A_Tensorboard.Rmd

---
title: "Atelier #A - Survol de Tensorboard"
author: "Mikaël SWAWOLA et Éric HAMEL"
date: "14/5/2019"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
Sys.setenv(TF_CPP_MIN_LOG_LEVEL = "3") # Car TensorFlow est extrèmement bavard...
```

```{r libraries}
library(png)
library(keras)
library(tensorflow)
use_implementation("tensorflow")
```

```{r}
K <- backend()
K$clear_session()
```



## 1. Visualiser l'entraînement et le graph

```{r}
fashion_mnist <- dataset_fashion_mnist()

c(train_images, train_labels) %<-% fashion_mnist$train
c(test_images, test_labels) %<-% fashion_mnist$test

class_names <- c('T-shirt/top',
                'Trouser',
                'Pullover',
                'Dress',
                'Coat', 
                'Sandal',
                'Shirt',
                'Sneaker',
                'Bag',
                'Ankle boot')

train_images_norm <- train_images / 255
test_images_norm <- test_images / 255
```


```{r}
model <- keras_model_sequential() %>%
  layer_flatten(input_shape = c(28, 28)) %>%
  layer_dense(units = 128, activation = 'relu') %>%
  layer_dense(units = 10, activation = 'softmax')
  
model %>% compile(
    optimizer = optimizer_adam(lr=0.001), 
    loss = 'sparse_categorical_crossentropy',
    metrics = c('accuracy')
  )

model %>% fit(train_images_norm, train_labels,
    batch_size=32,
    epochs=25,
    callbacks = callback_tensorboard("logs/demo_1", write_images = TRUE, histogram_freq = 2),
    validation_split = 0.1
    )

```



## 2. Visualiser les projections du jeu de données

Ce code, donné à titre indicatif, est pour le moment non fonctionnel en R. Pour la démo, les logs de Tensorboard correspondants ont été créés en Python.

```{r}
embedding_var <- tf$Variable(test_images, name='fmnist_embedding')
summary_writer <- tf$summary$FileWriter('logs')

config <- tf$contrib$tensorboard$plugins$projector$ProjectorConfig()
config$model_checkpoint_path = "logs/demo_2/model.ckpt"
embedding <- config$embeddings$add()
embedding$tensor_name <- embedding_var$name

embedding$metadata_path = 'logs/demo_2/metadata.tsv'

embedding$sprite$image_path = 'logs/demo_2/sprite.png' # os.path.join(logdir, 'sprite.png')
embedding$sprite$single_image_dim$extend(c(as.integer(28), as.integer(28)))

tf$contrib$tensorboard$plugins$projector$visualize_embeddings(summary_writer,config)
```

```{r}
sesh <- tf$Session()

sesh$run(tf$global_variables_initializer())
saver <- tf$train$Saver()
saver$save(sess = sesh, save_path = "logs/demo_2/model.ckpt")
sesh$close()
```



### Création de l'image des Sprites

```{r}
rows <- 28
cols <- 28

sprite_dim <- as.integer(sqrt(dim(test_images)[1]))
sprite_image <- matrix(1, ncol = cols*sprite_dim, nrow = rows*sprite_dim)

index <- 1
labels <- c()
for(i in 1:sprite_dim){
  for(j in 1:sprite_dim){
    
    labels <- c(labels, class_names[test_labels[index]+1])
    sprite_image[
      as.integer((i-1) * cols+1): as.integer(i * cols),
      as.integer((j-1) * rows + 1): as.integer(j * rows)
    ] <- test_images[index,,] * - 1 + 1

    index <- index + 1
  }
}

rotate <- function(x) t(apply(x, 2, rev))#

image <- rotate(sprite_image)

writePNG(image, target = 'logs/demo_2/sprite.png')

```



### Création du fichier de metadata

```{r}
fileConn<-file("logs/demo_2/metadata.tsv")
writeLines(paste(as.character(1:10000), as.character(test_labels), sep="\t"), fileConn)
close(fileConn)
```



### Pour aller dans Tensorboard

```{bash}
tensorboard --logdir=logs/demo_2/
```

Ou bien:

```{r}
tensorboard( log_dir = "logs/demo_2/")
```