def update(self):
state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.replay_buffer.sample(
self.batch_size)
with autograd.record():
# get the Q(s,a)
all_current_q_value = self.main_network(state_batch)
main_q_value = nd.pick(all_current_q_value, action_batch)
# get the maxQ(s',a')
all_next_q_value = self.target_network(
next_state_batch).detach() # only get gradient of main network
max_next_q_value = nd.max(all_next_q_value, axis=1)
target_q_value = reward_batch + (
1 - done_batch) * self.gamma * max_next_q_value
# record loss
loss = gloss.L2Loss()
value_loss = loss(target_q_value, main_q_value)
self.main_network.collect_params().zero_grad()
value_loss.backward()
self.optimizer.step(batch_size=self.batch_size)
def training2(features, labels, position, fire_point, title, batch_size=11):
dataset = gdata.ArrayDataset(features, labels)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
net = nn.Sequential()
net.add(nn.Dense(1))
net.initialize()
loss = gloss.L2Loss()
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.01})
epoch = 0
m = 100
while epoch < 1000:
epoch += 1
for X, y in data_iter:
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
l = loss(net(features), labels)
if epoch % 100 == 0 or epoch == 1:
print('epcoh: %d loss: %s' % (epoch, l.mean().asscalar()))
m = l.mean().asscalar()
print("epoch:" + str(epoch))
print("m: " + str(m))
print("w: " + str(net[0].weight.data().asnumpy()))
print("b: " + str(net[0].bias.data().asnumpy()))
print("error square: " +
str(((net(features).reshape(batch_size, 1) -
labels.reshape(batch_size, 1))**2).sum().asscalar()))
p_test = nd.arange(0, 1 / fire_point, 0.1)
q_test = net(p_test)
plt.subplot(1, 3, position)
plt.scatter(features.asnumpy(), labels.asnumpy(), color='#FF4700')
plt.plot(p_test.asnumpy(), q_test.asnumpy(), color='b')
plt.xlabel("1/p (1/kPa)")
plt.ylabel("1/q (g carbon/mmol)")
plt.title(title)
def optimize(batch_size, trainer, num_epochs, decay_epoch, log_interval,
features, labels, net):
"""Optimize an objective function."""
dataset = gdata.ArrayDataset(features, labels)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
loss = gloss.L2Loss()
ls = [loss(net(features), labels).mean().asnumpy()]
for epoch in range(1, num_epochs + 1):
# Decay the learning rate.
if decay_epoch and epoch > decay_epoch:
trainer.set_learning_rate(trainer.learning_rate * 0.1)
for batch_i, (X, y) in enumerate(data_iter):
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
if batch_i * batch_size % log_interval == 0:
ls.append(loss(net(features), labels).mean().asnumpy())
# To print more conveniently, use numpy.
print('w:', net[0].weight.data(), '\nb:', net[0].bias.data(), '\n')
es = np.linspace(0, num_epochs, len(ls), endpoint=True)
semilogy(es, ls, 'epoch', 'loss')
def train_gluon_ch7(trainer_name,
trainer_hyperparams,
features,
labels,
batch_size=10,
num_epochs=2):
# Iniatial model
net = nn.Sequential()
net.add(nn.Dense(1))
net.initialize(init.Normal(sigma=0.01))
loss = gloss.L2Loss()
#Store the loss
def eval_loss():
return loss(net(features), labels).mean().asscalar()
ls = [eval_loss()]
data_iter = gdata.DataLoader(gdata.ArrayDataset(features, labels),
batch_size,
shuffle=True)
# Create Trainer
trainer = gluon.Trainer(net.collect_params(), trainer_name,
trainer_hyperparams)
for _ in range(num_epochs):
start = time.time()
for batch_i, (X, y) in enumerate(data_iter):
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size) # Average the gradient
if (batch_i + 1) * batch_size % 100 == 0:
ls.append(eval_loss())
# Print result and graph
print('loss: %f, %f sec per epoch' % (ls[-1], time.time() - start))
d2l.set_figsize()
d2l.plt.plot(np.linspace(0, num_epochs, len(ls)), ls)
d2l.plt.xlabel('epoch')
d2l.plt.ylabel('loss')
def house_prise_gulon():
"""
使用gulon模型构建房价预估
:return:
"""
features = nd.array(nd.array([[120, 2], [100, 1], [130, 3]]))
labels = nd.array([1200000, 1000000, 1300000])
logger.info(features)
logger.info(labels)
# labels += nd.random.normal(scale=0.01, shape=labels.shape)
batch_size = 10
# 将训练数据的特征和标签组合
dataset = gdata.ArrayDataset(features, labels)
# 随机读取小批量
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
net = nn.Sequential()
net.add(nn.Dense(1))
net.initialize(init.Normal(sigma=0.01))
loss = gloss.L2Loss()
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.03})
num_epochs = 3
for epoch in range(1, num_epochs + 1):
for X, y in data_iter:
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
l = loss(net(features), labels)
logger.info('epoch %d, loss: %f' % (epoch, l.mean().asnumpy()))
dense = net[0]
logger.info("预测数据")
logger.info(dense.weight.data())
logger.info(dense.bias.data())
logger.info(net(features))
def update(self):
state_batch, action_batch, reward_batch, next_state_batch = self.replay_buffer.sample(
self.batch_size)
with autograd.record():
# get the Q(s,a)
all_current_q_value = self.main_network(state_batch)
main_q_value = nd.pick(all_current_q_value, action_batch)
# different from DQN
# get next action from main network, then get its Q value from target network
all_next_q_value = self.target_network(
next_state_batch).detach() # only get gradient of main network
max_action = nd.argmax(all_current_q_value, axis=1)
target_q_value = nd.pick(all_next_q_value, max_action).detach()
target_q_value = reward_batch + self.gamma * target_q_value
# record loss
loss = gloss.L2Loss()
value_loss = loss(target_q_value, main_q_value)
self.main_network.collect_params().zero_grad()
value_loss.backward()
self.optimizer.step(batch_size=self.batch_size)
def optimize(batch_size, trainer, num_epochs, decay_epoch, log_interval,
features, labels, net):
i = 0
dataset = gdata.ArrayDataset(features, labels)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
loss = gloss.L2Loss()
ls = [loss(net(features), labels).mean().asnumpy()]
for epoch in range(1, num_epochs + 1):
# 学习率自我衰减。
if decay_epoch and epoch > decay_epoch:
trainer.set_learning_rate(trainer.learning_rate * 0.1)
for batch_i, (X, y) in enumerate(data_iter):
with autograd.record():
l = loss(net(X), y)
l.backward()
i += 1
trainer.step(batch_size)
if batch_i * batch_size % log_interval == 0:
ls.append(loss(net(features), labels).mean().asnumpy())
# 为了便于打印，改变输出形状并转化成 numpy 数组。
print('i:', i, 'w:', net[0].weight.data(), '\nb:', net[0].bias.data(),
'\n')
es = np.linspace(0, num_epochs, len(ls), endpoint=True)
gb.semilogy(es, ls, 'epoch', 'loss')
def train_gluon(trainer_name, hyperparams, batch_size):
#【读取数据】
data_iter = gdata.DataLoader(gdata.ArrayDataset(features, labels),
batch_size,
shuffle=True)
#【定义模型】
net = nn.Sequential()
net.add(nn.Dense(1))
#【初始化模型参数】
net.initialize(init=init.Normal(sigma=0.01))
#【定义损失函数】
loss = gloss.L2Loss()
#【定义优化算法】
trainer = gluon.Trainer(net.collect_params(), trainer_name, hyperparams)
#【训练模型】
num_epochs = 2
start = 0
ls = []
for _ in range(num_epochs):
start = time.time()
for epoch_i, (X, y) in enumerate(data_iter):
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
trainer.step(batch_size)
if (epoch_i + 1) * batch_size % 100 == 0:
ls.append(loss(net(features), labels).mean().asscalar())
print('loss %f,%f sec per epoch' % (ls[-1],
(time.time() - start) / len(ls)))
plt.plot(np.linspace(0, num_epochs, len(ls)), ls)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()
def optimize(batch_size, trainer, num_epochs, decay_epoch, log_interval, X, y,
net):
# num_examples = 1000
# X, y = genData(num_examples)
dataset = gdata.ArrayDataset(X, y)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
square_loss = gloss.L2Loss()
y_vals = [square_loss(net(X), y).mean().asnumpy()]
for epoch in range(1, num_epochs + 1):
if decay_epoch and epoch > decay_epoch:
trainer.set_learning_rate(trainer.learning_rate * 0.1)
for batch_i, (features, label) in enumerate(data_iter):
with autograd.record():
output = net(features)
loss = square_loss(output, label)
loss.backward()
trainer.step(batch_size)
if batch_i * batch_size % log_interval == 0:
y_vals.append(square_loss(net(X), y).mean().asnumpy())
# 为了便于打印，改变输出形状并转化成numpy数组。
print('w:', net[0].weight.data(), '\nb:', net[0].bias.data(), '\n')
x_vals = np.linspace(0, num_epochs, len(y_vals), endpoint=True)
utils.semilogy(x_vals, y_vals, 'epoch', 'loss')
def optimize_gluon(trainer, features, labels, net, decay_epoch=None,
batch_size=10, log_interval=10, num_epochs=3):
"""Optimize an objective function with a Gluon trainer."""
dataset = gdata.ArrayDataset(features, labels)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
loss = gloss.L2Loss()
ls = [loss(net(features), labels).mean().asnumpy()]
for epoch in range(1, num_epochs + 1):
# Decay the learning rate.
if decay_epoch and epoch > decay_epoch:
trainer.set_learning_rate(trainer.learning_rate * 0.1)
for batch_i, (X, y) in enumerate(data_iter):
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
if batch_i * batch_size % log_interval == 0:
ls.append(loss(net(features), labels).mean().asnumpy())
print('w[0]=%.2f, w[1]=%.2f, b=%.2f'
% (net[0].weight.data()[0][0].asscalar(),
net[0].weight.data()[0][1].asscalar(),
net[0].bias.data().asscalar()))
es = np.linspace(0, num_epochs, len(ls), endpoint=True)
semilogy(es, ls, 'epoch', 'loss')
def update(self):
state = nd.array([t.state for t in self.buffer], ctx=self.ctx)
action = nd.array([t.action for t in self.buffer], ctx=self.ctx)
reward = [t.reward for t in self.buffer]
# next_state = nd.array([t.next_state for t in self.buffer], ctx=self.ctx)
old_action_log_prob = nd.array([t.a_log_prob for t in self.buffer],
ctx=self.ctx)
R = 0
Gt = []
for r in reward[::-1]:
R = r + self.gamma * R
Gt.insert(0, R)
Gt = nd.array(Gt, ctx=self.ctx)
# sample 'ppo_update_time' times
# sample 'batch_size' samples every time
for i in range(self.ppo_update_times):
assert len(self.buffer) >= self.batch_size
sample_index = random.sample(range(len(self.buffer)),
self.batch_size)
for index in sample_index:
# optimize the actor network
with autograd.record():
Gt_index = Gt[index]
V = self.critic_network(state[index].reshape(1,
-1)).detach()
advantage = (Gt_index - V)
all_action_prob = self.actor_network(state[index].reshape(
1, -1))
action_prob = nd.pick(all_action_prob, action[index])
ratio = action_prob / old_action_log_prob[index]
surr1 = ratio * advantage
surr2 = nd.clip(ratio, 1 - self.clip_param,
1 + self.clip_param) * advantage
action_loss = -nd.mean(nd.minimum(surr1,
surr2)) # attention
self.actor_network.collect_params().zero_grad()
action_loss.backward()
actor_network_params = [
p.data()
for p in self.actor_network.collect_params().values()
]
gb.grad_clipping(actor_network_params,
theta=self.clip_param,
ctx=self.ctx)
self.actor_optimizer.step(1)
# optimize the critic network
with autograd.record():
Gt_index = Gt[index]
V = self.critic_network(state[index].reshape(1, -1))
loss = gloss.L2Loss()
value_loss = nd.mean(loss(Gt_index, V))
self.critic_network.collect_params().zero_grad()
value_loss.backward()
critic_network_params = [
p.data()
for p in self.critic_network.collect_params().values()
]
gb.grad_clipping(critic_network_params,
theta=self.clip_param,
ctx=self.ctx)
self.critic_optimizer.step(1)
self.training_step += 1
# clear buffer
del self.buffer[:]
def __init__(self, net, ctx=mx.cpu()):
super(Regression, self).__init__(net=net, ctx=ctx)
self.loss_fun = gloss.L2Loss()
self.metric = mx.metric.PearsonCorrelation()
def update(self):
self.total_train_steps += 1
state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory_buffer.sample(
self.batch_size)
# --------------optimize the critic network--------------------
with autograd.record():
# choose next action according to target policy network
next_action_batch = self.target_actor_network(next_state_batch)
noise = nd.normal(loc=0,
scale=self.policy_noise,
shape=next_action_batch.shape,
ctx=self.ctx)
# with noise clip
noise = nd.clip(noise,
a_min=-self.noise_clip,
a_max=self.noise_clip)
next_action_batch = next_action_batch + noise
clipped_action = self.action_clip(next_action_batch)
# get target q value
target_q_value1 = self.target_critic_network1(
next_state_batch, clipped_action)
target_q_value2 = self.target_critic_network2(
next_state_batch, clipped_action)
target_q_value = nd.minimum(target_q_value1,
target_q_value2).squeeze()
target_q_value = reward_batch + (1.0 - done_batch) * (
self.gamma * target_q_value)
# get current q value
current_q_value1 = self.main_critic_network1(
state_batch, action_batch)
current_q_value2 = self.main_critic_network2(
state_batch, action_batch)
loss = gloss.L2Loss()
value_loss1 = loss(current_q_value1, target_q_value.detach())
value_loss2 = loss(current_q_value2, target_q_value.detach())
self.main_critic_network1.collect_params().zero_grad()
value_loss1.backward()
self.critic1_optimizer.step(self.batch_size)
self.main_critic_network2.collect_params().zero_grad()
value_loss2.backward()
self.critic2_optimizer.step(self.batch_size)
# ---------------optimize the actor network-------------------------
if self.total_train_steps % self.policy_update == 0:
with autograd.record():
pred_action_batch = self.main_actor_network(state_batch)
actor_loss = -nd.mean(
self.main_critic_network1(state_batch, pred_action_batch))
self.main_actor_network.collect_params().zero_grad()
actor_loss.backward()
self.actor_optimizer.step(1)
self.soft_update(self.target_actor_network,
self.main_actor_network)
self.soft_update(self.target_critic_network1,
self.main_critic_network1)
self.soft_update(self.target_critic_network2,
self.main_critic_network2)
from mxnet.gluon import loss loss_softmax = loss.SoftmaxCrossEntropyLoss(sparse_label=False) loss_mse = loss.L2Loss()
labels = true_w[0] * features[:, 0] + true_w[1] * features[:, 1] + true_b
labels += nd.random.normal(scale=0.01, shape=labels.shape)
batch_size = 10
dataset = gdata.ArrayDataset(features, labels)
#随机读取小批量
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
net = nn.Sequential()
net.add(nn.Dense(1)) #Dense定义该层输出个数为 1。全连接：Dense
#初始化模型参数
net.initialize(init.Normal(sigma=0.01)) #指定权重参数每个元素将在初始化时随机采样于均值为 0 标准差为 0.01 的正态分布。
#定义损失函数
loss = gloss.L2Loss() # 平⽅损失⼜称 L2 范数损失。
#定义优化算法 学习率的数值一般设置为1/batch_size
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.03}) #指定学习率为 0.03 的小批量随机梯度下降（sgd）为优化算法 些参数可以通过 collect_params 函数获取
#训练模型
num_epochs = 3
for epoch in range(1, num_epochs + 1):
for X, y in data_iter:
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size) #迭代模型参数 指明批量⼤小，从而对批量中样本梯度求平均
l = loss(net(features), labels)
print('epoch %d, loss: %f' % (epoch, l.mean().asnumpy()))
batch_size = 10
dataset = gdata.ArrayDataset(features, labels)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)
from mxnet.gluon import nn
net = nn.Sequential() #建立一个net容器
net.add(nn.Dense(1)) # 自动计算Input点数
from mxnet import init
net.initialize(init.Normal(sigma=0.01)) # initial
from mxnet.gluon import loss as gloss
loss = gloss.L2Loss() # L2 loss
from mxnet import gluon
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.03})
num_epoch = 3
for epoch in range(1, num_epoch + 1):
for X, y in data_iter:
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
print('epoch {0} , loss {1}'.format(
epoch,
loss(net(features), labels).mean().asnumpy()))
def __init__(self, **kwargs):
super(PolynomialRegressionGluon, self).__init__(**kwargs)
self.net = None
self.loss = gloss.L2Loss()
self.regularization = kwargs.get("regularization")
def __init__(self, feature_dict, args, ctx, task, **kwargs):
"""{"sparse":[SingleFeat],"dense":[SingleFeat]}"""
super(xDeepFM, self).__init__(**kwargs) # ??
util.mkdir_if_not_exist(args.SAVE_PARAMS_PATH_PREFIX)
# self.feature_sizes = args.FEATURE_SIZE
self.field_size = args.FIELD_NUM
self.feature_dict = feature_dict
print('field_size:')
print(self.field_size)
if args.TASK == 'finish':
self.embedding_size = args.FINISH_EMBEDDING_SIZE
self.batch_size = args.FINISH_BATCH_SIZE
else:
self.embedding_size = args.LIKE_EMBEDDING_SIZE
self.batch_size = args.LIKE_BATCH_SIZE
self.config_name = args.CONFIG_NAME
# self.dropout_prob = args.DROPOUT_PROB
self.task = task
# self.loss = gloss.SigmoidBinaryCrossEntropyLoss()
if args.LOSS == 'l2loss':
self.loss = gloss.L2Loss()
else:
self.loss = gloss.SigmoidBinaryCrossEntropyLoss()
self.ctx = ctx
self.embedding_dict = OrderedDict()
self.dense_dict = OrderedDict()
with self.name_scope():
if self.task == 'finish':
self.layer_list = [np.int(x) for x in args.FINISH_LAYER]
self.dropout = args.FINISH_DROPOUT_PROB
else:
self.layer_list = [np.int(x) for x in args.LIKE_LAYER]
self.dropout = args.LIKE_DROPOUT_PROB
# self.params.get('v',shape=(self.field_size,self.embedding_size))
self.dnn_out = nn.Dense(1, use_bias=False)
self.register_child(self.dnn_out)
for feat in feature_dict['sparse']:
self.embedding_dict[feat.feat_name] = nn.Embedding(
feat.feat_num, self.embedding_size)
for feat in feature_dict['dense']:
self.dense_dict[feat.feat_name] = nn.Dense(self.embedding_size)
for emb_k, emb_v in self.embedding_dict.items():
self.register_child(emb_v)
for den_k, den_v in self.dense_dict.items():
self.register_child(den_v)
self.linear_logit_dense = nn.Dense(1, use_bias=False)
self.register_child(self.linear_logit_dense)
self.linear_logit_embedding_bn = nn.BatchNorm()
self.register_child(self.linear_logit_embedding_bn)
self.dense_list = []
self.dropout_list = []
self.bn_list = []
self.activation_list = []
for i in range(len(self.layer_list)):
self.dense_list.append(nn.Dense(self.layer_list[i]))
self.dropout_list.append(nn.Dropout(self.dropout))
self.bn_list.append(nn.BatchNorm())
self.activation_list.append(nn.Activation('relu'))
self.register_child(self.dense_list[i])
self.register_child(self.dropout_list[i])
self.register_child(self.bn_list[i])
self.register_child(self.activation_list[i])
# if True:
print('true')
self.layer_size = [np.int(x) for x in args.CONV1D_LAYER]
# self.cin_net = CIN(self.embedding_size,self.field_size, (128, 64), self.ctx)
# print('oo')
# self.cin_net.initialize()
# print('uu')
# self.register_child(self.cin_net)
self.cin_dense = nn.Dense(1)
self.register_child(self.cin_dense)
self.cin_bn = nn.BatchNorm()
self.register_child(self.cin_bn)
self.field_nums = [self.field_size]
self.conv_list = []
for idx, size in enumerate(self.layer_size):
self.conv_list.append(
nn.Conv1D(channels=size,
kernel_size=1,
strides=1,
padding=0,
activation='relu',
in_channels=self.field_nums[0] *
self.field_nums[-1],
weight_initializer=init.Uniform()))
self.field_nums.append(size)
self.register_child(self.conv_list[idx])
# 1. prepare training data set
n_train, n_test, true_w, true_b = 100, 100, [1.2, -3.4, 5.6], 5
features = nd.random.normal(shape=(n_train + n_test, 1))
poly_features = nd.concat(features, nd.power(features, 2),
nd.power(features, 3))
labels = (true_w[0] * poly_features[:, 0] + true_w[1] * poly_features[:, 1] +
true_w[2] * poly_features[:, 2] + true_b)
labels += nd.random.normal(loc=0, scale=0.01, shape=labels.shape)
print(features[:2], poly_features[:2], labels[:2])
print("features.shape={}, poly_features.shape={} ".format(
features.shape, poly_features.shape))
print("labels.shape={}".format(labels.shape))
num_epochs, loss = 100, gloss.L2Loss()
# 2. training model
def fit_and_plot(train_features, test_features, train_labels, test_labels):
net = nn.Sequential()
net.add(nn.Dense(1))
net.initialize()
batch_size = min(10, train_labels.shape[0])
train_iter = gdata.DataLoader(gdata.ArrayDataset(train_features,
train_labels),
batch_size,
shuffle=True)
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.01})
train_ls, test_ls = [], []
modelname = 'semi_pi_simple2'
basemodel_zoo = 'simple2'
net = symbols.get_model('simple2')
net.initialize(mx.init.Xavier(magnitude=2.24))
#net.load_parameters(os.path.join('symbols','para','%s.params'%(modelname)))
# g(x) : stochastic input augmentation function
def g(x):
return x + nd.random.normal(0, stochastic_ratio, shape=x.shape)
# loss function
l_logistic = gloss.SoftmaxCrossEntropyLoss()
l_l2loss = gloss.L2Loss()
metric = mx.metric.Accuracy()
# train
def test():
metric = mx.metric.Accuracy()
for data, label in val_data:
X = data.reshape((-1, 1, 28, 28))
#img = nd.concat(X,X,X,dim=1)
output = net(X)
metric.update([label], [output])
return metric.get()
def train(epochs, alpha):
def log_rmse(net, features, labels):
l2_loss = gloss.L2Loss()
clipped_preds = nd.clip(net(features), 1, float('inf'))
rmse = nd.sqrt(2 * l2_loss(clipped_preds.log(), labels.log()).mean())
return rmse
def __init__(self):
super(LinRegGluon, self).__init__()
self.net = None
if not self.loss:
self.loss = g_loss.L2Loss()
import mxnet from mxnet.gluon import nn, loss as gloss import mxnet as mx import mxnet.ndarray as nd from mxnet import nd, autograd, gluon from mxnet.gluon.data.vision import transforms # L2 Loss loss2 = gloss.L2Loss() # sample data x = nd.ones((2, )) y = nd.ones((2, )) * 2 loss2(x, y) # Huber loss loss_huber = gloss.HuberLoss(rho=0.85) # threshold rho loss = gloss.SoftmaxCrossEntropyLoss() x = nd.array([[1, 10], [8, 2]]) y = nd.array([0, 1]) loss(x, y)
# 全连接层
net.add(nn.Dense(1))
# 【初始化模型参数】
# 默认初始化的是权重参数,采用标准差为0.01的正态分布
# 偏差参数默认初始化为0
# 这时net已经有了w和b，后续操作中只需要将输入输入层即可
# 这个net(X)的操作就是求出预测值的过程！
net.initialize(init.Normal(sigma=0.01))
# 【定义损失函数】
loss = loss.L2Loss() # 平方损失又称为L2范数损失
# 【定义优化算法】
# 定义一个模型参数优化算法一共需要4个参数，w,b,learning_rate,batch_size
# 其中batch_size这个参数在最后的step函数中给出！
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.03})
# 【训练模型】
num_epochs = 3
for epoch in range(num_epochs):
for X, y in data_iter:
with autograd.record():
l = loss(net(X), y)
def build_model(self):
# DataLoader
train_transform = transforms.Compose([
transforms.RandomFlipLeftRight(),
transforms.Resize((self.img_size + 30, self.img_size + 30)),
transforms.RandomResizedCrop(self.img_size),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
test_transform = transforms.Compose([
transforms.Resize((self.img_size, self.img_size)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
self.trainA = ImageFolder(
os.path.join('dataset', self.dataset, 'trainA'), train_transform)
self.trainB = ImageFolder(
os.path.join('dataset', self.dataset, 'trainB'), train_transform)
self.testA = ImageFolder(
os.path.join('dataset', self.dataset, 'testA'), test_transform)
self.testB = ImageFolder(
os.path.join('dataset', self.dataset, 'testB'), test_transform)
self.trainA_loader = DataLoader(self.trainA,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers)
self.trainB_loader = DataLoader(self.trainB,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers)
self.testA_loader = DataLoader(self.testA, batch_size=1, shuffle=False)
self.testB_loader = DataLoader(self.testB, batch_size=1, shuffle=False)
""" Define Generator, Discriminator """
self.genA2B = ResnetGenerator(input_nc=3,
output_nc=3,
ngf=self.ch,
n_blocks=self.n_res,
img_size=self.img_size,
light=self.light)
self.genB2A = ResnetGenerator(input_nc=3,
output_nc=3,
ngf=self.ch,
n_blocks=self.n_res,
img_size=self.img_size,
light=self.light)
self.disGA = Discriminator(input_nc=3, ndf=self.ch, n_layers=7)
self.disGB = Discriminator(input_nc=3, ndf=self.ch, n_layers=7)
self.disLA = Discriminator(input_nc=3, ndf=self.ch, n_layers=5)
self.disLB = Discriminator(input_nc=3, ndf=self.ch, n_layers=5)
self.whole_model = nn.HybridSequential()
self.whole_model.add(*[
self.genA2B, self.genB2A, self.disGA, self.disGB, self.disLA,
self.disLB
])
self.whole_model.hybridize(static_alloc=False, static_shape=False)
""" Define Loss """
self.L1_loss = gloss.L1Loss()
self.MSE_loss = gloss.L2Loss(weight=2)
self.BCE_loss = gloss.SigmoidBCELoss()
""" Initialize Parameters"""
params = self.whole_model.collect_params()
block = self.whole_model
if not self.debug:
force_init(block.collect_params('.*?_weight'), KaimingUniform())
force_init(block.collect_params('.*?_bias'),
BiasInitializer(params))
block.collect_params('.*?_rho').initialize()
block.collect_params('.*?_gamma').initialize()
block.collect_params('.*?_beta').initialize()
block.collect_params('.*?_state_.*?').initialize()
else:
pass
block.collect_params().reset_ctx(self.dev)
""" Trainer """
self.G_params = param_dicts_merge(
self.genA2B.collect_params(),
self.genB2A.collect_params(),
)
self.G_optim = gluon.Trainer(
self.G_params,
'adam',
dict(learning_rate=self.lr,
beta1=0.5,
beta2=0.999,
wd=self.weight_decay),
)
self.D_params = param_dicts_merge(self.disGA.collect_params(),
self.disGB.collect_params(),
self.disLA.collect_params(),
self.disLB.collect_params())
self.D_optim = gluon.Trainer(
self.D_params,
'adam',
dict(learning_rate=self.lr,
beta1=0.5,
beta2=0.999,
wd=self.weight_decay),
)
""" Define Rho clipper to constraint the value of rho in AdaILN and ILN"""
self.Rho_clipper = RhoClipper(0, 1)
break
from mxnet.gluon import nn
net = nn.Sequential()
net.add(nn.Dense(1))
from mxnet import init
net.initialize(init.Normal(sigma=0.01))
from mxnet.gluon import loss as gloss
loss = gloss.L2Loss() # 平方损失又称 L2 范数损失。
from mxnet import gluon
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.03})
num_epochs = 3
for epoch in range(1, num_epochs + 1):
for X, y in data_iter:
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
l = loss(net(features), labels)
print('epoch %d, loss: %f' % (epoch, l.mean().asnumpy()))
def train_GAS_ch9(model, data_utils, batch_size, lr, num_epochs, ctx):
model.initialize(init.Xavier(), force_reinit=True, ctx=ctx)
trainer = gluon.Trainer(model.collect_params(), 'adam',
{'learning_rate': lr})
loss = d2l.MaskedSoftmaxCELoss()
loss1 = gloss.SigmoidBinaryCrossEntropyLoss(from_sigmoid=True)
loss2 = gloss.L2Loss()
animator = d2l.Animator(xlabel='epoch',
ylabel='loss',
xlim=[1, num_epochs],
ylim=[0, 0.25])
for epoch in range(1, num_epochs + 1):
timer = d2l.Timer()
metric = d2l.Accumulator(2) # loss_sum, num_tokens
l_sum, l_class_sum, l_score_sum, n, acc_sum = 0.0, 0.0, 0.0, 0, 0.0
rmse_sum = nd.array([0, 0, 0])
data_iter = data_utils.get_batch_train(batch_size)
# ti=0
for i, (X, Y) in enumerate(data_iter):
# print(X.shape,Y.shape) ##X=(128, 10, 16) (128, 5)
# exit()
X, Y = nd.array(X).as_np_ndarray(), nd.array(Y).as_np_ndarray()
# print("after reshape",X.shape, Y.shape) ##after reshape (128, 10, 16) (128, 5)
# exit()
# vlinz=nd.random.randint(0,10,(X.shape[0],)).as_np_ndarray()
# vlinz=nd.ones((X.shape[0],)).as_np_ndarray()
valid_len = np.repeat(np.array([X.shape[1]]), X.shape[0])
# print(valid_len.shape) ##(128,)
# exit()
# ti+=1
# print('keepup',ti)
with autograd.record():
dec_output = model(X, valid_len)
# print(dec_output.shape) ##(128, 10, 5)
# exit()
# ###################
# l = loss(dec_output, Y, vlinz)
# l.backward()
# d2l.grad_clipping(model, 1)
# num_tokens = vlinz.sum()
# trainer.step(num_tokens)
# metric.add(l.sum(), num_tokens)
# # exit()
# if epoch % 10 == 0:
# animator.add(epoch, (metric[0] / metric[1],))
# print(f'loss {metric[0] / metric[1]:.3f}, {metric[1] / timer.stop():.1f} '
# f'tokens/sec on {str(ctx)}')
# ##########################
output = dec_output.as_nd_ndarray()
cl_res, score_res = class_and_score_forward(output)
# print("shape of cl_res:",cl_res.shape,"shape of Y[0][:,:3]:",Y.shape)
# print("shape of score_res:",score_res.shape,"shape of Y[0][:,3:]:",Y.shape)
cl_weight, conc_weight = nd.ones_like(cl_res), nd.ones_like(
score_res)
l_class = loss1(cl_res.as_np_ndarray(), Y[:, :3],
cl_weight.as_np_ndarray()).sum()
l_conc = loss2(score_res.as_np_ndarray(), Y[:, 3:],
conc_weight.as_np_ndarray()).sum()
n = Y.shape[0]
l = (l_class / n) + (l_conc / n)
l.backward()
d2l.grad_clipping(model, 1)
num_tokens = n
trainer.step(num_tokens)
metric.add(l.sum(), num_tokens)
if epoch % 10 == 0:
animator.add(epoch, (metric[0] / metric[1], ))
print(f'loss {metric[0] / metric[1]:.3f}, {metric[1] / timer.stop():.1f} '
f'tokens/sec on {str(ctx)}')
return gluon.data.DataLoader(dataset,batch_size,shuffle=is_train)
if (__name__=='__main__'):
true_w=nd.array([2,-3.4])
true_b=4.2
features,labels=synthetic_data(true_w,true_b,1000)
batch_size=10
data_iter=load_array((features,labels),batch_size)
net=nn.Sequential()
net.add(nn.Dense(1))
net.initialize(init.Normal(sigma=0.01))
loss=gloss.L2Loss() #The squared loss is known as the L2 norm loss
trainer=gluon.Trainer(net.collect_params(),'sgd',{'learning_rate':0.03})
num_epochs=3
for epoch in range(1,num_epochs+1):
for X,y in data_iter:
with autograd.record():
l=loss(net(X),y)
l.backward()
trainer.step(batch_size)
l=loss(net(features),labels)
print('epoch %d,loss: %f'%(epoch,l.mean().asnumpy()))
print('After regression, w is ',net[0].weight.data())
print('After regression, b is ',net[0].bias.data())
y2_vals=None,
legend=None,
figsize=(3.5, 2.5)):
gb.plt.rcParams['figure.figsize'] = figsize
set_matplotlib_formats('retina')
gb.plt.xlabel(x_label)
gb.plt.ylabel(y_label)
gb.plt.semilogy(x_vals, y_vals)
if x2_vals and y2_vals:
gb.plt.semilogy(x2_vals, y2_vals)
gb.plt.legend(legend)
gb.plt.show()
num_epochs = 100
loss = gloss.L2Loss()
def fit_and_plot(train_features, test_features, train_labels, test_labels):
net = nn.Sequential()
net.add(nn.Dense(1))
net.initialize()
batch_size = min(100, train_labels.shape[0])
train_iter = gdata.DataLoader(gdata.ArrayDataset(train_features,
train_labels),
batch_size,
shuffle=True)
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.01})
train_ls, test_ls = [], []
def main() -> None:
"""
Main execution of the module.
"""
# Setup the same initial constraints as our previous linear regression model.
true_weights = np.array([2, -3.4])
true_bias = 4.2
features, targets = d2l.synthetic_data(true_weights, true_bias, 1000)
batch_size = 10
data_iterator = load_array((features, targets), batch_size, True)
# Create a seuqential neural network with one output layer. gluon will infer
# the input shape the first time data is passed through to it.
net = nn.Sequential()
net.add(nn.Dense(1))
# Initialize the weights with a random sample from a normal distribution
# with a mean of 0 and a standard deviation of 0.01. bias is initialized as
# by default. The initialization is deferred until the first attempt to pass
# data through the network.
net.initialize(init.Normal(sigma=0.01))
# The squared loss is also known as the L2 norm loss.
l2_loss = loss.L2Loss()
# Setup our SGD optimizer through the trainer class.
trainer = gluon.Trainer(net.collect_params(), "sgd",
{"learning_rate": 0.03})
num_epochs = 3
# Training loop time
for epoch in range(1, num_epochs + 1):
for feature_batch, target_batch in data_iterator:
with autograd.record():
predicted_targets = net(feature_batch)
batch_loss = l2_loss(predicted_targets, target_batch)
# Compute the gradients for all of our weights and bias. The trainer
# initialized the parameters for us already, allowing us to not worry
# about manually attaching gradients.
batch_loss.backward()
# Because we're passing in a number of batches, we need to compute
# reduction of all gradients in order to update our model
# accordingly.
trainer.step(batch_size)
# Compute the overall loss for the epoch.
epoch_loss = l2_loss(net(features), targets)
print(f"epoch {epoch}, loss: {epoch_loss.mean().asnumpy()}")
# Obtain the weights and biases from the first (and only) layer inside of
# our model.
first_layer_weights = net[0].weight.data()
first_layer_bias = net[0].bias.data()
print(
f"Error in estimating the weights: {true_weights.reshape(first_layer_weights.shape) - first_layer_weights}"
)
print(f"Error in estimating the bias: {true_bias - first_layer_bias}")
