def fit(self, epochs=1000, alpha=0.001, beta=0, verbose=False):
"""
Find optimal weights to fit to samples.
:param epochs: Number of epochs to train model for \\
:param alpha: Learning rate hyperparameter \\
:param beta: Regularization hyperparameter \\
:param verbose: Flag for verbose logging \\
:return: Optimal weights \\
"""
X, y = self.data, self.labels
optimizer = tf.train.GradientDescentOptimizer(learning_rate=alpha)
gradient = tfe.implicit_gradients(self._loss)
for epoch in range(epochs):
optimizer.apply_gradients(gradient(beta))
if verbose:
once_every = epochs // 10
if (epoch + 1) % once_every == 0:
loss = self._loss(beta)
print('Epoch: %i --> loss = %0.3f' % (epoch + 1, loss))
return self.w.numpy()
def train():
batch_size = 10
num_epoches = 2
data_size = 500 * 10
display_iter = 100
dir = r"E:\code\python\deeplearning\tensorflow1.x\data\mnist_digits_images"
save_dir = r"E:\\code\\python\\deeplearning\\tensorflow1.x\\data\\ck\\"
dataset = get_dataset(directory=dir, size=[28, 28], batch_size=batch_size)
iterator = dataset.make_one_shot_iterator()
data = iterator.get_next()
model = MNISTModel(name='net')
global_step = tf.train.get_or_create_global_step()
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
grad_fn = tfe.implicit_gradients(compute_loss)
while global_step * batch_size / data_size < num_epoches:
step = int(global_step * batch_size / data_size)
x, y = tf.cast(data[0], dtype=tf.float32), data[1]
grads_and_vars = grad_fn(model, x, y)
optimizer.apply_gradients(
grads_and_vars=grads_and_vars,
global_step=tf.train.get_or_create_global_step())
# 获取要保存的变量
if global_step % display_iter == 0:
all_variables = (model.variables + optimizer.variables() +
[global_step])
tfe.Saver(all_variables).save(save_dir,
global_step=global_step) # 检查点文件
print("Epoch:{}, Iteration:{}, loss:{}".format(
step, global_step, compute_loss(model, x, y)))
global_step = tf.train.get_or_create_global_step()
def optimize_policy(self,
policy,
env,
n_samples_per_rollout,
time_horizon,
n_iter=1000,
verbose=1):
if policy.type == 'discrete':
loss = self.loss_discrete
elif policy.type == 'continuous':
loss = self.loss_continuous
grads = tfe.implicit_gradients(loss)
for i in range(n_iter):
observations, actions, q_values, mean_reward = self.policy_rollout(
policy, env, n_samples_per_rollout, time_horizon)
if verbose >= 1 and (i+1) % 1 == 0:
print('Iteration {0}. Loss: {1}. Average reward: {2}.'
.format(i+1,
loss(policy, observations, actions, q_values),
mean_reward))
if verbose >= 2 and (i+1) % 10 == 0:
render_policy(policy, env)
self.optimizer.apply_gradients(
grads(policy, observations, actions, q_values))
return n_iter
def call(self, inputs):
"""Calculates logits and action.
Args:
inputs: Observations from a step in the cart-pole environment, of shape
`(batch_size, input_size)`
Returns:
logits: the logits output by the output layer. This can be viewed as the
likelihood vales of choosing the left (0) action. Shape:
`(batch_size, 1)`.
actions: randomly selected actions ({0, 1}) based on the logits. Shape:
`(batch_size, 1)`.
"""
hidden = self._hidden_layer(inputs)
logits = self._output_layer(hidden)
left_prob = tf.nn.sigmoid(logits)
action_probs = tf.concat([left_prob, 1.0 - left_prob], 1)
self._grad_fn = eager.implicit_gradients(
self._get_cross_entropy_and_save_actions)
actions = tf.multinomial(tf.log(action_probs), 1)
return logits, actions
def __init__(self, hidden_size):
super(EagerPolicyNetwork, self).__init__()
self._hidden_layer = tf.keras.layers.Dense(hidden_size,
activation=tf.nn.elu)
self._output_layer = tf.keras.layers.Dense(1)
self._grad_fn = eager.implicit_gradients(
self._get_cross_entropy_and_save_actions)
def __init__(self, params):
super(SimpleNN, self).__init__()
self.params = params
self.hidden_layers = [
self.track_layer(tf.layers.Dense(params.state_size, activation=tf.nn.relu))
for _ in range(params.num_layers)]
self.output_layer = self.track_layer(tf.layers.Dense(params.num_classes, name='output_layer'))
self.optimizer = tf.train.AdamOptimizer(1e-3)
self.grads_fn = tfe.implicit_gradients(self.loss_fn)
def __init__(self, env):
self.actions = list(range(env.action_space.n))
self.num_states = env.observation_space.n
self.rewards = tfe.Variable(tf.random_uniform(
[env.observation_space.n]),
name="rewards")
"""
self.rewards = tfe.Variable(initial_value=[0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 1.0,],
name="rewards")
"""
self._updater = tfe.implicit_gradients(self.loss)
def train():
datasize = 100
train_X, train_y = generate_data(datasize=datasize)
dataset = gen_datast(train_X, train_y)
iterator = dataset.make_one_shot_iterator()
data = iterator.get_next()
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)
grad_fn = tfe.implicit_gradients(compute_loss)
global_step = tf.train.get_or_create_global_step()
iter = 0
training_epoches = 20
display_step = 2
batch_size = 10
plot_steps = []
plot_losses = []
while global_step * batch_size / datasize < training_epoches:
step = int(global_step * batch_size / datasize)
X = tf.cast(data['X'], dtype=tf.float32)
y = tf.cast(data['y'], dtype=tf.float32)
grads_and_vars = grad_fn(linear_regression, X, y)
optimizer.apply_gradients(
grads_and_vars=grads_and_vars, global_step=global_step
) # compute_cost的输入为linear_regression, data['X'], data['y']
loss = compute_loss(linear_regression, X, y)
if step % display_step == 0:
print("Epoch:", step + 1, "loss=", loss, "weight=", weight, "b=",
b)
if not (loss == "NA"):
# plot_steps.append(global_step.eval()) 不行
plot_steps.append(iter)
plot_losses.append(loss)
data = iterator.get_next()
global_step = tf.train.get_or_create_global_step()
iter += 1
print("Done!!!")
print(plot_steps)
show_data(plot_steps, plot_losses)
def optimize_policy(self,
env,
n_samples_per_rollout,
time_horizon=200,
n_iter=1000,
early_stop=True,
verbose=True):
grads = tfe.implicit_gradients(self.loss)
for i in range(n_iter):
observations, actions, q_values, mean_reward = self.policy_rollout(
env, n_samples_per_rollout, time_horizon)
if verbose and (i + 1) % 10 == 0:
print('Iteration {0}. Average reward: {1}'.format(
i + 1, mean_reward))
if mean_reward == 200 and early_stop:
return i + 1
self.optimizer.apply_gradients(
grads(observations, actions, q_values))
return n_iter
#定义优化器
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
#定义函数计算loss
def _grad_fn(yolo_v3, images_tensor, list_y_trues):
logits = yolo_v3(images_tensor)
loss = loss_fn(list_y_trues,
logits,
anchors=COCO_ANCHORS,
image_size=[imgsize, imgsize])
return loss
grad = tfe.implicit_gradients(_grad_fn) #获得计算梯度的函数
history = []
for i in range(num_epoches):
_loop_train(yolo_v3, optimizer, generator, grad) #训练
loss_value = _loop_validation(yolo_v3, generator) #验证
print("{}-th loss = {}".format(i, loss_value))
#收集loss
history.append(loss_value)
if loss_value == min(history): #只有loss创新低时再保存模型
print(" update weight {}".format(loss_value))
yolo_v3.save_weights("{}.h5".format(save_fname))
################################################################
#使用模型
def all_loss(encoder, decoder, img_tensor, target):
loss = 0
hidden = decoder.reset_state(batch_size=target.shape[0])
dec_input = tf.expand_dims([word_index['<start>']] * target.shape[0], 1)
feature = encoder(img_tensor) # (batch_size,49,256)
for i in range(1, target.shape[1]):
predictions, hidden, _ = decoder(dec_input, feature, hidden)
loss += loss_mask(target[:, i], predictions)
dec_input = tf.expand_dims(target[:, i], 1)
return loss
grad = tfe.implicit_gradients(all_loss)
# 实现单步训练过程
def train_one_epoch(encoder, decoder, optimizer, step_counter, dataset, epoch):
total_loss = 0
for (step, (img_tensor, target)) in enumerate(dataset):
loss = 0
optimizer.apply_gradients(grad(encoder, decoder, img_tensor, target), step_counter)
loss = all_loss(encoder, decoder, img_tensor, target)
total_loss += (loss / int(target.shape[1]))
if step % 5 == 0:
print('Epoch {} Batch {} Loss {:.4f}'.format(epoch + 1,
step,
# Using sparse_softmax cross entropy
return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=inference_fn(inputs), labels=labels))
# Calculate accuracy
def accuracy_fn(inference_fn, inputs, labels):
prediction = tf.nn.softmax(inference_fn(inputs))
correct_pred = tf.equal(tf.argmax(prediction, 1), labels)
return tf.reduce_mean(tf.cast(correct_pred, tf.float32))
# SGD Optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
# Compute gradients
grad = tfe.implicit_gradients(loss_fn)
# Training
average_loss = 0.
average_acc = 0.
for step in range(num_steps):
# Iterate through the dataset
try:
d = dataset_iter.next()
except StopIteration:
# Refill queue
dataset_iter = tfe.Iterator(dataset)
d = dataset_iter.next()
# Images
global_step = tf.train.get_or_create_global_step()
def getcost(x, y): #定义函数，计算loss值
# 前向结构
z = tf.cast(tf.multiply(np.asarray(x, dtype=np.float32), W) + b,
dtype=tf.float32)
cost = tf.reduce_mean(tf.square(y - z)) #loss值
return cost
learning_rate = 0.01
# 随机梯度下降法作为优化器
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
grad = tfe.implicit_gradients(getcost) #获得计算梯度的函数
#定义saver，演示两种方法处理检查点文件
savedir = "logeager/"
savedirx = "logeagerx/"
saver = tf.train.Saver([W, b],
max_to_keep=1) #生成saver。 max_to_keep=1，表明最多只保存一个检查点文件
saverx = tfe.Saver([W, b]) #生成saver。 max_to_keep=1，表明最多只保存一个检查点文件
kpt = tf.train.latest_checkpoint(savedir) #找到检查点文件
kptx = tf.train.latest_checkpoint(savedirx) #找到检查点文件
if kpt != None:
saver.restore(None, kpt) #两种加载方式都可以
saverx.restore(kptx)
# Create Linear Regression function
def linear_regression(theta):
return theta * W + b
# Create MSE function (y-yi)^2 / 2*n_samples
def MSE(model, theta, labels):
#return tf.reduce_sum(tf.pow(model(theta) - labels, 2) / (2*n_samples))
return tf.reduce_mean(tf.square(tf.subtract(model(theta), labels)))
# Define SGD Optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
# Computation of gradients
grad = tfe.implicit_gradients(MSE)
# Print initial cost
cost = MSE(linear_regression, train_X, train_Y)
print("Cost: %.9f" % cost)
print("Weight: %s" % W.numpy())
print("Bias: %s" % b.numpy())
for step in range(num_steps):
optimizer.apply_gradients(grad(linear_regression, train_X, train_Y))
if (step + 1) % display_step == 0 or step == 0:
print("Epoch:", '%04d' % (step + 1), "cost=",
"{:.9f}".format(MSE(linear_regression, train_X, train_Y)), "W=",
W.numpy(), "b=", b.numpy())
if batch < n_batch - 1:
batch_X = train_list[batch * batch_size:(batch + 1) *
batch_size, :, :, :]
batch_Y = train_label[batch * batch_size:(batch + 1) *
batch_size, :, :]
else:
batch_X = train_list[batch * batch_size:, :, :, :]
batch_Y = train_label[batch * batch_size:, :, :]
return batch_X, batch_Y
# Adam Optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
# Compute gradients
grad = tfe.implicit_gradients(mean_square_fn)
# Initial cost, before optimizing
print("Initial cost = {:.9f}".format(mean_square_fn(pred, batch_Y)))
# Training
costs = []
params = stage_1_util_v1.params
for epoch in range(epochs):
batch_cost = 0.0
for batch in range(n_batch):
batch_X, batch_Y = next_batch(batch, batch_size)
pred = stage_1_model(batch_X, n_joints, is_training,
def train(self, global_step):
if self.use_difference_critic:
# Sample replay buffer for training
batchA = self.memory.sample(batch_size=self.batch_size * 2)
batchB = self.memory.sample(batch_size=self.batch_size * 2)
# Prepare training data for critic
actionA1s = self.actor(tf.convert_to_tensor(batchA['obs1']))
actionB1s = self.actor(tf.convert_to_tensor(batchB['obs1']))
diff_Qs = self.difference_critic(
tf.convert_to_tensor(batchA['obs1']), actionA1s,
tf.convert_to_tensor(batchB['obs1']), actionB1s)
target_diff_Qs = batchA['rewards'] - batchB[
'rewards'] + self.gamma * diff_Qs
target_diff_Qs = tf.clip_by_value(
tf.convert_to_tensor(target_diff_Qs), 2 * self.return_range[0],
2 * self.return_range[1])
target_diff_Qs = tf.constant(
target_diff_Qs
) # target_diff_Qs should be treated as a constant (like supervised learning labels)
# Gradient descent step on critic
diff_critic_grad_fn = tfe.implicit_gradients(
self.difference_critic_loss)
difference_critic_grad = diff_critic_grad_fn(
tf.convert_to_tensor(batchA['obs0']),
tf.convert_to_tensor(batchA['actions']),
tf.convert_to_tensor(batchB['obs0']),
tf.convert_to_tensor(batchB['actions']), target_diff_Qs)
difference_critic_grad = filter_grads_and_vars_by_scope(
difference_critic_grad, 'difference_critic')
self.difference_critic_optimizer.apply_gradients(
difference_critic_grad, global_step=global_step)
# Gradient descent step on actor
actor_grad_fn = tfe.implicit_gradients(
self.actor_loss_with_difference_critic)
actor_grad = actor_grad_fn(tf.convert_to_tensor(batchA['obs0']),
tf.convert_to_tensor(batchB['obs0']),
tf.convert_to_tensor(batchB['actions']))
actor_grad = filter_grads_and_vars_by_scope(
actor_grad,
'actor') # We only want to optimize the actor at this stage
self.actor_optimizer.apply_gradients(actor_grad,
global_step=global_step)
else:
# Sample replay buffer for training
batch = self.memory.sample(batch_size=self.batch_size)
# Prepare training data for critic
action1s = self.actor(tf.convert_to_tensor(batch['obs1']))
Qs = self.critic(tf.convert_to_tensor(batch['obs1']), action1s)
terminals = batch['terminals1'].astype('float32')
target_Qs = batch['rewards'] + (1. - terminals) * self.gamma * Qs
target_Qs = tf.clip_by_value(tf.convert_to_tensor(target_Qs),
self.return_range[0],
self.return_range[1])
target_Qs = tf.constant(
target_Qs
) # target_Qs should be treated as a constant (like supervised learning labels)
# Gradient descent step on critic_loss
critic_grad_fn = tfe.implicit_gradients(self.critic_loss)
critic_grad = critic_grad_fn(
tf.convert_to_tensor(batch['obs0']),
tf.convert_to_tensor(batch['actions']), target_Qs)
critic_grad = filter_grads_and_vars_by_scope(critic_grad, 'critic')
self.critic_optimizer.apply_gradients(critic_grad,
global_step=global_step)
# Gradient descent step on actor
actor_grad_fn = tfe.implicit_gradients(self.actor_loss)
actor_grad = actor_grad_fn(tf.convert_to_tensor(batch['obs0']))
actor_grad = filter_grads_and_vars_by_scope(actor_grad, 'actor')
self.actor_optimizer.apply_gradients(actor_grad,
global_step=global_step)
grad_f = tfe.gradients_function(f)
print(f(2., 1.)) # 2.0
print(grad_f(2., 1.)) # [Tensor(3.0), Tensor(-2.0)]
# It's also possible to calculate partial derivatives wrt. Variables implicitly
# involved in the computation of the function
x = tfe.Variable(1.0, name='x')
y = tfe.Variable(2.0, name='y')
def g(z):
return z * x + z * y**2
grad_g = tfe.gradients_function(g)
imp_grad_g = tfe.implicit_gradients(g)
print(g(3.)) # Tensor(15.0)
print(grad_g(3.)) # [Tensor(5.0)] is [dg/dz]
print(imp_grad_g(3.)) # [(Tensor(3.0), Variable(name='x:0')),
# (Tensor(12.0), Variable(name='y:0'))] is [dg/dx, dg/dy]
# You can also calculate partial derivatives of one computation inside a
# GradientTape context manager
x = tf.constant(3.)
y = tfe.Variable(3.)
with tfe.GradientTape() as tape:
tape.watch(x) # Trainable Variables are automatically watched
f = x**2 + y**2
print(tape.gradient(f, [x, y]))
# Linear regression (Wx + b)
def linear_regression(inputs):
return inputs * W + b
# Mean square error
def mean_square_fn(model_fn, inputs, labels):
return tf.reduce_sum(tf.pow(model_fn(inputs) - labels, 2)) / (2 * n_samples)
# SGD Optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
# Compute gradients
grad = tfe.implicit_gradients(mean_square_fn)
# Initial cost, before optimizing
print("Initial cost= {:.9f}".format(
mean_square_fn(linear_regression, train_X, train_Y)),
"W=", W.numpy(), "b=", b.numpy())
# Training
for step in range(num_steps):
optimizer.apply_gradients(grad(linear_regression, train_X, train_Y))
if (step + 1) % display_step == 0 or step == 0:
print("Epoch:", '%04d' % (step + 1), "cost=",
"{:.9f}".format(mean_square_fn(linear_regression, train_X, train_Y)),
"W=", W.numpy(), "b=", b.numpy())
# 必须在loss_fn函数里面进行前项传播，然后经implicit_gradients包装后，才能进行梯度计算
def loss_fn(forward, inputs, ys, keep_prob, is_training, momentum=0.9):
# tf.nn.softmax_cross_entropy_with_logits_v2等价于tf.nn.softmax+cross_entropy
# 1.最后一层不需要使用激活函数softmax，直接使用softmax_cross_entropy_with_logits_v2即可
return tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=ys,
logits=forward(
inputs, keep_prob,
is_training, momentum)))
'''--------定义loss函数 end--------'''
'''--------定义优化器与梯度函数 start--------'''
# ys与out_layer都是one-hot类型的向量
optimizer = tf.train.AdamOptimizer(0.01)
grad_fn = tfe.implicit_gradients(loss_fn)
'''--------定义优化器与梯度函数 end--------'''
'''--------训练start--------'''
model = CNNModel()
for i in range(100000):
# 获取要训练的数据
batch_xs, batch_ys = mnist.train.next_batch(32)
# 计算每一个变量的梯度
grad = grad_fn(model.forward, batch_xs.reshape(-1, 28, 28, 1), batch_ys,
0.8, True)
# 优化梯度
optimizer.apply_gradients(grad)
if i % 50 == 0:
# 测试阶段
y_pre = model.forward(tf.cast(x_test, dtype=tf.float32), 1, False)
# 测试正确率
## Gradients - Automatic differentiation is built into eager execution
# Under the hood ...
# - Operations are recorded on a tap
# - The tape is Played back to compute gradients
# - This is reverse-mode differentiation (backpropagation)
# Eg: 01
def square(x):
return x**2
grad = tfe.gradients_function(square) # differentiation w.r.t input of square
print(square(3.))
print(grad(3.))
# Eg: 02
x = tfe.Variable(2.0) # use when eager execution is enabled
def loss(y):
return (y - x**2)**2
grad = tfe.implicit_gradients(
loss) # Differentiation w.r.t variables used to compute loss
print(loss(7.))
print(grad(7.)) # [gradient w.r.t x, x]
for i in x:
print(i)
def square(x):
return x**2
grad = tfe.gradients_function(square)
print(square(3.))
print(grad(3.))
x = tfe.Variable(2.0)
def loss(y):
return (y - x**2)**2
grad = tfe.implicit_gradients(loss)
print(loss(7.))
print(grad(7.))
#APIs for computing gradients work
#tfe.gradients_function()
#tfe.value_and_gradients_function()
#tfe.implicit_gradients()
#tfe.implicit_value_and_gradients()
def __init__(self):
super(MNISTModel, self).__init__()
self.layer1 = self.track_layer(tf.layers.Dense(units=10))
self.layer2 = self.track_layer(tf.layers.Dense(units=10))
def call(self, input):
"""Actually runs the model."""
result = self.layer1(input)
result = self.layer2(result)
return result
# Let's make up a blank input image
model = MNISTModel()
batch = tf.zeros([1, 1, 784])
print(batch.shape)
result = model(batch)
print(result)
def loss_function(model, x, y):
y_ = model(x)
return tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=y_)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
for (x, y) in tfe.Iterator(batch):
grads = tfe.implicit_gradients(loss_function)(model, x, y)
optimizer.apply_gradients(grads)
