Merge pull request #108 from mlverse/examples

Examples

.gitignore

@@ -7,3 +7,5 @@ inst/doc
 docs
 doc
 Meta
+mnist
+pets

R/metrics.R

@@ -206,7 +206,7 @@ luz_metric_loss_average <- luz_metric(
     self$values <- list()
   },
   update = function(preds, targets) {
-    if (length(ctx$loss) == 1)
+    if (length(ctx$loss) == 1 && is.list(ctx$loss))
       loss <- ctx$loss[[1]]
     else
       loss <- ctx$loss

vignettes/examples/.gitignore

@@ -4,3 +4,5 @@ triplet.pt
 mnist-virtual-batch_size.pt
 mnist-autoencoder.pt
 model-dogs-and-cats.pt
+dogs-vs-cats
+mnist

vignettes/examples/dogs-vs-cats-binary-classification.Rmd

@@ -16,12 +16,11 @@ torch_manual_seed(1)
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/dogs-vs-cats" # caching directory
-kaggle_token <- "~/Downloads/kaggle.json"
+dir <- "./dogs-vs-cats" # caching directory
 
 ds <- torchdatasets::dogs_vs_cats_dataset(
   dir,
-  token = kaggle_token,
+  download = TRUE,
   transform = . %>%
     torchvision::transform_to_tensor() %>%
     torchvision::transform_resize(size = c(224, 224)) %>% 
@@ -65,7 +64,7 @@ net <- torch::nn_module(
 fitted <- net %>%
   setup(
     loss = nn_bce_with_logits_loss(),
-    optimizer = madgrad::optim_madgrad,
+    optimizer = optim_adam,
     metrics = list(
       luz_metric_binary_accuracy_with_logits()
     )

vignettes/examples/mnist-autoencoder.Rmd

@@ -12,7 +12,7 @@ library(luz)
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/mnist" # caching directory
+dir <- "./mnist" # caching directory
 
 # Modify the MNIST dataset so the target is identical to the input.
 mnist_dataset2 <- torch::dataset(

vignettes/examples/mnist-cnn-virtual-batch-size.Rmd

@@ -5,15 +5,14 @@ category: 'advanced'
 ---
 
 ```{r, eval = FALSE}
-
 # Packages ----------------------------------------------------------------
 library(torch)
 library(torchvision)
 library(luz)
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/mnist" # caching directory
+dir <- "./mnist" # caching directory
 
 train_ds <- mnist_dataset(
   dir,
@@ -38,8 +37,8 @@ net <- nn_module(
   initialize = function(accumulate_batches = 2) {
     self$conv1 <- nn_conv2d(1, 32, 3, 1)
     self$conv2 <- nn_conv2d(32, 64, 3, 1)
-    self$dropout1 <- nn_dropout2d(0.25)
-    self$dropout2 <- nn_dropout2d(0.5)
+    self$dropout1 <- nn_dropout(0.25)
+    self$dropout2 <- nn_dropout(0.5)
     self$fc1 <- nn_linear(9216, 128)
     self$fc2 <- nn_linear(128, 10)
 
@@ -60,14 +59,25 @@ net <- nn_module(
     x
   },
   step = function() {
-    ctx$pred <- ctx$model(ctx$input)
-    loss <- ctx$model$loss(ctx$pred, ctx$target)
-
+    # we implement a custom step method that runs for every 
+    # batch in training and validation.
+    
+    # calculate predictions. we save them in `ctx$pred` so other parts of luz
+    # can use it.
+    ctx$pred <- ctx$model(ctx$input) 
+    
+    # we now calculate the loss. also save it in `ctx$loss` so, for example,
+    # it's correctly logged.
+    ctx$loss <- ctx$model$loss(ctx$pred, ctx$target)
+
+    # `ctx$training` is set automatically to `TRUE` during the training phase
     if (ctx$training) {
-      loss <- loss/self$accumulate_batches
-      loss$backward()
+      ctx$loss <- ctx$loss/self$accumulate_batches
+      ctx$loss$backward()
     }
 
+    # only after `accumulate_batches` that we do a optimizer step, so we use 
+    # the virtual batch_size.
     if (ctx$training && (ctx$iter %% self$accumulate_batches == 0)) {
       opt <- ctx$optimizers[[1]]
       opt$step()
@@ -87,7 +97,7 @@ fitted <- net %>%
       luz_metric_accuracy()
     )
   ) %>%
-  fit(train_dl, epochs = 10, valid_data = test_dl)
+  fit(train_dl, valid_data = test_dl, epochs = 10)
 
 
 # Serialization -----------------------------------------------------------

vignettes/examples/mnist-cnn.Rmd

@@ -16,7 +16,7 @@ torch_manual_seed(1)
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/mnist" # caching directory
+dir <- "./mnist" # caching directory
 
 train_ds <- mnist_dataset(
   dir,
@@ -41,8 +41,8 @@ net <- nn_module(
   initialize = function() {
     self$conv1 <- nn_conv2d(1, 32, 3, 1)
     self$conv2 <- nn_conv2d(32, 64, 3, 1)
-    self$dropout1 <- nn_dropout2d(0.25)
-    self$dropout2 <- nn_dropout2d(0.5)
+    self$dropout1 <- nn_dropout(0.25)
+    self$dropout2 <- nn_dropout(0.5)
     self$fc1 <- nn_linear(9216, 128)
     self$fc2 <- nn_linear(128, 10)
   },

vignettes/examples/mnist-dcgan.Rmd

@@ -34,7 +34,7 @@ plot_callback <- luz_callback(
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/mnist" # caching directory
+dir <- "./mnist" # caching directory
 
 train_ds <- mnist_dataset(
   dir,
@@ -153,9 +153,8 @@ dcgan <- torch::nn_module(
   }
 )
 
-dcgan <- setup(dcgan)
-
 res <- dcgan %>%
+  setup() %>% 
   set_hparams(latent_dim = 100, channels = 1) %>%
   fit(train_dl, epochs = 10, valid_data = test_dl, callbacks = list(plot_callback()))
 
@@ -172,6 +171,5 @@ plot(img)
 # Serialization ----------------------------------------------------
 
 luz_save(res, "mnist-dcgan.pt")
-
 ```
 

vignettes/examples/mnist-mixup.Rmd

@@ -0,0 +1,89 @@
+---
+title: 'MixUp augmentation'
+desc: 'Demonstrates using the MixUp callback'
+category: 'intermediate'
+---
+
+```{r, eval = FALSE}
+# Packages ----------------------------------------------------------------
+library(torch)
+library(torchvision)
+library(luz)
+
+set.seed(1)
+torch_manual_seed(1)
+
+# Datasets and loaders ----------------------------------------------------
+
+dir <- "./mnist" # caching directory
+
+train_ds <- mnist_dataset(
+  dir,
+  download = TRUE,
+  transform = transform_to_tensor
+)
+
+test_ds <- mnist_dataset(
+  dir,
+  train = FALSE,
+  transform = transform_to_tensor
+)
+
+train_dl <- dataloader(train_ds, batch_size = 128, shuffle = TRUE)
+test_dl <- dataloader(test_ds, batch_size = 128)
+
+
+# Building the network ---------------------------------------------------
+
+net <- nn_module(
+  "Net",
+  initialize = function() {
+    self$conv1 <- nn_conv2d(1, 32, 3, 1)
+    self$conv2 <- nn_conv2d(32, 64, 3, 1)
+    self$dropout1 <- nn_dropout(0.25)
+    self$dropout2 <- nn_dropout(0.5)
+    self$fc1 <- nn_linear(9216, 128)
+    self$fc2 <- nn_linear(128, 10)
+  },
+  forward = function(x) {
+    x %>% 
+      self$conv1() %>% 
+      nnf_relu() %>% 
+      self$conv2() %>% 
+      nnf_relu() %>% 
+      nnf_max_pool2d(2) %>% 
+      self$dropout1() %>% 
+      torch_flatten(start_dim = 2) %>% 
+      self$fc1() %>% 
+      nnf_relu() %>% 
+      self$dropout2() %>% 
+      self$fc2()
+  }
+)
+
+# Train -------------------------------------------------------------------
+
+fitted <- net %>%
+  setup(
+    loss = nn_mixup_loss(nn_cross_entropy_loss()),
+    optimizer = optim_adam
+  ) %>%
+  fit(
+    train_dl, epochs = 10, valid_data = test_dl, 
+    callbacks = list(
+      luz_callback_mixup(alpha = 0.4)
+    )
+  )
+
+# Making predictions ------------------------------------------------------
+
+preds <- predict(fitted, test_dl)
+preds$shape
+
+
+# Serialization -----------------------------------------------------------
+
+luz_save(fitted, "mnist-mixup.pt")
+```
+
+

vignettes/examples/mnist-triplet.Rmd

@@ -12,7 +12,7 @@ library(luz)
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/mnist" # caching directory
+dir <- "./mnist" # caching directory
 
 triplet_mnist_dataset <- dataset(
   inherit = mnist_dataset,
@@ -56,8 +56,8 @@ net <- nn_module(
   initialize = function(embedding_dim) {
     self$conv1 <- nn_conv2d(1, 32, 3, 1)
     self$conv2 <- nn_conv2d(32, 64, 3, 1)
-    self$dropout1 <- nn_dropout2d(0.25)
-    self$dropout2 <- nn_dropout2d(0.5)
+    self$dropout1 <- nn_dropout(0.25)
+    self$dropout2 <- nn_dropout(0.5)
     self$fc1 <- nn_linear(9216, 512)
     self$fc2 <- nn_linear(512, embedding_dim)
   },
@@ -100,6 +100,5 @@ fitted <- triplet_model %>%
 # Serializing
 
 luz_save(fitted, "triplet.pt")
-
 ```
 

vignettes/examples/pets-unet.Rmd

@@ -15,7 +15,7 @@ library(luz)
 
 # Datasets and loaders ----------------------------------------------------
 
-dir <- "~/Downloads/" #caching directory
+dir <- "./pets" #caching directory
 
 # A light wrapper around the `oxford_pet_dataset` that resizes and transforms 
 # input images and masks to the specified `size` and introduces the `augmentation`
@@ -73,8 +73,6 @@ valid_ds <- pet_dataset(
 train_dl <- dataloader(train_ds, batch_size = 32, shuffle = TRUE)
 valid_dl <- dataloader(valid_ds, batch_size = 32)
 
-x <- coro::collect(train_dl, 1)
-
 # Define the network ------------------------------------------------------
 
 # We use a pre-trained mobile net encoder. We take intermediate layers to use