def curiosity(world):
world = ActionNoise(world, stddev=0.2)
memory = Cache(max_size=100)
log_dir = "__oracle"
if not os.path.exists(log_dir):
os.mkdir(log_dir)
agent = build_agent()
agent_opt = Adams(np.random.randn(agent.n_params), lr=0.00015, memory=0.5)
oracle = build_oracle()
oracle_opt = Adam(np.random.randn(oracle.n_params) * 0.1,
lr=0.05,
memory=0.95)
for episode in range(1000):
agent.load_params(agent_opt.get_value())
oracle.load_params(oracle_opt.get_value())
agent_trajs = world.trajectories(agent, 4)
for_oracle = [[(np.asarray([o1, o2, o3]).flatten(), a1, r1)
for (o1, a1, r1), (o2, a2,
r2), (o3, a3,
r3) in zip(t, t[1:], t[2:])]
for t in agent_trajs]
memory.add_trajectories(for_oracle)
predictions = retrace(for_oracle, model=oracle)
save_plot(log_dir + "/%04d.png" % (episode + 1), agent_trajs,
predictions)
np.save(log_dir + "/%04d.npy" % (episode + 1), agent_opt.get_value())
curiosity_trajs = [[
(o1, a1, np.log(np.mean(np.square((o2 - o1) - delta_p))))
for (o1, a1, r1), (o2, a2, r2), delta_p in zip(t, t[10:], p)
] for t, p in zip(agent_trajs, predictions)]
#curiosity_trajs = replace_rewards(curiosity_trajs,
# episode=lambda rs: np.max(rs))
print_reward(curiosity_trajs, max_value=5000.0)
print_reward(agent_trajs, max_value=90.0, episode=np.sum)
curiosity_trajs = discount(curiosity_trajs, horizon=500)
curiosity_trajs = normalize(curiosity_trajs)
agent_trajs = discount(agent_trajs, horizon=500)
agent_trajs = normalize(agent_trajs)
agent_trajs = [traj[:-10] for traj in agent_trajs]
agent_weight = 0.5 # + 0.4*(0.5 * (1 - np.cos(np.pi * episode / 20)))
curiosity_weight = 1. - agent_weight
comb_trajs = combine_rewards([curiosity_trajs, agent_trajs],
[curiosity_weight, agent_weight])
grad = policy_gradient(comb_trajs, policy=agent)
agent_opt.apply_gradient(grad)
oracle_trajs = [[(o1, (o2 - o1)[:2], 1.0)
for (o1, a1, r1), (o2, a2, r2) in zip(t, t[10:])]
for t in memory.trajectories(None, 4)]
grad = policy_gradient(oracle_trajs, policy=oracle)
oracle_opt.apply_gradient(grad)
def run():
model = Input(4)
model = Affine(model, 128)
model = LReLU(model)
model = Affine(model, 2)
model = Softmax(model)
world = StochasticPolicy(Gym(make_env, max_steps=500))
opt = Adam(np.random.randn(model.n_params) * 0.1, lr=0.01)
for _ in range(50):
model.load_params(opt.get_value())
trajs = world.trajectories(model, 16)
print_reward(trajs, max_value=5000)
trajs = discount(trajs, horizon=500)
trajs = normalize(trajs)
grad = policy_gradient(trajs, policy=model)
opt.apply_gradient(grad)
while True:
world.render(model)
def __init__(self, X, Y):
# Normalize data
self.Xmean, self.Xstd = X.mean(0), X.std(0)
self.Ymean, self.Ystd = Y.mean(0), Y.std(0)
X = (X - self.Xmean) / self.Xstd
Y = (Y - self.Ymean) / self.Ystd
self.X = X
self.Y = Y
self.n = X.shape[0]
# Randomly initialize weights and noise variance
w = np.random.randn(X.shape[1], Y.shape[1])
sigma_sq = np.array([np.log([1e-3])])
# Concatenate all parameters in a single vector
self.theta = np.concatenate([w.flatten(), sigma_sq.flatten()])
# Count total number of parameters
self.num_params = self.theta.shape[0]
# Define optimizer
self.optimizer = Adam(self.num_params, lr=1e-3)
# Define loss gradient function using autograd
self.grad_loss = grad(self.loss)
def __init__(self, X, layers_Q, layers_P):
# Normalize data
self.Xmean, self.Xstd = X.mean(0), X.std(0)
X = (X - self.Xmean) / self.Xstd
self.X = X
self.layers_Q = layers_Q
self.layers_P = layers_P
self.X_dim = X.shape[1]
self.Z_dim = layers_Q[-1]
# Initialize encoder
params = self.initialize_NN(layers_Q)
self.idx_Q = np.arange(params.shape[0])
# Initialize decoder
params = np.concatenate([params, self.initialize_NN(layers_P)])
self.idx_P = np.arange(self.idx_Q[-1] + 1, params.shape[0])
self.params = params
# Total number of parameters
self.num_params = self.params.shape[0]
# Define optimizer
self.optimizer = Adam(self.num_params, lr=1e-3)
# Define gradient function using autograd
self.grad_elbo = grad(self.ELBO)
def train_one(carrOpt):
nonlocal oldTrajs
classOpt = Adam(
np.random.randn(classifier.n_params) * 1.,
lr=0.5,
memory=0.9,
)
if carrOpt == None:
carrOpt = Adam(
np.random.randn(curCarr.n_params),
lr=0.10,
memory=0.5,
)
curScore = 0.
curAccuracy = 0.
for i in range(250):
classifier.load_params(classOpt.get_value())
curCarr.load_params(carrOpt.get_value())
oldTrajIdx = np.random.choice(len(oldTrajs), size=50)
trajs = [oldTrajs[i] for i in oldTrajIdx]
trajs += world.trajectories(curCarr, 50)
trajsForClass = [tag_traj(traj, [1, 0]) for traj in trajs[:50]]
trajsForClass += [tag_traj(traj, [0, 1]) for traj in trajs[50:]]
plot_tagged_trajs(trajsForClass)
accTrajs = accuracy(trajsForClass, model=classifier)
print_reward(accTrajs,
max_value=1.0,
episode=np.mean,
label="Cla reward: ")
curAccuracy = np.mean(get_rewards(accTrajs, episode=np.mean))
if curAccuracy > 1. - i / 500:
break
grad = policy_gradient(trajsForClass, policy=classifier)
classOpt.apply_gradient(grad)
trajs2 = learn_from_classifier(classifier, trajs[50:], 1)
print_reward(trajs2,
max_value=1.0,
episode=np.max,
label="Car reward: ")
curScore = np.mean(get_rewards(trajs2, episode=np.max))
trajs2 = replace_rewards(trajs2, episode=np.max)
trajs2 = normalize(trajs2)
grad2 = policy_gradient(trajs2, policy=curCarr)
carrOpt.apply_gradient(grad2)
if i % 10 == 0:
print("%d episodes in." % i)
oldTrajs += world.trajectories(curCarr, 800)
world.render(curCarr)
if curScore > 0.11:
return carrOpt
else:
return None
def train_one(carrOpt):
if carrOpt == None:
carrOpt = Adam(
np.random.randn(curCarr.n_params),
lr=0.10,
memory=0.5,
)
nextBreak = 5
for i in range(250):
curCarr.load_params(carrOpt.get_value())
realTrajs, curiosityTrajs = world.trajectories(curCarr, 50)
curScore = np.mean(get_rewards(realTrajs, episode=np.sum)) / 90.
print_reward(realTrajs,
max_value=90.0,
episode=np.sum,
label="Real reward: ")
print_reward(curiosityTrajs,
max_value=1.0,
episode=np.max,
label="Curiosity reward: ")
curCuriosity = np.mean(get_rewards(curiosityTrajs, episode=np.max))
if curCuriosity > 0.98:
if nextBreak == 0:
break
else:
nextBreak -= 1
else:
nextBreak = np.min([nextBreak + 1, 5])
realTrajs = replace_rewards(realTrajs, episode=np.sum)
realTrajs = normalize(realTrajs)
curiosityTrajs = replace_rewards(curiosityTrajs, episode=np.max)
#this is stupid, we should care more(?) if the costs are to high
realWeight = 0.001 + np.max([np.min([curScore, 0.2]), 0.
]) * 0.998 / 0.2
curiosityWeight = 1. - realWeight
print('RWeight: %f, CWeight: %f' % (realWeight, curiosityWeight))
trajs = combine_rewards([realTrajs, curiosityTrajs],
[realWeight, curiosityWeight])
trajs = normalize(trajs)
grad = policy_gradient(trajs, policy=curCarr)
carrOpt.apply_gradient(grad)
if i % 10 == 0:
print("%d episodes in." % i)
world.remember_agent(curCarr)
world.render(curCarr)
if curScore > 0.01:
return carrOpt
else:
return None
def __init__(self,
num_layers,
units_list=None,
initializer=None,
optimizer='adam'):
self.weight_num = num_layers - 1
# 根据传入的初始化方法初始化参数，本次实验只实现xavier和全0初始化
self.params = xavier(num_layers,
units_list) if initializer == 'xavier' else zero(
num_layers, units_list)
self.optimizer = Adam(
weights=self.params,
weight_num=self.weight_num) if optimizer == 'adam' else SGD()
self.bn_param = {}
def train(world, model):
opt = Adam(np.random.randn(model.n_params), lr=0.3, memory=0.9)
for _ in range(20):
model.load_params(opt.get_value())
trajs = world.trajectories(None, 100)
grad = policy_gradient(trajs, policy=model)
opt.apply_gradient(grad)
trajs = cross_entropy(trajs, model=model)
print_reward(trajs,
episode=np.mean,
label="Surprise/byte:",
max_value=8.0)
def __init__(self, algorithm, config):
self.config = config
super(MujocoAgent, self).__init__(algorithm)
weights = self.get_weights()
assert len(
weights) == 1, "There should be only one model in the algorithm."
self.weights_name = list(weights.keys())[0]
weights = list(weights.values())[0]
self.weights_shapes = [x.shape for x in weights]
self.weights_total_size = np.sum(
[np.prod(x) for x in self.weights_shapes])
self.optimizer = Adam(self.weights_total_size, self.config['stepsize'])
def run_autoencoder(optimizer):
""" Runs the autoencoder model using the specified optimizer.
Parameters
----------
optimizer : RMSProp/Adam
Optimization algorithm to be used for parameter learning
"""
optimizer = Adam(learning_rate=0.03) if optimizer == 'adam' else RMSProp(
learning_rate=0.05)
train_matrix, val_matrix = get_training_and_val_data()
model = Autoencoder(input_dim=train_matrix.shape[1])
model.print_summary()
model.compile(optimizer)
errors = model.fit(train_matrix,
train_matrix,
num_epochs=60,
val_set=(val_matrix, val_matrix),
early_stopping=True)
plot_losses(errors['training'], errors['validation'])
neuron_num = model.model.layers[0].optimizer.reference_index
learning_rates = model.model.layers[0].optimizer.learning_rates
plot_learning_rates(learning_rates['weights'], learning_rates['bias'],
neuron_num)
def remember(agent):
nonlocal history, classOpt
history.add_trajectory(*inner.trajectories(agent, history_length))
classOpt = Adam(
np.random.randn(classifier.n_params) * 1.,
lr=0.06,
memory=0.9,
)
def curiosity(world):
world = ActionNoise(world, stddev=0.1)
memory = Cache(max_size=100)
log_dir = "__oracle"
if not os.path.exists(log_dir):
os.mkdir(log_dir)
agent = build_agent()
agent_opt = Adams(np.random.randn(agent.n_params), lr=0.00015, memory=0.5)
oracle = build_oracle()
oracle_opt = Adam(np.random.randn(oracle.n_params) * 0.1,
lr=0.05,
memory=0.95)
for episode in range(1000):
agent.load_params(agent_opt.get_value())
oracle.load_params(oracle_opt.get_value())
agent_trajs = world.trajectories(agent, 4)
memory.add_trajectories(agent_trajs)
predictions = retrace(agent_trajs, model=oracle)
save_plot(log_dir + "/%04d.png" % (episode + 1), agent_trajs,
predictions)
np.save(log_dir + "/%04d.npy" % (episode + 1), agent_opt.get_value())
agent_trajs = [[
(o1, a1, np.log(np.mean(np.square((o2 - o1) - delta_p))))
for (o1, a1, r1), (o2, a2, r2), delta_p in zip(t, t[10:], p)
] for t, p in zip(agent_trajs, predictions)]
agent_trajs = replace_rewards(agent_trajs,
episode=lambda rs: np.max(rs) / len(rs))
print_reward(agent_trajs, max_value=10.0)
agent_trajs = normalize(agent_trajs)
grad = policy_gradient(agent_trajs, policy=agent)
agent_opt.apply_gradient(grad)
oracle_trajs = [[(o1, o2 - o1, 1.0)
for (o1, a1, r1), (o2, a2, r2) in zip(t, t[10:])]
for t in memory.trajectories(None, 4)]
grad = policy_gradient(oracle_trajs, policy=oracle)
oracle_opt.apply_gradient(grad)
def train_one():
gaussOpt = Adam(
[0., 0.],
lr=0.010,
memory=0.5,
)
classOpt = Adam(np.random.randn(classifier.n_params) * 0.1,
lr=0.5,
memory=0.99)
gaussCenterer = Constant(2)
gausses.append(gaussCenterer)
curAccuracy = 0.
while curAccuracy < 0.98:
classifier.load_params(classOpt.get_value())
gaussCenterer.load_params(gaussOpt.get_value())
trajs = [[(gauss_observation(gausses[:-1]), [1, 0], 1.)]
for _ in range(500)]
trajs += [[(gauss_observation(gausses[-1:]), [0, 1], 1.)]
for _ in range(500)]
accTrajs = accuracy(trajs, model=classifier)
print_reward(accTrajs, max_value=1.0)
accs = [traj[0][2] for traj in accTrajs]
curAccuracy = np.mean(accs)
grad = policy_gradient(trajs, policy=classifier)
classOpt.apply_gradient(grad)
trajs2 = learn_from_classifier(classifier, trajs[500:], 1)
trajs2 = normalize(trajs2)
grad2 = policy_gradient(trajs2, policy=gaussCenterer)
gaussOpt.apply_gradient(grad2)
plt.clf()
plt.grid()
plt.gcf().axes[0].set_ylim([-1, 1])
plt.gcf().axes[0].set_xlim([-1, 1])
x, y = zip(*[o for ((o, _, _), ) in trajs[:500]])
plt.scatter(x, y, color="blue")
x, y = zip(*[o for ((o, _, _), ) in trajs[500:]])
plt.scatter(x, y, color="red")
plt.pause(0.01)
def reset_agent():
nonlocal agentOpt, trainTimeLeft, lastScores, curAgentId, motivation
if agentOpt is not None:
save_agent()
print("Resetting agent %d." % curAgentId)
agentOpt = Adam(
np.random.randn(agent.n_params) * 1.5,
lr=0.05,
memory=0.9,
)
trainTimeLeft = MAX_TRAIN_TIME
lastScores = [-0.4]
curAgentId += 1
motivation = MAX_MOTIVATION
def __init__(self, X, Y, hidden_dim):
# X has the form lags x data x dim
# Y has the form data x dim
self.X = X
self.Y = Y
self.X_dim = X.shape[-1]
self.Y_dim = Y.shape[-1]
self.hidden_dim = hidden_dim
self.lags = X.shape[0]
# Define and initialize neural network
self.params = self.initialize_RNN()
# Total number of parameters
self.num_params = self.params.shape[0]
# Define optimizer
self.optimizer = Adam(self.num_params, lr=1e-3)
# Define gradient function using autograd
self.grad_loss = grad(self.loss)
def init_optimizer_q_v(self, var_params_q_v):
cfg_optimizer_q_v = self['optimizer_q_v']['args']
return Adam([{
'params': [var_params_q_v['mu']],
'lr': cfg_optimizer_q_v['lr_mu']
}, {
'params': [var_params_q_v['log_var']],
'lr': cfg_optimizer_q_v['lr_log_var']
}, {
'params': [var_params_q_v['u']],
'lr': cfg_optimizer_q_v['lr_u']
}],
lr_decay=cfg_optimizer_q_v['lr_decay'])
def __init__(self, X, Y, layers):
# Normalize data
self.Xmean, self.Xstd = X.mean(0), X.std(0)
self.Ymean, self.Ystd = Y.mean(0), Y.std(0)
X = (X - self.Xmean) / self.Xstd
Y = (Y - self.Ymean) / self.Ystd
self.X = X
self.Y = Y
self.layers = layers
# Define and initialize neural network
self.params = self.initialize_NN(self.layers)
# Total number of parameters
self.num_params = self.params.shape[0]
# Define optimizer
self.optimizer = Adam(self.num_params, lr=1e-3)
# Define gradient function using autograd
self.grad_loss = grad(self.loss)
def init_optimizer_reg(self, reg_loss):
if self['optimizer_reg']['type'] != 'Adam':
print(
'only the Adam optimiser is supported for the regularisation, exiting..'
)
raise
cfg_optimizer_reg = self['optimizer_reg']['args']
if reg_loss.__class__.__name__ == 'RegLoss_LogNormal':
optimizer_reg = Adam([{
'params': [reg_loss.loc],
'lr': cfg_optimizer_reg['lr_loc']
}, {
'params': [reg_loss.log_scale],
'lr': cfg_optimizer_reg['lr_log_scale']
}],
lr_decay=cfg_optimizer_reg['lr_decay'])
elif reg_loss.__class__.__name__ == 'RegLoss_L2':
optimizer_reg = Adam(reg_loss.parameters(),
lr=cfg_optimizer_reg['lr_log_w_reg'],
lr_decay=cfg_optimizer_reg['lr_decay'])
return optimizer_reg
def init_from_str(self):
r = r"([a-zA-Z]*)=([^,)]*)"
opt_str = self.param.lower()
kwargs = dict([(i, eval(j)) for (i, j) in re.findall(r, opt_str)])
if "sgd" in opt_str:
optimizer = SGD(**kwargs)
elif "adagrad" in opt_str:
optimizer = AdaGrad(**kwargs)
elif "rmsprop" in opt_str:
optimizer = RMSProp(**kwargs)
elif "adam" in opt_str:
optimizer = Adam(**kwargs)
else:
raise NotImplementedError("{}".format(opt_str))
return optimizer
def model_builder(n_inputs, n_outputs):
model = NeuralNetwork(optimizer=Adam(), loss=CrossEntropy)
model.add(Dense(64, input_shape=(n_inputs, )))
model.add(Activation('relu'))
model.add(Dense(128))
model.add(Activation('relu'))
model.add(Dense(128))
model.add(Activation('relu'))
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dense(1))
model.add(Activation('linear'))
return model
def init_optimizer_GMM(self, data_loss):
if self['optimizer_GMM']['type'] != 'Adam':
print(
'only the Adam optimiser is supported for the GMM, exiting..')
raise
cfg_optimizer_GMM = self['optimizer_GMM']['args']
return Adam([{
'params': [data_loss.log_std],
'lr': cfg_optimizer_GMM['lr_log_std']
}, {
'params': [data_loss.logits],
'lr': cfg_optimizer_GMM['lr_logits']
}],
lr_decay=cfg_optimizer_GMM['lr_decay'])
def init_from_dict(self):
O = self.param
cc = O["cache"] if "cache" in O else None
op = O["hyperparameters"] if "hyperparameters" in O else None
if op is None:
raise ValueError("Must have `hyperparemeters` key: {}".format(O))
if op and op["id"] == "SGD":
optimizer = SGD().set_params(op, cc)
elif op and op["id"] == "RMSProp":
optimizer = RMSProp().set_params(op, cc)
elif op and op["id"] == "AdaGrad":
optimizer = AdaGrad().set_params(op, cc)
elif op and op["id"] == "Adam":
optimizer = Adam().set_params(op, cc)
elif op:
raise NotImplementedError("{}".format(op["id"]))
return optimizer
def run_GNN_Adam(train_data,
valid_data,
W,
A,
b,
B,
alpha=0.0001,
eps=0.001,
n_vector=8,
gnn_steps=2,
n_epochs=100):
"""GNN with Adamで学習, 評価を行う"""
beta1 = 0.9
beta2 = 0.999
m_W = np.zeros(W.shape)
m_A = np.zeros(A.shape)
m_b = 0.0
v_W = np.zeros(W.shape)
v_A = np.zeros(A.shape)
v_b = 0.0
# loss, precisionの保存用
params = []
# W, A, bの保存用
weights = []
for epoch in range(n_epochs):
W, A, b, loss_train = Adam(train_data, n_vector, B, W, A, b, gnn_steps,
epoch, alpha, beta1, beta2, eps, m_W, m_A,
m_b, v_W, v_A, v_b)
precision_train = mean_precision(train_data, W, A, b, n_vector,
gnn_steps)
precision_val = mean_precision(valid_data, W, A, b, n_vector,
gnn_steps)
loss_val = valid_loss(data, W, A, b, n_vector, gnn_steps)
print(
'epoch: {}, train loss: {}, train precision: {}, valid loss: {}, valid precision: {}'
.format(epoch + 1, loss_train, precision_train, loss_val,
precision_val))
params.append((loss_train, precision_train, loss_val, precision_val))
weights.append((W, A, b))
return params, weights
def train(params):
env = gym.make(params['env_name'])
params['ob_dim'] = env.observation_space.shape[0]
params['ac_dim'] = env.action_space.shape[0]
m, v = 0, 0
master = Learner(params)
n_eps = 0
n_iter = 0
ts_cumulative = 0
ts, rollouts, rewards = [], [], []
while n_iter < params['max_iter']:
reward = master.policy.rollout(env, params['steps'])
rewards.append(reward)
rollouts.append(n_eps)
ts.append(ts_cumulative)
print('Iter: %s, Eps: %s, R: %s' %(n_iter, n_eps, np.round(reward,4)))
params['n_iter'] = n_iter
gradient, timesteps = aggregate_rollouts(master, params)
ts_cumulative += timesteps
n_eps += 2 * params['sensings']
gradient /= (np.linalg.norm(gradient) / master.policy.N + 1e-8)
n_iter += 1
update, m, v = Adam(gradient, m, v, params['learning_rate'], n_iter)
master.policy.update(update)
out = pd.DataFrame({'Rollouts': rollouts, 'Reward': rewards, 'Timesteps': ts})
out.to_csv('data/%s/results/%s_Seed%s.csv' %(params['dir'], params['filename'], params['seed']), index=False)
def train(model):
world = Mnist()
opt = Adam(np.random.randn(model.n_params), lr=0.1)
for i in range(600):
model.load_params(opt.get_value() +
np.random.randn(model.n_params) * 0.01)
trajs = world.trajectories(None, 256)
grad = policy_gradient(trajs, policy=model)
opt.apply_gradient(grad)
if i % 20 == 19:
print("%4d) " % (i + 1), flush=True, end="")
trajs = world.trajectories(None, 2000)
trajs = accuracy(trajs, model=model, percent=True)
print_reward(trajs, max_value=100, label="Train accuracy:")
return opt.get_value()
class CVAE:
def __init__(self, X, Y, layers_P, layers_Q, layers_R):
# Normalize data
self.Xmean, self.Xstd = X.mean(0), X.std(0)
self.Ymean, self.Ystd = Y.mean(0), Y.std(0)
X = (X - self.Xmean) / self.Xstd
Y = (Y - self.Ymean) / self.Ystd
self.X = X
self.Y = Y
self.layers_P = layers_P
self.layers_Q = layers_Q
self.layers_R = layers_R
self.X_dim = X.shape[1]
self.Y_dim = Y.shape[1]
self.Z_dim = layers_Q[-1]
# Initialize encoder
params = self.initialize_NN(layers_P)
self.idx_P = np.arange(params.shape[0])
# Initialize decoder
params = np.concatenate([params, self.initialize_NN(layers_Q)])
self.idx_Q = np.arange(self.idx_P[-1] + 1, params.shape[0])
# Initialize prior
params = np.concatenate([params, self.initialize_NN(layers_R)])
self.idx_R = np.arange(self.idx_Q[-1] + 1, params.shape[0])
self.params = params
# Total number of parameters
self.num_params = self.params.shape[0]
# Define optimizer
self.optimizer = Adam(self.num_params, lr=1e-3)
# Define gradient function using autograd
self.grad_elbo = grad(self.ELBO)
def initialize_NN(self, Q):
hyp = np.array([])
layers = len(Q)
for layer in range(0, layers - 2):
A = -np.sqrt(6.0 / (Q[layer] + Q[layer + 1])) + 2.0 * np.sqrt(
6.0 / (Q[layer] + Q[layer + 1])) * np.random.rand(
Q[layer], Q[layer + 1])
b = np.zeros((1, Q[layer + 1]))
hyp = np.concatenate([hyp, A.ravel(), b.ravel()])
A = -np.sqrt(6.0 / (Q[-2] + Q[-1])) + 2.0 * np.sqrt(
6.0 / (Q[-2] + Q[-1])) * np.random.rand(Q[-2], Q[-1])
b = np.zeros((1, Q[-1]))
hyp = np.concatenate([hyp, A.ravel(), b.ravel()])
A = -np.sqrt(6.0 / (Q[-2] + Q[-1])) + 2.0 * np.sqrt(
6.0 / (Q[-2] + Q[-1])) * np.random.rand(Q[-2], Q[-1])
b = np.zeros((1, Q[-1]))
hyp = np.concatenate([hyp, A.ravel(), b.ravel()])
return hyp
def forward_pass(self, X, Q, params):
H = X
idx_3 = 0
layers = len(Q)
for layer in range(0, layers - 2):
idx_1 = idx_3
idx_2 = idx_1 + Q[layer] * Q[layer + 1]
idx_3 = idx_2 + Q[layer + 1]
A = np.reshape(params[idx_1:idx_2], (Q[layer], Q[layer + 1]))
b = np.reshape(params[idx_2:idx_3], (1, Q[layer + 1]))
H = np.tanh(np.matmul(H, A) + b)
idx_1 = idx_3
idx_2 = idx_1 + Q[-2] * Q[-1]
idx_3 = idx_2 + Q[-1]
A = np.reshape(params[idx_1:idx_2], (Q[-2], Q[-1]))
b = np.reshape(params[idx_2:idx_3], (1, Q[-1]))
mu = np.matmul(H, A) + b
idx_1 = idx_3
idx_2 = idx_1 + Q[-2] * Q[-1]
idx_3 = idx_2 + Q[-1]
A = np.reshape(params[idx_1:idx_2], (Q[-2], Q[-1]))
b = np.reshape(params[idx_2:idx_3], (1, Q[-1]))
Sigma = np.exp(np.matmul(H, A) + b)
return mu, Sigma
def ELBO(self, params):
X = self.X_batch
Y = self.Y_batch
# Prior: p(z|x)
mu_0, Sigma_0 = self.forward_pass(X, self.layers_R, params[self.idx_R])
# Encoder: q(z|x,y)
mu_1, Sigma_1 = self.forward_pass(np.concatenate([X, Y], axis=1),
self.layers_Q, params[self.idx_Q])
# Reparametrization trick
epsilon = np.random.randn(X.shape[0], self.Z_dim)
Z = mu_1 + epsilon * np.sqrt(Sigma_1)
# Decoder: p(y|x,z)
mu_2, Sigma_2 = self.forward_pass(np.concatenate([X, Z], axis=1),
self.layers_P, params[self.idx_P])
# Log-determinants
log_det_0 = np.sum(np.log(Sigma_0))
log_det_1 = np.sum(np.log(Sigma_1))
log_det_2 = np.sum(np.log(Sigma_2))
# KL[q(z|x,y) || p(z|x)]
KL = 0.5 * (np.sum(Sigma_1 / Sigma_0) + np.sum(
(mu_0 - mu_1)**2 / Sigma_0) - self.Z_dim + log_det_0 - log_det_1)
# -log p(y|x,z)
NLML = 0.5 * (np.sum((Y - mu_2)**2 / Sigma_2) + log_det_2 +
np.log(2. * np.pi) * self.Y_dim * X.shape[0])
return NLML + KL
# Fetches a mini-batch of data
def fetch_minibatch(self, X, Y, N_batch):
N = X.shape[0]
idx = np.random.choice(N, N_batch, replace=False)
X_batch = X[idx, :]
Y_batch = Y[idx, :]
return X_batch, Y_batch
# Trains the model
def train(self, nIter=10000, batch_size=100):
start_time = timeit.default_timer()
for it in range(nIter):
# Fetch minibatch
self.X_batch, self.Y_batch = self.fetch_minibatch(
self.X, self.Y, batch_size)
# Evaluate loss using current parameters
params = self.params
elbo = self.ELBO(params)
# Update parameters
grad_params = self.grad_elbo(params)
self.params = self.optimizer.step(params, grad_params)
# Print
if it % 10 == 0:
elapsed = timeit.default_timer() - start_time
print('It: %d, ELBO: %.3e, Time: %.2f' % (it, elbo, elapsed))
start_time = timeit.default_timer()
def generate_samples(self, X_star, N_samples):
X_star = (X_star - self.Xmean) / self.Xstd
# Encode X_star
mu_0, Sigma_0 = self.forward_pass(X_star, self.layers_R,
self.params[self.idx_R])
# Reparametrization trick
epsilon = np.random.randn(N_samples, self.Z_dim)
Z = mu_0 + epsilon * np.sqrt(Sigma_0)
# Decode
mean_star, var_star = self.forward_pass(
np.concatenate([X_star, Z], axis=1), self.layers_P,
self.params[self.idx_P])
# De-normalize
mean_star = mean_star * self.Ystd + self.Ymean
var_star = var_star * self.Ystd**2
return mean_star, var_star
from keras.callbacks import ModelCheckpoint, CSVLogger, TensorBoard, LearningRateScheduler
from optimizers import Adam, schedule
from layers import get_loss_funcs, show_gpus
from model import thin_model
from dataloader import data_gen_train, data_gen_val
from config import logs_dir, weights_best_file, training_log, base_lr, max_iter, batch_size
show_gpus()
train_samples = data_gen_train.size()
val_samples = data_gen_val.size()
iterations_per_epoch = train_samples // batch_size
adam = Adam(lr=base_lr)
loss_funcs = get_loss_funcs(batch_size)
thin_model.compile(loss=loss_funcs, optimizer=adam, metrics=["accuracy"])
checkpoint = ModelCheckpoint(weights_best_file,
monitor='loss',
verbose=0,
save_best_only=True,
save_weights_only=True,
mode='min',
period=1)
csv_logger = CSVLogger(training_log, append=True)
tb = TensorBoard(log_dir=logs_dir,
histogram_freq=0,
write_graph=True,
write_images=False)
lrate = LearningRateScheduler(schedule)
callbacks_list = [checkpoint, csv_logger, tb, lrate]
# plt.subplot(1, 4, 2)
# plt.imshow(img[1])
# plt.subplot(1, 4, 3)
# plt.imshow(img[2])
# plt.subplot(1, 4, 4)
# plt.imshow(img[3])
model = MNISTNet()
loss = SoftmaxCrossEntropy(num_class=10)
# define your learning rate sheduler
def func(lr, iteration):
if iteration % 1000 == 0:
return lr * 0.5
else:
return lr
adam = Adam(lr=0.01, decay=0, sheduler_func=func)
l2 = L2(w=0.001) # L2 regularization with lambda=0.001
model.compile(optimizer=adam, loss=loss, regularization=l2)
train_results, val_results, test_results = model.train(mnist,
train_batch=30,
val_batch=1000,
test_batch=1000,
epochs=2,
val_intervals=100,
test_intervals=300,
print_intervals=100)
train_y = convert_to_one_hot(train_y, num_classes)
test_x = np.reshape(test_x, (len(test_x), 1, img_rows, img_cols)).astype(skml_config.config.i_type)
test_y = convert_to_one_hot(test_y, num_classes)
train_x, valid_x, train_y, valid_y = train_test_split(train_x, train_y)
filters = 64
model = Sequential()
model.add(Convolution(filters, 3, input_shape=input_shape))
model.add(BatchNormalization())
model.add(ReLU())
model.add(MaxPooling(2))
model.add(Convolution(filters, 3))
model.add(BatchNormalization())
model.add(ReLU())
model.add(GlobalAveragePooling())
model.add(Affine(num_classes))
model.compile(SoftmaxCrossEntropy(), Adam())
train_batch_size = 100
valid_batch_size = 1
print("訓練開始: {}".format(datetime.now().strftime("%Y/%m/%d %H:%M")))
model.fit(train_x, train_y, train_batch_size, 20, validation_data=(valid_batch_size, valid_x, valid_y), validation_steps=1)
print("訓練終了: {}".format(datetime.now().strftime("%Y/%m/%d %H:%M")))
model.save(save_path)
loss, acc = model.evaluate(test_x, test_y)
print("Test loss: {}".format(loss))
print("Test acc: {}".format(acc))
class NeuralNetwork:
# Initialize the class
def __init__(self, X, Y, layers):
# Normalize data
self.Xmean, self.Xstd = X.mean(0), X.std(0)
self.Ymean, self.Ystd = Y.mean(0), Y.std(0)
X = (X - self.Xmean) / self.Xstd
Y = (Y - self.Ymean) / self.Ystd
self.X = X
self.Y = Y
self.layers = layers
# Define and initialize neural network
self.params = self.initialize_NN(self.layers)
# Total number of parameters
self.num_params = self.params.shape[0]
# Define optimizer
self.optimizer = Adam(self.num_params, lr=1e-3)
# Define gradient function using autograd
self.grad_loss = grad(self.loss)
# Initializes the network weights and biases using Xavier initialization
def initialize_NN(self, Q):
params = np.array([])
num_layers = len(Q)
for layer in range(0, num_layers - 1):
weights = -np.sqrt(6.0 /
(Q[layer] + Q[layer + 1])) + 2.0 * np.sqrt(
6.0 /
(Q[layer] + Q[layer + 1])) * np.random.rand(
Q[layer], Q[layer + 1])
biases = np.zeros((1, Q[layer + 1]))
params = np.concatenate([params, weights.ravel(), biases.ravel()])
return params
# Evaluates the forward pass
def forward_pass(self, X, Q, params):
H = X
idx_3 = 0
num_layers = len(self.layers)
# All layers up to last
for layer in range(0, num_layers - 2):
idx_1 = idx_3
idx_2 = idx_1 + Q[layer] * Q[layer + 1]
idx_3 = idx_2 + Q[layer + 1]
weights = np.reshape(params[idx_1:idx_2], (Q[layer], Q[layer + 1]))
biases = np.reshape(params[idx_2:idx_3], (1, Q[layer + 1]))
H = np.tanh(np.matmul(H, weights) + biases)
# Last linear layer
idx_1 = idx_3
idx_2 = idx_1 + Q[-2] * Q[-1]
idx_3 = idx_2 + Q[-1]
weights = np.reshape(params[idx_1:idx_2], (Q[-2], Q[-1]))
biases = np.reshape(params[idx_2:idx_3], (1, Q[-1]))
mu = np.matmul(H, weights) + biases
return mu
# Evaluates the mean square error loss
def loss(self, params):
X = self.X_batch
Y = self.Y_batch
mu = self.forward_pass(X, self.layers, params)
return np.mean((Y - mu)**2)
# Fetches a mini-batch of data
def fetch_minibatch(self, X, Y, N_batch):
N = X.shape[0]
idx = np.random.choice(N, N_batch, replace=False)
X_batch = X[idx, :]
Y_batch = Y[idx, :]
return X_batch, Y_batch
# Trains the model by minimizing the MSE loss
def train(self, nIter=10000, batch_size=100):
start_time = timeit.default_timer()
for it in range(nIter):
# Fetch minibatch
self.X_batch, self.Y_batch = self.fetch_minibatch(
self.X, self.Y, batch_size)
# Evaluate loss using current parameters
params = self.params
loss = self.loss(params)
# Update parameters
grad_params = self.grad_loss(params)
self.params = self.optimizer.step(params, grad_params)
# Print
if it % 10 == 0:
elapsed = timeit.default_timer() - start_time
print('It: %d, Loss: %.3e, Time: %.2f' % (it, loss, elapsed))
start_time = timeit.default_timer()
# Evaluates predictions at test points
def predict(self, X_star):
# Normalize inputs
X_star = (X_star - self.Xmean) / self.Xstd
y_star = self.forward_pass(X_star, self.layers, self.params)
# De-normalize outputs
y_star = y_star * self.Ystd + self.Ymean
return y_star