Add comments

Author: bpesquet

pytorch_tutorial/convolutional_neural_network/README.md

@@ -54,7 +54,7 @@ conv2d_kernel_size = 3  # Size of the 2D convolution kernels
 
 We use [Fashion-MNIST](https://github.com/zalandoresearch/fashion-mnist), a classic dataset for image recognition. Each example is a 28x28 grayscale image,  associated with a label (fashion category) from 10 possible classes: t-shirt, trouser, pullover...
 
-This dataset is provided by PyTorch through the [FashionMNIST](https://pytorch.org/vision/0.19/generated/torchvision.datasets.FashionMNIST.html) class. In order to evaluate the trained model performance on unseen data, this class lets us download either the training set (60,000 images) or the test set (10,000 images).
+No need to download it manually: this dataset is provided by PyTorch through the [FashionMNIST](https://pytorch.org/vision/0.19/generated/torchvision.datasets.FashionMNIST.html) class, which lets us download either the training set (60,000 images) or the test set (10,000 images).
 
 Alongside download, a [transform](https://pytorch.org/vision/main/transforms.html) operation is applied to turn images into PyTorch tensors of shape `(color_depth, height, width)`, with pixel values scaled to the $[0,1]$ range.
 
@@ -64,7 +64,9 @@ Alongside download, a [transform](https://pytorch.org/vision/main/transforms.htm
 # Directory for downloaded files
 DATA_DIR = "./_output"
 
-# Download and construct the Fashion-MNIST images dataset
+# Download and construct the Fashion-MNIST images dataset.
+# ToTensor() converts a PIL Image or NumPy array (H x W x C) in the range [0, 255]
+# to a PyTorch tensor of shape (C x H x W) in the range [0.0, 1.0].
 # The training set is used to train the model
 train_dataset = datasets.FashionMNIST(
     root=DATA_DIR,

pytorch_tutorial/convolutional_neural_network/images/convolutional_neural_network.png

@@ -70,7 +70,9 @@ def test_convolutional_neural_network(show_plots=False):
     batch_size = 64  # Number of samples used for one gradient descent step
     conv2d_kernel_size = 3  # Size of the 2D convolution kernels
 
-    # Download and construct the Fashion-MNIST images dataset
+    # Download and construct the Fashion-MNIST images dataset.
+    # ToTensor() converts a PIL Image or NumPy array (H x W x C) in the range [0, 255]
+    # to a PyTorch tensor of shape (C x H x W) in the range [0.0, 1.0].
     # The training set is used to train the model
     train_dataset = datasets.FashionMNIST(
         root=DATA_DIR,

pytorch_tutorial/convolutional_neural_network/test_convolutional_neural_network.py

@@ -30,10 +30,10 @@ def test_multilayer_perceptron(show_plots=False):
 
     # Hyperparameters
     n_samples = 500  # Number of data samples
-    hidden_layer_dim = 3  # Number of neurons on the hidden layer of the MLP
     n_epochs = 50  # Number of training iterations on the whole dataset
     learning_rate = 0.1  # Rate of parameter change during gradient descent
     batch_size = 5  # Number of samples used for one gradient descent step
+    hidden_layer_dim = 3  # Number of neurons on the hidden layer of the MLP
 
     # Generate 2D data (a large circle containing a smaller circle)
     inputs, targets = make_circles(n_samples=n_samples, noise=0.1, factor=0.3)

pytorch_tutorial/multilayer_perceptron/test_multilayer_perceptron.py

@@ -173,7 +173,9 @@ def plot_fashion_images(dataset, device, model=None):
         img, label = dataset[sample_idx]
         figure.add_subplot(rows, cols, i)
         plt.axis("off")
-        plt.imshow(img.cpu().detach().numpy().squeeze(), cmap="gray")
+
+        # The color depth channel is removed to obtain a (H x W) tensor ready for plotting
+        plt.imshow(img.squeeze(), cmap="binary")
 
         # Title is either the true or predicted fashion item
         if model is None: