaluate(valid_gen, steps=valid_steps)
new_model = Model(inputs=vgg_model.input, outputs=predict)
layer_index = 17
action_cnt = base_model.layers[layer_index].get_weights()[0].shape[-1]
h = base_model.layers[layer_index].get_weights()[0].shape[:3]
print("layer {} have {} filters".format(layer_index, action_cnt))
ag = agent((action_cnt, h[0] * h[1] * h[2], 1),
action_cnt,
lr=1e-3,
factor=0.01)
env = environment(base_model,
new_model,
x_train,
y_train,
x_test,
y_test,
steps_per_epoch,
valid_steps,
batch_size=batch_size,
b=0.05,
layer_index=layer_index,
epochs=1)
episode_len = 5
writer = tf.summary.create_file_writer("./logs/mylogs_7")
epsilon = 0.1
gamma = 0.9
target_replace_iter = 100
memory_capacity = 1000
n_actions = 7
n_states = 42
'''Runner'''
class runner():
def __init__(self, env: environment, mode: str, total_episode: int):
self.env = env
self.mode = mode
self.total_episode = total_episode
def start(self):
if self.mode == 'train':
self.env.train(self.total_episode)
self.env.test(trained_agent="none", total_episode_trained=self.total_episode)
self.env.test(trained_agent="first_mover", total_episode_trained=self.total_episode)
self.env.test(trained_agent="second_mover", total_episode_trained=self.total_episode)
self.env.test(trained_agent="both agents", total_episode_trained=self.total_episode)
self.env.com_as_player(com_as='first_mover', total_episode_trained=self.total_episode)
self.env.com_as_player(com_as='second_mover', total_episode_trained=self.total_episode)
if __name__ == "__main__":
first_mover = agent(1, total_episode, epsilon, learning_rate, gamma, batch_size, target_replace_iter, memory_capacity, n_actions, n_states)
second_mover = agent(-1, total_episode, epsilon, learning_rate, gamma, batch_size, target_replace_iter, memory_capacity, n_actions, n_states)
env = environment(first_mover, second_mover)
runner = runner(env, mode, total_episode)
runner.start()
agentList = []
wallList = []
maplist = "pacman-large.txt", "huge.txt", "complex2.txt", "complex.txt", "default.txt", "empty-large.txt", "pacman.txt", "test.txt"
i = random.randint(0, 7)
chosenMap = "maps/" + maplist[i]
c_map = maps.Map(chosenMap,
c_gameType=random.randint(0, 1)) #"maps/complex2.txt")
for i in range(0, 10):
r = random.randint(0, 10)
if r == 0:
_role = "runner"
agentList.append(
agents.agent(c_map=c_map,
c_agent_list=agentList,
c_alg="Reflex",
_role=_role,
_index=(len(agentList)),
_rand=2))
elif r >= 1 and r <= 3:
_role = "runner"
agentList.append(
agents.agent(c_map=c_map,
c_agent_list=agentList,
c_alg="BFS",
_role=_role,
_index=(len(agentList))))
elif r > 3 and r < 6:
_role = "runner"
agentList.append(
agents.agent(c_map=c_map,
c_agent_list=agentList,
from crypto_fitness import *
import agents
temp = agents.agent()
temp.load_gene([22, 10, 16, 25, 20, 18, 2, 26, 11, 3, 19, 8, 24, 4, 23, 7, 14, 1, 9, 12, 15, 5, 17, 13, 6, 21])
print fitness(temp ,26)
fhCrypt = open("crypto.txt")
newAlpha = convert(temp.gene)
print flip(fhCrypt.readline(),newAlpha)
for x in range(1, nooflocaion + 1):
location_name = input("Enter location_name: ")
merchant_Id = input("Enter merchant_Id: ")
location = location(location_name, merchant_Id)
locationmerchant_Id = insertlocationToDb(location)
location.merchant_Id = locationmerchant_Id
noofagents = int(input("Enter noofagents: "))
agents = []
for x in range(1, noofagents + 1):
agent_name = input("Enter agent_name: ")
location_Id = input("Enter location_Id: ")
agents = agent(agent_name, location_Id)
agentsagent_name = insertagentsToDb(agent)
agents.agent_name = agentsagent_name
nooftransactions = int(input("Enter nooftransactions: "))
transactions = []
for x in range(1, nooftransactions + 1):
amount = int(input("Enter amount: "))
balance = int(input("Enter balance: "))
member_Id = input("Enter mermber_Id: ")
clubcards_Id = input("Enter clubcards_Id: ")
agents_Id = input("Enter agents_Id: ")
transactions = transactions(amount, balance, member_Id, clubcards_Id,
agents_Id)