---
title: 'Lab #8: Neural networks'
output:
  html_document:
    df_print: paged
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
set.seed(24)
```

This lab is closely modeled after the neural network tutorial presented in the Medium post by the PyTorch team: https://medium.com/pytorch/please-allow-me-to-introduce-myself-torch-for-r-9ea0f361ea7e.

1. Install and load the torch package.
```{r installTorch}
# install.packages("torch")
library(torch)
```

2. Check that you can make a 1-D tensor with value 1 using the command `torch_tensor(1)`.
```{r}
```

3. Install and load the `torchvision` package as well.
```{r installTorchVision}
# install.packages("torchvision")
library(torchvision)
```

4. Download the mnist dataset by running the following code. The mnist dataset is composed of black and white images of digits 0-9, where each image is 28x28 pixels. Each observation is labeled with the true label. The digits are evenly distributed in both the training and testing datasets.
```{r mnist}
train_ds <- mnist_dataset(
  ".",
  download = FALSE,
  train = TRUE,
  transform = transform_to_tensor
)

test_ds <- mnist_dataset(
  ".",
  download = FALSE,
  train = FALSE,
  transform = transform_to_tensor
)
```

5. How many train and test observations are there? We'll only use the first 3200 observations in each dataset. Subset the data using the `dataset_subset` function.
```{r}
```

6. You can access the first observation in the training dataset using the typical numeric indexing: `train_ds[1]`. What is the dimension of the first observation? What is the label of this first observation?
```{r}
```

7. To train a model in torch, you need to first wrap the subsetted data within a `dataloader`. Do this using the following commands.
```{r}
```

8. You can access the data in the data loader using its `iter` attribute.
```{r}
```

9. Let's plot the first 32 images that we retrieve from the training-set dataloader.
```{r}
```

10. Let's create a neural network with 2 convolutional layers, one max pooling layer, and one fully connected layer. Use the Relu activation function for the hidden nodes.
```{r}
```

11. To instantiate a neural network, you need to call `net()`.
```{r}
```

12. Create an optimizer using the Adam algorithm to minimize the training loss. Do this by running the following code.
```{r}
```

13. Train the neural network for 4 epochs.
```{r}
```

14. Set the trained neural network in evaluation/testing mode using the following code.
```{r}
```

15. Evaluate your trained neural network on the test data. What is the test accuracy?
```{r}
```

16. Is this better than random guessing? How would you further improve the model?