Welcome back to the Advanced Neural Tuning course. In the last lesson, you explored how the choice of optimizer can impact your neural network’s training. Now, we will focus on another key aspect of building effective neural networks: weight initialization.
When you create a neural network, each layer has weights and biases that start with some initial values. These values are important because they set the starting point for learning. If the weights are not initialized well, your model might struggle to learn, or it could even fail to train at all. For example, poor initialization can cause gradients to vanish (become too small) or explode (become too large) as they move through the network. This can make training very slow or unstable. Good initialization helps your model start off on the right foot, making learning smoother and more reliable.
In PyTorch, each layer in your neural network, such as nn.Linear, has its own set of weights and biases. These are stored as parameters inside the layer. When you define a model, you can access these parameters directly. For example, if you have a linear layer called layer, you can access its weights with layer.weight and its biases with layer.bias. These parameters are what the optimizer updates during training.
When you build a model using nn.Module, you can loop through all the layers and access their parameters. This is useful when you want to apply a specific initialization method to every layer of a certain type, such as all fully connected (nn.Linear) layers in your model.
One of the most popular ways to initialize weights in fully connected layers is called Xavier initialization, also known as Glorot initialization. This method was designed to keep the scale of the gradients roughly the same in all layers, which helps prevent the vanishing and exploding gradient problems mentioned earlier.
Xavier initialization works by setting the initial weights so that the variance of the outputs of each layer is the same as the variance of its inputs. In practice, this means the weights are drawn from a uniform or normal distribution with a specific range, depending on the number of input and output units in the layer.
The key to Xavier initialization is the use of two values: fan_in and fan_out.
- fan_in is the number of input units (neurons) to a layer.
- fan_out is the number of output units (neurons) from a layer.
For a fully connected (nn.Linear) layer in PyTorch, if the layer is defined as nn.Linear(in_features, out_features), then:
fan_in = in_featuresfan_out = out_features
These values matter because they determine how much signal flows into and out of each neuron. If the weights are too large relative to fan_in or fan_out, the outputs can become too large (exploding gradients). If they are too small, the outputs can shrink towards zero (vanishing gradients).
For Xavier uniform initialization, the weights are drawn from a uniform distribution bounded by:
Let’s look at how you can apply Xavier initialization to all the fully connected layers in your model using PyTorch. Suppose you have a model with several nn.Linear layers. You can loop through each layer, check if it is a linear layer, and then apply Xavier uniform initialization to its weights and set its biases to zero. Here is how you can do it:
In this code, you loop through all the modules in your model using model.modules(). For each layer, you check if it is an instance of nn.Linear. If it is, you apply Xavier uniform initialization to the layer’s weights using nn.init.xavier_uniform_, and you set the biases to zero with nn.init.zeros_.
Setting biases to zero is a common and safe default for most layers. This is because, at the start of training, you want the output of each neuron to be determined only by the weighted sum of its inputs. A nonzero bias could introduce an unintended offset before learning begins. Since the optimizer will quickly learn the appropriate bias values during training, starting with zero biases helps keep the initial activations balanced and avoids introducing unnecessary asymmetry.
After running the initialization code, you can print the weights and biases of a linear layer to verify the initialization:
If you run this code, you might see output like this:
This output shows that the weights have been set to values drawn from the Xavier uniform distribution (with bounds determined by fan_in and fan_out), and the biases have been set to zero.
After initializing your weights and biases, it is a good idea to check that the values have been set as expected. You can do this by printing out a few of the weights and biases from your model. For example, after running the initialization code, you can print the weights and biases of the first linear layer like this:
This will display the actual values stored in the weights and biases. You should see that the weights are now small random numbers (from the Xavier uniform distribution), and the biases are all zeros. This quick check helps you confirm that your initialization code is working as intended.
In this lesson, you learned why weight and bias initialization are important for training neural networks. You saw how PyTorch stores these parameters in each layer and how to access them. You were introduced to Xavier initialization, a widely used method for setting the initial weights of fully connected layers, and learned how it uses fan_in and fan_out to determine the appropriate distribution bounds. You also learned why zero is a good default for biases. Finally, you practiced applying Xavier uniform initialization and zeroing biases in PyTorch, and learned how to verify that the initialization was successful.
In the upcoming practice exercises, you will get hands-on experience initializing weights and biases in your own models. This will help you build stronger neural networks and prepare you for even more advanced tuning techniques later in the course.
