Sequential: Sequential neural networks

In frhl/neuralnetr: Neural Nets in R

Description

Setup a sequential (uni-directional) deep feed forward neural network. The Sequential is used as a wrapper for the modules.

Super class

neuralnetr::ClassModule -> Sequential

Public fields

loss

The loss module

modules

the list of current modules

Methods

Public methods

• Sequential$new()

• Sequential$sgd()

• Sequential$mini_gd()

• Sequential$forward()

• Sequential$backward()

• Sequential$sgd_step()

• Sequential$classify()

• Sequential$print_accuarcy()

• Sequential$clone()

Inherited methods

---

Method new()

setup the neural network

Usage

Sequential$new(modules, loss)

Arguments

modules

list of modules

loss

the loss module

---

Method sgd()

train neural network using stochastic gradient descent

Usage

Sequential$sgd(X, Y, iters = 100, lrate = 0.005, verbose = F, seed = 1)

Arguments

X

the X input (m x b)

Y

the Y (target) input

iters

amount of iterations.

lrate

the learning rate.

verbose

print results every epoch.

seed

random seed.

Returns

cumulative loss for every iteration.

---

Method mini_gd()

train neural network using minibatch gradient descent.

Usage

Sequential$mini_gd(
  X,
  Y,
  iters = 100,
  lrate = 0.005,
  K = 5,
  verbose = F,
  seed = 1
)

Arguments

X

the X input (m x b)

Y

the Y (target) input

iters

amount of iterations.

lrate

the learning rate.

K

the size of the minibatch.

verbose

print results every epoch.

seed

random seed.

Returns

cumulative loss for every iteration.

---

Method forward()

Compute Ypred

Usage

Sequential$forward(Xt)

Arguments

Xt

the input at time t

---

Method backward()

Update dLdW and dLdW0

Usage

Sequential$backward(delta)

Arguments

delta

the backpropagated error

---

Method sgd_step()

gradient descent step

Usage

Sequential$sgd_step(lrate)

Arguments

lrate

the learning rate

---

Method classify()

classify the labels of the input

Usage

Sequential$classify(X)

Arguments

X

input X

---

Method print_accuarcy()

print accuracy during training

Usage

Sequential$print_accuarcy(it, X, Y, cur_loss, every = 250)

Arguments

it

iteration number

X

data X

Y

target Y

cur_loss

the current loss

every

how often should the function return feeddback?

---

Method clone()

The objects of this class are cloneable with this method.

Usage

Sequential$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Note

Note that delta can refer to dLdA or dLdZ over the course of the for loop, depending on the module m

See Also

Other architecture: BatchNorm, Linear

Examples

## Not run: 
## classification problem
nn = Sequential$new(list(
Linear$new(2,10), ReLU$new(),
Linear$new(10,10), Tanh$new(),
Linear$new(10,2), SoftMax$new()),
NLL$new())

data(xor_example)
X = xor_example$X
Y = xor_example$Y

nn$mini_gd(X, Y, 2500, 0.05,  K = 2)


## End(Not run)

frhl/neuralnetr documentation built on Nov. 9, 2020, 2:24 p.m.