Plot training history

.pylintrc

@@ -18,7 +18,7 @@ good-names=i,j,k,x,y
 max-args = 6
 
 # Maximum number of locals for function / method body
-max-locals = 17
+max-locals = 18
 
 [TYPECHECK]
 

mlcourse/feedforward_neural_networks/test_feedforward_neural_network_fashion_images.py

@@ -3,6 +3,7 @@
 """
 
 import matplotlib.pyplot as plt
+import seaborn as sns
 import torch
 from torch import nn
 from torch.utils.data import DataLoader
@@ -50,6 +51,31 @@ def plot_fashion_images(data, device, model=None):
  plt.show()
 
 
+def plot_loss_acc(history, dataset="Training"):
+ """Plot training loss and accuracy. Takes a Keras-like History object as parameter"""
+
+ loss_values = history["loss"]
+ recorded_epochs = range(1, len(loss_values) + 1)
+
+ fig, (ax1, ax2) = plt.subplots(2, 1)
+ ax1.plot(recorded_epochs, loss_values, ".--", label=f"{dataset} loss")
+ ax1.set_ylabel("Loss")
+ ax1.legend()
+
+ acc_values = history["acc"]
+ ax2.plot(recorded_epochs, acc_values, ".--", label=f"{dataset} accuracy")
+ ax2.set_xlabel("Epochs")
+ ax2.set_ylabel("Accuracy")
+ plt.legend()
+
+ final_loss = loss_values[-1]
+ final_acc = acc_values[-1]
+ fig.suptitle(
+ f"{dataset} loss: {final_loss:.5f}. {dataset} accuracy: {final_acc*100:.2f}%"
+ )
+ plt.show()
+
+
 def fetch_fashion_dataset(data_folder):
  """Download the Fashion-MNIST images dataset"""
 
@@ -118,12 +144,17 @@ def forward(self, x):
 def fit(model, dataloader, criterion, optimizer, n_epochs, device):
  """Train a model on a dataset, using a predefined gradient descent optimizer"""
 
+ # Object storing training history
+ history = {"loss": [], "acc": []}
+
  # Number of samples
  n_samples = len(dataloader.dataset)
 
  # Number of batches in an epoch (= n_samples / batch_size, rounded up)
  n_batches = len(dataloader)
 
+ print(f"Training started! {n_samples} samples. {n_batches} batches per epoch")
+
  # Train the model
  for epoch in range(n_epochs):
  # Total loss for epoch, divided by number of batches to obtain mean loss
@@ -160,6 +191,12 @@ def fit(model, dataloader, criterion, optimizer, n_epochs, device):
  f"Epoch [{(epoch + 1):3}/{n_epochs:3}] finished. Mean loss: {mean_loss:.5f}. Accuracy: {epoch_acc * 100:.2f}%"
  )
 
+ # Record epoch metrics for later plotting
+ history["loss"].append(mean_loss)
+ history["acc"].append(epoch_acc)
+
+ return history
+
 
 def evaluate(model, dataloader, device):
  """Evaluate a model in inference mode"""
@@ -221,12 +258,18 @@ def test_feedforward_neural_network_fashion_images(show_plots=False):
  # Adam optimizer for GD
  optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 
- fit(model, train_loader, criterion, optimizer, n_epochs, device)
+ history = fit(model, train_loader, criterion, optimizer, n_epochs, device)
 
  # Evaluate model performance on test data
  evaluate(model, test_loader, device)
 
  if show_plots:
+ # Improve plots appearance
+ sns.set_theme()
+
+ # Plot training history
+ plot_loss_acc(history)
+
  # Plot model predictions for some test images
  plot_fashion_images(test_dataset, device, model)