---
title: "MNIST Example"
author: "Prof. Eric A. Suess"
format: 
  html:
    embed-resources: true
---

## MNIST Feedforward neural network with **one** hidden layer

Load the packages

```{r}
library(tensorflow)
library(keras)
```

Load the MNIST dataset.

```{r}
mnist <- dataset_mnist()
train_images <- mnist$train$x
train_labels <- mnist$train$y
test_images <- mnist$test$x
test_labels <- mnist$test$y
```

Examine the data.

```{r}
str(train_images)
str(train_labels)

str(test_images)
str(test_labels)

```

Preprocess the data.

```{r}
train_images <- array_reshape(train_images, c(60000, 28 * 28))
train_images <- train_images / 255
test_images <- array_reshape(test_images, c(10000, 28 * 28))
test_images <- test_images / 255
```

Specify a neural network model.

```{r}
model <- keras_model_sequential(list(
  layer_dense(units = 512, activation = "relu"),
  layer_dense(units = 10, activation = "softmax")
))
```


```{r}
## -------------------------------------------------------------------------
compile(model,
        optimizer = "rmsprop",
        loss = "sparse_categorical_crossentropy",
        metrics = "accuracy")

```

Train the model.

```{r}
fit(model, train_images, train_labels, epochs = 5, batch_size = 128)
```

Make predictions.  Inference.

```{r}
test_digits <- test_images[1:10, ]
predictions <- predict(model, test_digits)
str(predictions)
predictions[1, ]

which.max(predictions[1, ])
predictions[1, 8]

test_labels[1]

```

Evaluate the model.

```{r}
metrics <- evaluate(model, test_images, test_labels)
metrics["accuracy"]
```

## MNIST Feedforward neural network with **two** hidden layers

Reload the images.

```{r}
mnist <- dataset_mnist()
train_images <- mnist$train$x
train_labels <- mnist$train$y
test_images <- mnist$test$x
test_labels <- mnist$test$y
```


View an image.

```{r}
digit <- train_images[1, , ]
par(mar = c(0, 0, 0, 0))
plot(as.raster(abs(255 - digit), max = 255))
```

```{r}
train_labels[1]
```

```{r}
digit <- test_images[1, , ]
par(mar = c(0, 0, 0, 0))
plot(as.raster(abs(255 - digit), max = 255))
```

```{r}
test_labels[1]
```

Preprocess the data.

```{r}
train_images <- array_reshape(train_images, c(60000, 28 * 28))
train_images <- train_images / 255

test_images <- array_reshape(test_images, c(10000, 28 * 28))
test_images <- test_images / 255

train_labels <- mnist$train$y
test_labels <- mnist$test$y
```


New model.

```{r}
model <- keras_model_sequential(list(
  layer_dense(units = 512, activation = "relu"),
  layer_dense(units = 10, activation = "softmax")
))
```

```{r}
compile(model,
        optimizer = "rmsprop",
        loss = "sparse_categorical_crossentropy",
        metrics = c("accuracy"))
```

Train the model.

```{r}
fit(model, train_images, train_labels, epochs = 5, batch_size = 128)
```

Inference

```{r}
predictions <- predict(model, test_images)
str(predictions)
predictions[1, ]

which.max(predictions[1, ]) - 1
predictions[1, 8]

test_labels[1]


predicted_labels <- max.col(predictions) - 1
matches <- predicted_labels == test_labels
cat(sprintf("accuracy: %.2f\n", mean(matches)))

```