mnist_fashion_inference

Source: https://www.kaggle.com/ysachit/inference-and-validation-ipynb

Original title: Inference and Validation

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
library(rTorch)
nn         <- torch$nn
transforms <- torchvision$transforms
datasets   <- torchvision$datasets
builtins   <- import_builtins()

torch$manual_seed(123)

Load datasets

As usual, let's start by loading the dataset through torchvision.

local_folder <- "../datasets/mnist_fashion"
# Define a transform to normalize the data
transform = transforms$Compose(list(transforms$ToTensor(),
                                transforms$Normalize(list(0.5), list(0.5) )))

# Download and load the training data
trainset = datasets$FashionMNIST(local_folder, download=TRUE, train=TRUE,  
        transform=transform)
train_loader = torch$utils$data$DataLoader(trainset, batch_size=64L, shuffle=TRUE)

# Download and load the test data
testset = datasets$FashionMNIST(local_folder, download=TRUE, train=FALSE,
    transform=transform)
test_loader = torch$utils$data$DataLoader(testset, batch_size=64L, shuffle=TRUE)

py_len(trainset)

Model

optim <- torch$optim
F     <- torch$nn$functional

main <- py_run_string("
from torch import nn, optim
import torch.nn.functional as F

class Classifier(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(784, 256)
        self.fc2 = nn.Linear(256, 128)
        self.fc3 = nn.Linear(128, 64)
        self.fc4 = nn.Linear(64, 10)

    def forward(self, x):
        # make sure input tensor is flattened
        x = x.view(x.shape[0], -1)

        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = F.relu(self.fc3(x))
        x = F.log_softmax(self.fc4(x), dim=1)

        return x
")


model = main$Classifier()
model

Instrospection

random <- import("random")

# number of elements in test
py_len(test_loader)    # len: 157

# get random object from the test_loader
i_rand = random$randrange(0, py_len(test_loader)-1)   # get a random data point
i_rand
enum_test_examples = builtins$enumerate(test_loader)
enum_test_examples
# <enumerate object at 0x7f793ce486c0>

iter_test_examples <- iterate(enum_test_examples)

# iter_test_examples[[3]]
iter_test_examples[[i_rand]][[1]]   # index number
iter_test_examples[[i_rand]][[2]][[1]]$size()  # image tensor .Size([64,1,28,28])
iter_test_examples[[i_rand]][[2]][[2]]$shape    # labels torch.Size([64])

idx    <-  iter_test_examples[[i_rand]][[1]]
images <- iter_test_examples[[i_rand]][[2]][[1]]
labels <- iter_test_examples[[i_rand]][[2]][[2]]

# Get the class probabilities
ps = torch$exp(model(images))

# Make sure the shape is appropriate, we should get 10 class probabilities for 64 examples
print(ps$shape)
# torch.Size([64, 10])

Most likely class

With the probabilities, we can get the most likely class using the ps.topk method. This returns the k highest values. Since we just want the most likely class, we can use ps.topk(1). This returns a tuple of the top-k values and the top-k indices. If the highest value is the fifth element, we'll get back 4 as the index.

top_ = ps$topk(1L, dim=1L)

top_p     <- top_[0]
top_class <- top_[1]

top_p$shape
top_class$shape

# Look at the most likely classes for the first 10 examples
print(top_class[1:10,])

Compare predicted vs true labels

Now we can check if the predicted classes match the labels. This is simple to do by equating top_class and labels, but we have to be careful of the shapes. Here top_class is a 2D tensor with shape (64, 1) while labels is 1D with shape (64). To get the equality to work out the way we want, top_class and labels must have the same shape.

equals will have shape (64, 64), try it yourself. What it's doing is comparing the one element in each row of top_class with each element in labels which returns 64 True/False boolean values for each row.

equals = top_class == labels
equals$shape
equals = top_class == labels$view(top_class$shape)
equals

Untrained model

Now we need to calculate the percentage of correct predictions. equals has binary values, either 0 or 1. This means that if we just sum up all the values and divide by the number of values, we get the percentage of correct predictions.

we'll need to convert equals to a float tensor. Note that when we take torch.mean it returns a scalar tensor, to get the actual value as a float we'll need to do accuracy.item().

accuracy = torch$mean(equals$type(torch$FloatTensor))
cat(sprintf("Accuracy: %f %%", accuracy$item()*100))

The network is untrained so it's making random guesses and we should see an accuracy around 10%.

Train the model

Now let's train our network and include our validation pass so we can measure how well the network is performing on the test set. Since we're not updating our parameters in the validation pass, we can speed up our code by turning off gradients using torch.no_grad():

model     <- main$Classifier()
criterion <- nn$NLLLoss()
optimizer <- optim$Adam(model$parameters(), lr = 0.003)

epochs <- 5
steps  <- 0

train_losses <- vector("list"); test_losses  <- vector("list")
for (e in 1:epochs) {
  i <- 1    # counter for training loop
  running_loss <- 0
  iter_train_dataset <- builtins$enumerate(train_loader) # reset iterator
  for (train_obj in iterate(iter_train_dataset)) {
      images <- train_obj[[2]][[1]] # extract images
      labels <- train_obj[[2]][[2]] # extract labels

      optimizer$zero_grad()

      log_ps <- model(images)
      loss <- criterion(log_ps, labels)
      loss$backward()
      optimizer$step()

      running_loss <- running_loss + loss$item()  
  }
  test_loss <- 0
  accuracy  <- 0

  with(torch$no_grad(), {
    iter_test_dataset <- builtins$enumerate(test_loader) # reset iterator
    for (test_obj in iterate(iter_test_dataset)) {
      images <- test_obj[[2]][[1]]  # extract images
      labels <- test_obj[[2]][[2]]  # extract labels
      output <- model(images)
      test_loss <- test_loss + criterion(output, labels)

      ps       <- torch$exp(model(images))
      top_     <- ps$topk(1L, dim=1L)
      top_p    <- top_[0]; top_class <- top_[1]
      equals   <- top_class == labels$view(top_class$shape)
      accuracy <- accuracy + torch$mean(equals$type(torch$FloatTensor))
      # Look at the most likely classes for the first 10 examples
    }
  })
  train_losses[[i]] <- running_loss / py_len(train_loader)
  test_losses[[i]]  <- test_loss / py_len(test_loader)
  cat(sprintf("\n Epoch: %3d Training Loss: %8.3f Test Loss: %8.3f Test Accuracy: %8.3f", 
              e,
              running_loss / py_len(train_loader), 
              test_loss$item() / py_len(test_loader),
              accuracy$item() / py_len(test_loader)       
  ))
}

Overfitting

If we look at the training and validation losses as we train the network, we can see a phenomenon known as overfitting.

The network learns the training set better and better, resulting in lower training losses. However, it starts having problems generalizing to data outside the training set leading to the validation loss increasing. The ultimate goal of any deep learning model is to make predictions on new data, so we should strive to get the lowest validation loss possible. One option is to use the version of the model with the lowest validation loss, here the one around 8-10 training epochs. This strategy is called early-stopping. In practice, you'd save the model frequently as you're training then later choose the model with the lowest validation loss.

Model with dropout

The most common method to reduce overfitting (outside of early-stopping) is dropout, where we randomly drop input units. This forces the network to share information between weights, increasing it's ability to generalize to new data. Adding dropout in PyTorch is straightforward using the nn.Dropout module.

main <- py_run_string("
from torch import nn, optim      
import torch.nn.functional as F

class ClassifierDO(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(784, 256)
        self.fc2 = nn.Linear(256, 128)
        self.fc3 = nn.Linear(128, 64)
        self.fc4 = nn.Linear(64, 10)

        # Dropout module with 0.2 drop probability
        self.dropout = nn.Dropout(p=0.2)

    def forward(self, x):
        # make sure input tensor is flattened
        x = x.view(x.shape[0], -1)

        # Now with dropout
        x = self.dropout(F.relu(self.fc1(x)))
        x = self.dropout(F.relu(self.fc2(x)))
        x = self.dropout(F.relu(self.fc3(x)))

        # output so no dropout here
        x = F.log_softmax(self.fc4(x), dim=1)

        return x
")

ClassifierDO <- main$ClassifierDO()
ClassifierDO

Train the model with dropout

During training we want to use dropout to prevent overfitting, but during inference we want to use the entire network. So, we need to turn off dropout during validation, testing, and whenever we're using the network to make predictions. To do this, you use model.eval(). This sets the model to evaluation mode where the dropout probability is 0. You can turn dropout back on by setting the model to train mode with model.train(). In general, the pattern for the validation loop will look like this, where you turn off gradients, set the model to evaluation mode, calculate the validation loss and metric, then set the model back to train mode.

modelDO     <- main$ClassifierDO()
criterion   <- nn$NLLLoss()
optimizerDO <- optim$Adam(modelDO$parameters(), lr = 0.003)

epochs <- 5
steps  <- 0

train_losses <- vector("list"); test_losses  <- vector("list")
for (e in 1:epochs) {
  i <- 1    # counter for training loop
  running_loss <- 0
  iter_train_dataset <- builtins$enumerate(train_loader) # reset iterator
  for (train_obj in iterate(iter_train_dataset)) {
      images <- train_obj[[2]][[1]] # extract images
      labels <- train_obj[[2]][[2]] # extract labels

      optimizerDO$zero_grad()

      log_ps <- modelDO(images)
      loss <- criterion(log_ps, labels)
      loss$backward()
      optimizerDO$step()

      running_loss <- running_loss + loss$item()  
  }
  test_loss <- 0
  accuracy  <- 0

  with(torch$no_grad(), {
    iter_test_dataset <- builtins$enumerate(test_loader)    # reset iterator
    for (test_obj in iterate(iter_test_dataset)) {
      images <- test_obj[[2]][[1]]  # extract images
      labels <- test_obj[[2]][[2]]  # extract labels
      output <- modelDO(images)
      test_loss <- test_loss + criterion(output, labels)

      ps       <- torch$exp(modelDO(images))
      top_     <- ps$topk(1L, dim=1L)
      top_p    <- top_[0]; top_class <- top_[1]
      equals   <- top_class == labels$view(top_class$shape)
      accuracy <- accuracy + torch$mean(equals$type(torch$FloatTensor))
      # Look at the most likely classes for the first 10 examples
    }
  })
  train_losses[[i]] <- running_loss / py_len(train_loader)
  test_losses[[i]]  <- test_loss / py_len(test_loader)
  cat(sprintf("\n Epoch: %3d Training Loss: %8.3f Test Loss: %8.3f Test Accuracy: %8.3f", 
              e,
              running_loss / py_len(train_loader), 
              test_loss$item() / py_len(test_loader),
              accuracy$item() / py_len(test_loader)       
  ))
}

Inference on dropout model

Now that the model is trained, we can use it for inference. We've done this before, but now we need to remember to set the model in inference mode with model.eval(). You'll also want to turn off autograd with the torch.no_grad() context.

# Test out your network!

# Switch to evaluation mode
modelDO$eval()

# load test dataset and get one image
# using an iterator
dataiter = builtins$iter(test_loader)      # len(test_loader): 157
data_obj = iter_next(dataiter)  

images <- data_obj[[1]]       # images.shape: torch.Size([64, 1, 28, 28])
labels <- data_obj[[2]]       # labels.shape: torch.Size([64])
images$shape
labels$shape

# take first image
img = images[1L,,,]  # shape:  torch.Size([1, 28, 28])

# Convert 2D image to 1D vector
img_pick = img$view(1L, 784L)   # shape: torch.Size([1, 784])
img_pick$shape

# Calculate the class probabilities (softmax) for img
with(torch$no_grad(), {
    output = modelDO$forward(img_pick)
})

ps = torch$exp(output)
ps$shape


# Plot the image and probabilities
  # view_classify(img_pick.view(1, 28, 28), ps, version='Fashion')

# show image
rotate <- function(x) t(apply(x, 2, rev))   # function to rotate the matrix
img_np_rs = np$reshape(img_pick$numpy(), c(28L, 28L))
image(rotate(img_np_rs))
# show barplot with class probability


Try the rTorch package in your browser

Any scripts or data that you put into this service are public.

rTorch documentation built on Jan. 13, 2021, 4:32 p.m.