logistica_com_keras

library(keras)
library(dplyr)

logit <- function(p) log(p) - log(1 - p)
inv_logit <- function(x) 1/(1 + exp(-x))

n <- 10000
set.seed(19880923)
df <- data.frame(x = rnorm(n)) %>%
  mutate(y_1 = rbinom(n, 1, prob = inv_logit(-1 + 4 * x)) %>% as.factor,
         y_2 = rbinom(n, 1, prob = inv_logit(-1 + 1 * inv_logit(-3 * 3 * x))) %>% as.factor)

# modelo glm 1 ------------------------------------------------------
modelo_lm_1 <- glm(y_1 ~ x, data = df, family = binomial)

# coefficients
coef(modelo_lm_1)

# accuracy
conf_matrix_lm_1 <- table(modelo_lm_1$fitted.values > 0.5, df$y_1)
sum(diag(conf_matrix_lm_1))/sum(conf_matrix_lm_1)

# modelo keras 1 -------------------------------------------------------
modelo_keras_1 <- keras_model_sequential(name = "modelo_keras_1")  %>% 
  layer_dense(units = 1, input_shape = 1, name = "camada_unica", use_bias = TRUE, bias_constraint = NULL, bias_regularizer = NULL, bias_initializer = "RandomNormal") %>%
  layer_activation( activation = 'softmax', input_shape = 1, name = "softmax")

summary(modelo_keras_1)

modelo_keras_1 %>% compile(
  loss = 'binary_crossentropy',
  optimizer = optimizer_sgd(lr = 1),
  metrics = c('accuracy')
)

modelo_keras_1_fit <- modelo_keras_1 %>% fit(
  x = df$x, 
  y = df$y_1, 
  epochs = 10, 
  batch_size = 128
)

  # coefficients
modelo_keras_1 %>% get_layer("camada_unica") %>% get_weights

# accuracy
loss_and_metrics_1 <- modelo_keras_1 %>% evaluate(df$x, df$y_1)
loss_and_metrics_1[[2]]