NN_demo.Rmd

---
title: "NN_demo"
output: html_document
date: "2024-03-14"
---
```{r keras}
install.packages("keras")
keras::install_keras()
```

```{r sim_lin}
library(tidyverse)
n <- 1000
b <- 30
a <- 5

set.seed(123)
sim <- tibble(
  x = runif(n, min = -1, max = 1),
  y = b +a*x + rnorm(n)
)
```

```{r plot_lin}
ggplot(sim, aes(x,y)) +
  geom_point()+
  geom_smooth(method = lm)
```

```
```{r fit_lin}
ols_model <- lm(y ~ x, sim)

sigma(ols_model)

ols_model
```


```{r mod}
x <- matrix(sim$x, ncol = 1)

library(keras)

mod <- keras_model_sequential() %>%
  layer_dense(units = 1,
              activation = "linear",
              input_shape = ncol(x))
```

```{r mod_compile}
mod %>%
  compile(optimizer = "sgd", #stochastic gradient descent
          loss = "mse") #mean square error
```

```{r mod_hist}
hist <- mod %>%
  fit(x, sim$y, #data
      batch_size = 16, #mini-batch size
      epochs = 20, # n times through full training data
      validation_split = .2 
  )
```


```{r compare_lin}

plot(hist)

keras::get_weights(mod)
coef(ols_model)

history
sigma(ols_model)
```

```{r sim_sine}
set.seed(321)
df <- tibble(
  x = seq(from = -1, to = 2 * pi, length = n),
  e = rnorm(n, sd = 0.2),
  y = sin(x) + e
)

x <- matrix(sim$x, ncol= 1)
```