class NeuMF(BaseMF):
def __init__(self, model_config):
super(NeuMF, self).__init__(model_config)
self.GMF = GMF(model_config)
self.MLP = MLP(model_config)
self.mapping = nn.Linear(2, 1)
def fix_left(self, optimizer):
for param in self.GMF.parameters():
param.requires_grad = False
groups = optimizer.param_groups
lr_, betas_ = groups[0]["lr"], groups[0]["betas"]
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
self.parameters()),
lr=lr_,
betas=betas_)
return optimizer
def fix_right(self, optimizer):
for param in self.MLP.parameters():
param.requires_grad = False
groups = optimizer.param_groups
lr_, betas_ = groups[0]["lr"], groups[0]["betas"]
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
self.parameters()),
lr=lr_,
betas=betas_)
return optimizer
def load_pretrained_embedding(self, model_data):
model_data_1, model_data_2 = model_data[0], model_data[1]
if model_data_1 != "default":
print("Loading GMF embedding in NeuMF...")
self.GMF.load_embedding_from_file(model_data_1)
if model_data_2 != "default":
print("Loading MLP embedding in NeuMF...")
self.MLP.load_embedding_from_file(model_data_2)
def load_pretrained_model(self, model_data):
model_data_1, model_data_2 = model_data[0], model_data[1]
if model_data_1 != "default":
print("Loading GMF model in NeuMF...")
self.GMF.load_model_from_file(model_data_1)
if model_data_2 != "default":
print("Loading MLP model in NeuMF...")
self.MLP.load_model_from_file(model_data_2)
def get_similarity(self, input):
users, items = input[0], input[1]
sim_GMF = self.GMF([users, items])
sim_MLP = self.MLP([users, items])
features = torch.cat((sim_GMF, sim_MLP), dim=1)
sim = self.mapping(features)
return sim
def get_model():
if exp_model == 'MLP':
return MLP(hist_len, pred_len, in_dim)
elif exp_model == 'LSTM':
return LSTM(hist_len, pred_len, in_dim, city_num, batch_size, device)
elif exp_model == 'GRU':
return GRU(hist_len, pred_len, in_dim, city_num, batch_size, device)
elif exp_model == 'nodesFC_GRU':
return nodesFC_GRU(hist_len, pred_len, in_dim, city_num, batch_size, device)
elif exp_model == 'GC_LSTM':
return GC_LSTM(hist_len, pred_len, in_dim, city_num, batch_size, device, graph.edge_index)
elif exp_model == 'PM25_GNN':
return PM25_GNN(hist_len, pred_len, in_dim, city_num, batch_size, device, graph.edge_index, graph.edge_attr, wind_mean, wind_std)
elif exp_model == 'PM25_GNN_nosub':
return PM25_GNN_nosub(hist_len, pred_len, in_dim, city_num, batch_size, device, graph.edge_index, graph.edge_attr, wind_mean, wind_std)
else:
raise Exception('Wrong model name!')
def choose_model(f_dict, ter_dict):
tmp_net1 = Fed_Model()
tmp_net2 = Fed_Model()
tmp_net1.load_state_dict(f_dict)
tmp_net2.load_state_dict(ter_dict)
_, acc_1, _ = evaluate(tmp_net1, G_loss_fun, test_iter, Args)
_, acc_2, _ = evaluate(tmp_net2, G_loss_fun, test_iter, Args)
print('F: %.3f' % acc_1, 'TF: %.3f' % acc_2)
flag = False
if np.abs(acc_1 - acc_2) < 0.03:
flag = True
return ter_dict, flag
else:
return f_dict, flag
return result
data = File(u'../../../data')
data.extract()
data.shuffle()
data.convertClass(function=f, length=3)
return data.allInputs, data.allCorrect
if __name__ == '__main__':
train, target = get_data()
mlp = MLP(features = 4,
topology = [2,3])
best_mlp = mlp
def fenotype(individual):
genotype = individual.get_genotype()
pos = 0
for i, layer in enumerate(mlp.get_layers()):
for j, neuron in enumerate(layer.get_neurons()):
end_neuron_position = pos + len(neuron.get_weights())
try:
neuron.get_weights()[:] = genotype[pos: end_neuron_position]
test_num = len(x_test)
output_size = y_train.shape[1]
print(
f"feature number : {num_feature} | train_data_number : {total_num} | validation_data_number : {val_num} | test_data_number : {test_num} | class : {output_size}"
)
hidden = [80, 40]
#optimizer = SGD()
#optimizer = Momentum(0.6)
optimizer = Adam(beta_1=0.9, beta_2=0.999, weight_decay=weight_decay)
#optimizer = RMSProp(0.6, 1e-6)
model = MLP(optimizer=optimizer,
learning_rate=learning_rate,
input_size=num_feature,
output_size=output_size,
hidden=hidden,
initialize='kaiming')
acc_stack = []
loss_stack = []
val_acc_stack = []
st = time.time()
best_score = 0
for epoch in range(epochs + 1):
epoch_loss = 0.0
for iteration in range(total_num // batch_size):
seed = np.random.choice(total_num, batch_size)
x_batch = x_train[seed]
y_batch = y_train[seed]
return result
data = File(u'../../../data')
data.extract()
data.shuffle()
data.convertClass(function=f, length=3)
return data.allInputs, data.allCorrect
if __name__ == '__main__':
train, target = get_data()
mlp = MLP(features=4, topology=[2, 3])
best_mlp = mlp
def fenotype(individual):
genotype = individual.get_genotype()
pos = 0
for i, layer in enumerate(mlp.get_layers()):
for j, neuron in enumerate(layer.get_neurons()):
end_neuron_position = pos + len(neuron.get_weights())
try:
neuron.get_weights()[:] = genotype[pos:end_neuron_position]
def main(args):
# environment initialization
env = BaseEnv(size=args.env_size)
# embedding network initialization
f_s_a = MLP(env.state_size + env.action_size, args.s_a_hidden_size,
args.embedding_dim)
f_s = MLP(env.state_size, args.s_hidden_size, args.embedding_dim)
# buffer initialization
replay_buffer_1 = ReplayBuffer(args, env)
replay_buffer_2 = ReplayBuffer(args, env)
goal_buffer = GoalBuffer()
init_goal = tuple(np.random.randint(1, args.env_size[0] + 1, size=2))
goal_buffer.store(init_goal)
# agent initialization
agent = Agent(env_size=args.env_size)
# optimizer initialization
s_a_optimizer = torch.optim.Adam(f_s_a.parameters(), lr=args.s_a_lr)
s_optimizer = torch.optim.Adam(f_s.parameters(), lr=args.s_a_lr)
log_loss = []
for epoch in tqdm.tqdm(range(args.epoch_num)):
start_position = np.random.randint(1, args.env_size[0] + 1, size=2)
goal = goal_buffer.sample_batch_goal(size=1)[0]
print("goal point is :{}".format(goal))
g_feature = agent.get_state_feature(goal)
ns, r, terminate = env.reset(size=args.env_size,
start_pos=start_position)
for step in range(args.max_step):
s = ns
s_feature = agent.get_state_feature(s)
action, min_dist = agent.get_best_action(s_feature, f_s_a, f_s,
g_feature)
ns, r, terminate = env.step(action)
goal_buffer.store(ns)
ns_feature = agent.get_state_feature(ns)
vec_action = vectorize_action(action)
# print(s_feature.shape, vec_action.shape)
# store one step loss
# s_a_pred = f_s_a.predict(np.concatenate((s_feature, vec_action), axis=0).reshape((1, -1)))
# ns_pred = f_s.predict(np.array(ns_feature).reshape((1, -1)))
e_1 = agent.get_dist(s_feature, vec_action, ns_feature, f_s_a,
f_s)[0]
replay_buffer_1.add([s_feature, vec_action, ns_feature, None, e_1])
# store two step loss
sub_g = goal_buffer.sample_batch_goal(size=1, with_weights=False)
sub_g_feature = agent.get_states_feature(sub_g)[0]
na, min_dist = agent.get_best_action(ns_feature, f_s_a, f_s,
sub_g_feature)
dist_ns = agent.get_dist(ns_feature, vectorize_action(na),
sub_g_feature, f_s_a, f_s)
target = dist_ns + 1
dist_s = agent.get_dist(s_feature, vec_action, sub_g_feature,
f_s_a, f_s)
e_2 = abs(dist_s - target)
replay_buffer_2.add(
[s_feature, vec_action, ns_feature, sub_g_feature, e_2])
if terminate:
break
for step in range(args.random_step):
s = ns
s_feature = agent.get_state_feature(s)
action = agent.get_random_action()
ns, r, terminate = env.step(action)
goal_buffer.store(ns)
ns_feature = agent.get_state_feature(ns)
vec_action = vectorize_action(action)
# print(s_feature.shape, vec_action.shape)
# store one step loss
# s_a_pred = f_s_a.predict(np.concatenate((s_feature, vec_action), axis=0).reshape((1, -1)))
# ns_pred = f_s.predict(np.array(ns_feature).reshape((1, -1)))
e_1 = agent.get_dist(s_feature, vec_action, ns_feature, f_s_a,
f_s)[0]
replay_buffer_1.add([s_feature, vec_action, ns_feature, None, e_1])
# store two step loss
sub_g = goal_buffer.sample_batch_goal(size=1, with_weights=False)
sub_g_feature = agent.get_states_feature(sub_g)[0]
na, min_dist = agent.get_best_action(ns_feature, f_s_a, f_s,
sub_g_feature)
dist_ns = agent.get_dist(ns_feature, vectorize_action(na),
sub_g_feature, f_s_a, f_s)
target = dist_ns + 1
dist_s = agent.get_dist(s_feature, vec_action, sub_g_feature,
f_s_a, f_s)
e_2 = abs(dist_s - target)
replay_buffer_2.add(
[s_feature, vec_action, ns_feature, sub_g_feature, e_2])
batch_1, _, index_1 = replay_buffer_1.get_batch_data()
_, batch_2, index_2 = replay_buffer_2.get_batch_data()
loss_1 = torch.mean(
torch.norm((f_s_a(batch_1['sa']) - f_s(batch_1['ns'])), dim=1))
na = agent.get_best_actions(batch_2['ns'], f_s_a, f_s, batch_2['g'])
target = agent.get_dist(batch_2['ns'], na, batch_2['g'], f_s_a,
f_s) + 1
pred = torch.norm((f_s_a(batch_2['sa']) - f_s(batch_2['g'])), dim=1)
if (epoch + 1) % 100 == 0:
print(pred - target)
print(len(replay_buffer_1), len(replay_buffer_2))
# goal_buffer.goal_visualize()
loss_2 = torch.mean(torch.abs(pred - target))
# if epoch >= 1500:
# args.reg_term = 0.5
loss = (1 - args.reg_term) * loss_1 + args.reg_term * loss_2
log_loss.append(loss)
s_a_optimizer.zero_grad()
s_optimizer.zero_grad()
loss.backward()
# nn.utils.clip_grad_norm_(f_s.parameters(), args.grad_clip)
# nn.utils.clip_grad_norm_(f_s_a.parameters(), args.grad_clip)
s_a_optimizer.step()
s_optimizer.step()
# Update replay buffer
with torch.no_grad():
e_update_1 = torch.norm(torch.FloatTensor(
f_s_a.predict(batch_1['sa']) - f_s.predict(batch_1['ns'])),
dim=1)
na = agent.get_best_actions(batch_2['ns'], f_s_a, f_s,
batch_2['g'])
target = agent.get_dist(batch_2['ns'], na, batch_2['g'], f_s_a,
f_s) + 1
pred = torch.norm(torch.FloatTensor(
f_s_a.predict(batch_2['sa']) - f_s.predict(batch_2['g'])),
dim=1)
e_update_2 = torch.abs(pred - target)
replay_buffer_1.update_error(index_1, e_update_1)
replay_buffer_2.update_error(index_2, e_update_2)
if (epoch + 1) % 100 == 0:
print(
"epoch number: {}/{}, total_loss: {}, loss_normal: {}, loss_update: {}"
.format(epoch + 1, args.epoch_num, loss, loss_1, loss_2))
# if epoch == args.epoch_num - 1:
# print(s_a_embed, s_embed)
if (epoch + 1) % 1000 == 0 and loss < 1:
save_model("./model.pt", f_s_a, f_s, args.epoch_num, loss,
s_a_optimizer, s_optimizer)
print("Saving model at epoch: {}".format(epoch))
if (epoch + 1) % 10000 == 0:
plt.plot(range(epoch + 1), log_loss, color='red')
plt.savefig("./{}.pdf".format(int((epoch + 1) / 10000)),
dpi=1200,
bbox_inches='tight')
def visualize(args):
env = BaseEnv(size=args.env_size)
f_s_a = MLP(env.state_size + env.action_size, args.s_a_hidden_size,
args.embedding_dim)
f_s = MLP(env.state_size, args.s_hidden_size, args.embedding_dim)
agent = Agent(env_size=args.env_size)
checkpoint = torch.load(args.model_path)
f_s_a.load_state_dict(checkpoint['s_a_state_dict'])
f_s.load_state_dict(checkpoint['s_state_dict'])
f_s_a.eval()
f_s.eval()
vis_map = np.zeros(args.env_size)
dist_map = np.zeros(args.env_size)
goal = args.goal_pos
for i in range(vis_map.shape[0]):
for j in range(vis_map.shape[1]):
if (i + 1, j + 1) == goal:
vis_map[i, j] = -1
else:
state_feature = agent.get_state_feature((i + 1, j + 1))
goal_feature = agent.get_state_feature(goal)
best_action, min_dist = agent.get_best_action(
state_feature, f_s_a, f_s, goal_feature)
vis_map[i, j] = best_action
dist_map[i, j] = min_dist
for i in range(vis_map.shape[0]):
for j in range(vis_map.shape[1]):
if vis_map[i, j] == 0:
print("R", end=" ")
elif vis_map[i, j] == 1:
print("L", end=" ")
elif vis_map[i, j] == 2:
print("U", end=" ")
elif vis_map[i, j] == 3:
print("D", end=" ")
elif vis_map[i, j] == -1:
print("G", end=" ")
print("\n")
for i in range(dist_map.shape[0]):
for j in range(dist_map.shape[1]):
print(dist_map[i, j], end=" ")
print("\n")
xor = numpy.ndarray((4, 1))
xor[:, 0] = numpy.logical_xor(inputs[:, 0], inputs[:, 1])
return inputs, xor
def test(mlp):
inputs = numpy.ndarray((4, 2))
inputs[:, 0] = numpy.array([0, 0, 1, 1])
inputs[:, 1] = numpy.array([0, 1, 0, 1])
xor = numpy.ndarray((4, 1))
xor[:, 0] = numpy.logical_xor(inputs[:, 0], inputs[:, 1])
for inp, x in zip(inputs, xor):
print inp, x, mlp.output(inp)
if __name__ == '__main__':
inputs, xor = createData()
mlp = MLP(n=0.3, features=2, topology=[4, 2, 1])
errors = mlp.train_data(train=inputs, target=xor, epochs=10000, log=True)
print errors
# print mlp.get_weights()
test(mlp)
mlp.get_weights('xor.mlp')
flag = False
if np.abs(acc_1 - acc_2) < 0.03:
flag = True
return ter_dict, flag
else:
return f_dict, flag
if __name__ == '__main__':
torch.manual_seed(Args.seed)
C_iter, train_iter, test_iter, stats = data_utils.get_dataset(args=Args)
# build global network
G_net = Fed_Model()
print(G_net)
G_net.train()
G_loss_fun = torch.nn.CrossEntropyLoss()
# copy weights
w_glob = G_net.state_dict()
m = max(int(Args.frac * Args.num_C), 1)
gv_acc = []
net_best = None
val_acc_list, net_list = [], []
num_s2 = 0
# training
xor[:,0] = numpy.logical_xor(inputs[:,0], inputs[:,1])
return inputs, xor
def test(mlp):
inputs = numpy.ndarray((4,2))
inputs[:,0] = numpy.array([0,0,1,1])
inputs[:,1] = numpy.array([0,1,0,1])
xor = numpy.ndarray((4,1))
xor[:,0] = numpy.logical_xor(inputs[:,0], inputs[:,1])
for inp, x in zip(inputs, xor):
print inp, x, mlp.output(inp)
if __name__ == '__main__':
inputs, xor = createData()
mlp = MLP(n = 0.3,
features = 2,
topology = [4,2,1])
errors = mlp.train_data(train = inputs, target = xor, epochs = 10000, log=True)
print errors
# print mlp.get_weights()
test(mlp)
mlp.get_weights('xor.mlp')
def __init__(self, model_config):
super(NeuMF, self).__init__(model_config)
self.GMF = GMF(model_config)
self.MLP = MLP(model_config)
self.mapping = nn.Linear(2, 1)