• Docs >
  • Tutorial 3: Initialization and Optimization
Shortcuts

Tutorial 3: Initialization and Optimization

  • Author: Phillip Lippe

  • License: CC BY-SA

  • Generated: 2023-03-15T09:58:45.256359

In this tutorial, we will review techniques for optimization and initialization of neural networks. When increasing the depth of neural networks, there are various challenges we face. Most importantly, we need to have a stable gradient flow through the network, as otherwise, we might encounter vanishing or exploding gradients. This is why we will take a closer look at the following concepts: initialization and optimization. This notebook is part of a lecture series on Deep Learning at the University of Amsterdam. The full list of tutorials can be found at https://uvadlc-notebooks.rtfd.io.


Open in Open In Colab

Give us a ⭐ on Github | Check out the documentation | Join us on Slack

Setup

This notebook requires some packages besides pytorch-lightning.

[1]:
! pip install --quiet "torchmetrics>=0.7, <0.12" "matplotlib" "ipython[notebook]>=8.0.0, <8.12.0" "setuptools==67.4.0" "seaborn" "lightning>=2.0.0rc0" "torchvision" "torch>=1.8.1, <1.14.0" "pytorch-lightning>=1.4, <2.0.0"
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv

In the first half of the notebook, we will review different initialization techniques, and go step by step from the simplest initialization to methods that are nowadays used in very deep networks. In the second half, we focus on optimization comparing the optimizers SGD, SGD with Momentum, and Adam.

Let’s start with importing our standard libraries:

[2]:
import copy
import json
import math
import os
import urllib.request
from urllib.error import HTTPError

import lightning as L
import matplotlib.pyplot as plt

%matplotlib inline
import matplotlib_inline.backend_inline
import numpy as np
import seaborn as sns
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data as data
from matplotlib import cm
from torchvision import transforms
from torchvision.datasets import FashionMNIST
from tqdm.notebook import tqdm

matplotlib_inline.backend_inline.set_matplotlib_formats("svg", "pdf")  # For export
sns.set()

Instead of the set_seed function as in Tutorial 3, we can use Lightning’s build-in function L.seed_everything. We will reuse the path variables DATASET_PATH and CHECKPOINT_PATH as in Tutorial 3. Adjust the paths if necessary.

[3]:
# Path to the folder where the datasets are/should be downloaded (e.g. MNIST)
DATASET_PATH = os.environ.get("PATH_DATASETS", "data/")
# Path to the folder where the pretrained models are saved
CHECKPOINT_PATH = os.environ.get("PATH_CHECKPOINT", "saved_models/InitOptim/")

# Seed everything
L.seed_everything(42)

# Ensure that all operations are deterministic on GPU (if used) for reproducibility
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

# Fetching the device that will be used throughout this notebook
device = torch.device("cpu") if not torch.cuda.is_available() else torch.device("cuda:0")
print("Using device", device)
Global seed set to 42
Using device cuda:0

In the last part of the notebook, we will train models using three different optimizers. The pretrained models for those are downloaded below.

[4]:
# Github URL where saved models are stored for this tutorial
base_url = "https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/"
# Files to download
pretrained_files = [
    "FashionMNIST_SGD.config",
    "FashionMNIST_SGD_results.json",
    "FashionMNIST_SGD.tar",
    "FashionMNIST_SGDMom.config",
    "FashionMNIST_SGDMom_results.json",
    "FashionMNIST_SGDMom.tar",
    "FashionMNIST_Adam.config",
    "FashionMNIST_Adam_results.json",
    "FashionMNIST_Adam.tar",
]
# Create checkpoint path if it doesn't exist yet
os.makedirs(CHECKPOINT_PATH, exist_ok=True)

# For each file, check whether it already exists. If not, try downloading it.
for file_name in pretrained_files:
    file_path = os.path.join(CHECKPOINT_PATH, file_name)
    if not os.path.isfile(file_path):
        file_url = base_url + file_name
        print(f"Downloading {file_url}...")
        try:
            urllib.request.urlretrieve(file_url, file_path)
        except HTTPError as e:
            print(
                "Something went wrong. Please try to download the file from the GDrive folder, or contact the author with the full output including the following error:\n",
                e,
            )
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_SGD.config...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_SGD_results.json...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_SGD.tar...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_SGDMom.config...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_SGDMom_results.json...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_SGDMom.tar...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_Adam.config...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_Adam_results.json...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial4/FashionMNIST_Adam.tar...

Preparation

Throughout this notebook, we will use a deep fully connected network, similar to our previous tutorial. We will also again apply the network to FashionMNIST, so you can relate to the results of Tutorial 3. We start by loading the FashionMNIST dataset:

[5]:

# Transformations applied on each image => first make them a tensor, then normalize them with mean 0 and std 1 transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.2861,), (0.3530,))]) # Loading the training dataset. We need to split it into a training and validation part train_dataset = FashionMNIST(root=DATASET_PATH, train=True, transform=transform, download=True) train_set, val_set = torch.utils.data.random_split(train_dataset, [50000, 10000]) # Loading the test set test_set = FashionMNIST(root=DATASET_PATH, train=False, transform=transform, download=True)
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw/train-images-idx3-ubyte.gz
Extracting /__w/15/s/.datasets/FashionMNIST/raw/train-images-idx3-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw/train-labels-idx1-ubyte.gz
Extracting /__w/15/s/.datasets/FashionMNIST/raw/train-labels-idx1-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw/t10k-images-idx3-ubyte.gz
Extracting /__w/15/s/.datasets/FashionMNIST/raw/t10k-images-idx3-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw/t10k-labels-idx1-ubyte.gz
Extracting /__w/15/s/.datasets/FashionMNIST/raw/t10k-labels-idx1-ubyte.gz to /__w/15/s/.datasets/FashionMNIST/raw

We define a set of data loaders that we can use for various purposes later. Note that for actually training a model, we will use different data loaders with a lower batch size.

[6]:
train_loader = data.DataLoader(train_set, batch_size=1024, shuffle=True, drop_last=False)
val_loader = data.DataLoader(val_set, batch_size=1024, shuffle=False, drop_last=False)
test_loader = data.DataLoader(test_set, batch_size=1024, shuffle=False, drop_last=False)

In comparison to the previous tutorial, we have changed the parameters of the normalization transformation transforms.Normalize. The normalization is now designed to give us an expected mean of 0 and a standard deviation of 1 across pixels. This will be particularly relevant for the discussion about initialization we will look at below, and hence we change it here. It should be noted that in most classification tasks, both normalization techniques (between -1 and 1 or mean 0 and stddev 1) have shown to work well. We can calculate the normalization parameters by determining the mean and standard deviation on the original images:

[7]:
print("Mean", (train_dataset.data.float() / 255.0).mean().item())
print("Std", (train_dataset.data.float() / 255.0).std().item())
Mean 0.2860405743122101
Std 0.3530242443084717

We can verify the transformation by looking at the statistics of a single batch:

[8]:
imgs, _ = next(iter(train_loader))
print(f"Mean: {imgs.mean().item():5.3f}")
print(f"Standard deviation: {imgs.std().item():5.3f}")
print(f"Maximum: {imgs.max().item():5.3f}")
print(f"Minimum: {imgs.min().item():5.3f}")
Mean: 0.020
Standard deviation: 1.011
Maximum: 2.022
Minimum: -0.810

Note that the maximum and minimum are not 1 and -1 anymore, but shifted towards the positive values. This is because FashionMNIST contains a lot of black pixels, similar to MNIST.

Next, we create a linear neural network. We use the same setup as in the previous tutorial.

[9]:
class BaseNetwork(nn.Module):
    def __init__(self, act_fn, input_size=784, num_classes=10, hidden_sizes=[512, 256, 256, 128]):
        """
        Args:
            act_fn: Object of the activation function that should be used as non-linearity in the network.
            input_size: Size of the input images in pixels
            num_classes: Number of classes we want to predict
            hidden_sizes: A list of integers specifying the hidden layer sizes in the NN
        """
        super().__init__()

        # Create the network based on the specified hidden sizes
        layers = []
        layer_sizes = [input_size] + hidden_sizes
        for layer_index in range(1, len(layer_sizes)):
            layers += [nn.Linear(layer_sizes[layer_index - 1], layer_sizes[layer_index]), act_fn]
        layers += [nn.Linear(layer_sizes[-1], num_classes)]
        # A module list registers a list of modules as submodules (e.g. for parameters)
        self.layers = nn.ModuleList(layers)

        self.config = {
            "act_fn": act_fn.__class__.__name__,
            "input_size": input_size,
            "num_classes": num_classes,
            "hidden_sizes": hidden_sizes,
        }

    def forward(self, x):
        x = x.view(x.size(0), -1)
        for layer in self.layers:
            x = layer(x)
        return x

For the activation functions, we make use of PyTorch’s torch.nn library instead of implementing ourselves. However, we also define an Identity activation function. Although this activation function would significantly limit the network’s modeling capabilities, we will use it in the first steps of our discussion about initialization (for simplicity).

[10]:
class Identity(nn.Module):
    def forward(self, x):
        return x


act_fn_by_name = {"tanh": nn.Tanh, "relu": nn.ReLU, "identity": Identity}

Finally, we define a few plotting functions that we will use for our discussions. These functions help us to (1) visualize the weight/parameter distribution inside a network, (2) visualize the gradients that the parameters at different layers receive, and (3) the activations, i.e. the output of the linear layers. The detailed code is not important, but feel free to take a closer look if interested.

[11]:
##############################################################


def plot_dists(val_dict, color="C0", xlabel=None, stat="count", use_kde=True):
    columns = len(val_dict)
    fig, ax = plt.subplots(1, columns, figsize=(columns * 3, 2.5))
    fig_index = 0
    for key in sorted(val_dict.keys()):
        key_ax = ax[fig_index % columns]
        sns.histplot(
            val_dict[key],
            ax=key_ax,
            color=color,
            bins=50,
            stat=stat,
            kde=use_kde and ((val_dict[key].max() - val_dict[key].min()) > 1e-8),
        )  # Only plot kde if there is variance
        hidden_dim_str = (
            r"(%i $\to$ %i)" % (val_dict[key].shape[1], val_dict[key].shape[0]) if len(val_dict[key].shape) > 1 else ""
        )
        key_ax.set_title(f"{key} {hidden_dim_str}")
        if xlabel is not None:
            key_ax.set_xlabel(xlabel)
        fig_index += 1
    fig.subplots_adjust(wspace=0.4)
    return fig


##############################################################


def visualize_weight_distribution(model, color="C0"):
    weights = {}
    for name, param in model.named_parameters():
        if name.endswith(".bias"):
            continue
        key_name = f"Layer {name.split('.')[1]}"
        weights[key_name] = param.detach().view(-1).cpu().numpy()

    # Plotting
    fig = plot_dists(weights, color=color, xlabel="Weight vals")
    fig.suptitle("Weight distribution", fontsize=14, y=1.05)
    plt.show()
    plt.close()


##############################################################


def visualize_gradients(model, color="C0", print_variance=False):
    """
    Args:
        net: Object of class BaseNetwork
        color: Color in which we want to visualize the histogram (for easier separation of activation functions)
    """
    model.eval()
    small_loader = data.DataLoader(train_set, batch_size=1024, shuffle=False)
    imgs, labels = next(iter(small_loader))
    imgs, labels = imgs.to(device), labels.to(device)

    # Pass one batch through the network, and calculate the gradients for the weights
    model.zero_grad()
    preds = model(imgs)
    loss = F.cross_entropy(preds, labels)  # Same as nn.CrossEntropyLoss, but as a function instead of module
    loss.backward()
    # We limit our visualization to the weight parameters and exclude the bias to reduce the number of plots
    grads = {
        name: params.grad.view(-1).cpu().clone().numpy()
        for name, params in model.named_parameters()
        if "weight" in name
    }
    model.zero_grad()

    # Plotting
    fig = plot_dists(grads, color=color, xlabel="Grad magnitude")
    fig.suptitle("Gradient distribution", fontsize=14, y=1.05)
    plt.show()
    plt.close()

    if print_variance:
        for key in sorted(grads.keys()):
            print(f"{key} - Variance: {np.var(grads[key])}")


##############################################################


def visualize_activations(model, color="C0", print_variance=False):
    model.eval()
    small_loader = data.DataLoader(train_set, batch_size=1024, shuffle=False)
    imgs, labels = next(iter(small_loader))
    imgs, labels = imgs.to(device), labels.to(device)

    # Pass one batch through the network, and calculate the gradients for the weights
    feats = imgs.view(imgs.shape[0], -1)
    activations = {}
    with torch.no_grad():
        for layer_index, layer in enumerate(model.layers):
            feats = layer(feats)
            if isinstance(layer, nn.Linear):
                activations[f"Layer {layer_index}"] = feats.view(-1).detach().cpu().numpy()

    # Plotting
    fig = plot_dists(activations, color=color, stat="density", xlabel="Activation vals")
    fig.suptitle("Activation distribution", fontsize=14, y=1.05)
    plt.show()
    plt.close()

    if print_variance:
        for key in sorted(activations.keys()):
            print(f"{key} - Variance: {np.var(activations[key])}")


##############################################################

Initialization

Before starting our discussion about initialization, it should be noted that there exist many very good blog posts about the topic of neural network initialization (for example deeplearning.ai, or a more math-focused blog post). In case something remains unclear after this tutorial, we recommend skimming through these blog posts as well.

When initializing a neural network, there are a few properties we would like to have. First, the variance of the input should be propagated through the model to the last layer, so that we have a similar standard deviation for the output neurons. If the variance would vanish the deeper we go in our model, it becomes much harder to optimize the model as the input to the next layer is basically a single constant value. Similarly, if the variance increases, it is likely to explode (i.e. head to infinity) the deeper we design our model. The second property we look out for in initialization techniques is a gradient distribution with equal variance across layers. If the first layer receives much smaller gradients than the last layer, we will have difficulties in choosing an appropriate learning rate.

As a starting point for finding a good method, we will analyze different initialization based on our linear neural network with no activation function (i.e. an identity). We do this because initializations depend on the specific activation function used in the network, and we can adjust the initialization schemes later on for our specific choice.

[12]:
model = BaseNetwork(act_fn=Identity()).to(device)

Constant initialization

The first initialization we can consider is to initialize all weights with the same constant value. Intuitively, setting all weights to zero is not a good idea as the propagated gradient will be zero. However, what happens if we set all weights to a value slightly larger or smaller than 0? To find out, we can implement a function for setting all parameters below and visualize the gradients.

[13]:
def const_init(model, fill=0.0):
    for name, param in model.named_parameters():
        param.data.fill_(fill)


const_init(model, fill=0.005)
visualize_gradients(model)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_27_0.svg
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_27_1.svg
Layer 0 - Variance: 2.0582759380340576
Layer 2 - Variance: 13.489116668701172
Layer 4 - Variance: 22.100568771362305
Layer 6 - Variance: 36.20956802368164
Layer 8 - Variance: 14.831439018249512

As we can see, only the first and the last layer have diverse gradient distributions while the other three layers have the same gradient for all weights (note that this value is unequal 0, but often very close to it). Having the same gradient for parameters that have been initialized with the same values means that we will always have the same value for those parameters. This would make our layer useless and reduce our effective number of parameters to 1. Thus, we cannot use a constant initialization to train our networks.

Constant variance

From the experiment above, we have seen that a constant value is not working. So instead, how about we initialize the parameters by randomly sampling from a distribution like a Gaussian? The most intuitive way would be to choose one variance that is used for all layers in the network. Let’s implement it below, and visualize the activation distribution across layers.

[14]:
def var_init(model, std=0.01):
    for name, param in model.named_parameters():
        param.data.normal_(mean=0.0, std=std)


var_init(model, std=0.01)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_30_0.svg
Layer 0 - Variance: 0.07692592591047287
Layer 2 - Variance: 0.00398120516911149
Layer 4 - Variance: 0.0002155901602236554
Layer 6 - Variance: 0.00011756643652915955
Layer 8 - Variance: 8.103555592242628e-05

The variance of the activation becomes smaller and smaller across layers, and almost vanishes in the last layer. Alternatively, we could use a higher standard deviation:

[15]:
var_init(model, std=0.1)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_32_0.svg
Layer 0 - Variance: 8.010124206542969
Layer 2 - Variance: 42.5277099609375
Layer 4 - Variance: 107.65530395507812
Layer 6 - Variance: 294.1153564453125
Layer 8 - Variance: 360.99920654296875

With a higher standard deviation, the activations are likely to explode. You can play around with the specific standard deviation values, but it will be hard to find one that gives us a good activation distribution across layers and is very specific to our model. If we would change the hidden sizes or number of layers, you would have to search all over again, which is neither efficient nor recommended.

How to find appropriate initialization values

From our experiments above, we have seen that we need to sample the weights from a distribution, but are not sure which one exactly. As a next step, we will try to find the optimal initialization from the perspective of the activation distribution. For this, we state two requirements:

  1. The mean of the activations should be zero

  2. The variance of the activations should stay the same across every layer

Suppose we want to design an initialization for the following layer: y=Wx+b with y\in\mathbb{R}^{d_y}, x\in\mathbb{R}^{d_x}. Our goal is that the variance of each element of y is the same as the input, i.e. \text{Var}(y_i)=\text{Var}(x_i)=\sigma_x^{2}, and that the mean is zero. We assume x to also have a mean of zero, because, in deep neural networks, y would be the input of another layer. This requires the bias and weight to have an expectation of 0. Actually, as b is a single element per output neuron and is constant across different inputs, we set it to 0 overall.

Next, we need to calculate the variance with which we need to initialize the weight parameters. Along the calculation, we will need to following variance rule: given two independent variables, the variance of their product is \text{Var}(X\cdot Y) = \mathbb{E}(Y)^2\text{Var}(X) + \mathbb{E}(X)^2\text{Var}(Y) + \text{Var}(X)\text{Var}(Y) = \mathbb{E}(Y^2)\mathbb{E}(X^2)-\mathbb{E}(Y)^2\mathbb{E}(X)^2 (X and Y are not refering to x and y, but any random variable).

The needed variance of the weights, \text{Var}(w_{ij}), is calculated as follows:

\begin{split}
    y_i & = \sum_{j} w_{ij}x_{j}\hspace{10mm}\text{Calculation of a single output neuron without bias}\\
    \text{Var}(y_i) = \sigma_x^{2} & = \text{Var}\left(\sum_{j} w_{ij}x_{j}\right)\\
    & = \sum_{j} \text{Var}(w_{ij}x_{j}) \hspace{10mm}\text{Inputs and weights are independent of each other}\\
    & = \sum_{j} \text{Var}(w_{ij})\cdot\text{Var}(x_{j}) \hspace{10mm}\text{Variance rule (see above) with expectations being zero}\\
    & = d_x \cdot \text{Var}(w_{ij})\cdot\text{Var}(x_{j}) \hspace{10mm}\text{Variance equal for all $d_x$ elements}\\
    & = \sigma_x^{2} \cdot d_x \cdot \text{Var}(w_{ij})\\
    \Rightarrow \text{Var}(w_{ij}) = \sigma_{W}^2 & = \frac{1}{d_x}\\
\end{split}

Thus, we should initialize the weight distribution with a variance of the inverse of the input dimension d_x. Let’s implement it below and check whether this holds:

[16]:
def equal_var_init(model):
    for name, param in model.named_parameters():
        if name.endswith(".bias"):
            param.data.fill_(0)
        else:
            param.data.normal_(std=1.0 / math.sqrt(param.shape[1]))


equal_var_init(model)
visualize_weight_distribution(model)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_35_0.svg
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_35_1.svg
Layer 0 - Variance: 1.034169316291809
Layer 2 - Variance: 1.0524311065673828
Layer 4 - Variance: 1.0837327241897583
Layer 6 - Variance: 1.0072245597839355
Layer 8 - Variance: 0.9160612225532532

As we expected, the variance stays indeed constant across layers. Note that our initialization does not restrict us to a normal distribution, but allows any other distribution with a mean of 0 and variance of 1/d_x. You often see that a uniform distribution is used for initialization. A small benefit of using a uniform instead of a normal distribution is that we can exclude the chance of initializing very large or small weights.

Besides the variance of the activations, another variance we would like to stabilize is the one of the gradients. This ensures a stable optimization for deep networks. It turns out that we can do the same calculation as above starting from \Delta x=W\Delta y, and come to the conclusion that we should initialize our layers with 1/d_y where d_y is the number of output neurons. You can do the calculation as a practice, or check a thorough explanation in this blog post. As a compromise between both constraints, Glorot and Bengio (2010) proposed to use the harmonic mean of both values. This leads us to the well-known Xavier initialization:

W\sim \mathcal{N}\left(0,\frac{2}{d_x+d_y}\right)

If we use a uniform distribution, we would initialize the weights with:

W\sim U\left[-\frac{\sqrt{6}}{\sqrt{d_x+d_y}}, \frac{\sqrt{6}}{\sqrt{d_x+d_y}}\right]

Let’s shortly implement it and validate its effectiveness:

[17]:
def xavier_init(model):
    for name, param in model.named_parameters():
        if name.endswith(".bias"):
            param.data.fill_(0)
        else:
            bound = math.sqrt(6) / math.sqrt(param.shape[0] + param.shape[1])
            param.data.uniform_(-bound, bound)


xavier_init(model)
visualize_gradients(model, print_variance=True)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_37_0.svg
layers.0.weight - Variance: 0.0005894455825909972
layers.2.weight - Variance: 0.0009689269936643541
layers.4.weight - Variance: 0.0011539047118276358
layers.6.weight - Variance: 0.001938404282554984
layers.8.weight - Variance: 0.01793084293603897
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_37_2.svg
Layer 0 - Variance: 1.2016416788101196
Layer 2 - Variance: 1.5907800197601318
Layer 4 - Variance: 1.728979468345642
Layer 6 - Variance: 2.3041787147521973
Layer 8 - Variance: 3.7488739490509033

We see that the Xavier initialization balances the variance of gradients and activations. Note that the significantly higher variance for the output layer is due to the large difference of input and output dimension (128 vs 10). However, we currently assumed the activation function to be linear. So what happens if we add a non-linearity? In a tanh-based network, a common assumption is that for small values during the initial steps in training, the \tanh works as a linear function such that we don’t have to adjust our calculation. We can check if that is the case for us as well:

[18]:
model = BaseNetwork(act_fn=nn.Tanh()).to(device)
xavier_init(model)
visualize_gradients(model, print_variance=True)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_39_0.svg
layers.0.weight - Variance: 2.5025436116266064e-05
layers.2.weight - Variance: 3.932134495698847e-05
layers.4.weight - Variance: 4.471026113606058e-05
layers.6.weight - Variance: 5.619897638098337e-05
layers.8.weight - Variance: 0.00048347172560170293
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_39_2.svg
Layer 0 - Variance: 1.2273123264312744
Layer 2 - Variance: 0.5360878705978394
Layer 4 - Variance: 0.2593609690666199
Layer 6 - Variance: 0.2384992241859436
Layer 8 - Variance: 0.29864996671676636

Although the variance decreases over depth, it is apparent that the activation distribution becomes more focused on the low values. Therefore, our variance will stabilize around 0.25 if we would go even deeper. Hence, we can conclude that the Xavier initialization works well for Tanh networks. But what about ReLU networks? Here, we cannot take the previous assumption of the non-linearity becoming linear for small values. The ReLU activation function sets (in expectation) half of the inputs to 0 so that also the expectation of the input is not zero. However, as long as the expectation of W is zero and b=0, the expectation of the output is zero. The part where the calculation of the ReLU initialization differs from the identity is when determining \text{Var}(w_{ij}x_{j}):

\text{Var}(w_{ij}x_{j})=\underbrace{\mathbb{E}[w_{ij}^2]}_{=\text{Var}(w_{ij})}\mathbb{E}[x_{j}^2]-\underbrace{\mathbb{E}[w_{ij}]^2}_{=0}\mathbb{E}[x_{j}]^2=\text{Var}(w_{ij})\mathbb{E}[x_{j}^2]

If we assume now that x is the output of a ReLU activation (from a previous layer, x=max(0,\tilde{y})), we can calculate the expectation as follows:

\begin{split}
    \mathbb{E}[x^2] & =\mathbb{E}[\max(0,\tilde{y})^2]\\
                    & =\frac{1}{2}\mathbb{E}[{\tilde{y}}^2]\hspace{2cm}\tilde{y}\text{ is zero-centered and symmetric}\\
                    & =\frac{1}{2}\text{Var}(\tilde{y})
\end{split}

Thus, we see that we have an additional factor of 1/2 in the equation, so that our desired weight variance becomes 2/d_x. This gives us the Kaiming initialization (see He, K. et al. (2015)). Note that the Kaiming initialization does not use the harmonic mean between input and output size. In their paper (Section 2.2, Backward Propagation, last paragraph), they argue that using d_x or d_y both lead to stable gradients throughout the network, and only depend on the overall input and output size of the network. Hence, we can use here only the input d_x:

[19]:
def kaiming_init(model):
    for name, param in model.named_parameters():
        if name.endswith(".bias"):
            param.data.fill_(0)
        elif name.startswith("layers.0"):  # The first layer does not have ReLU applied on its input
            param.data.normal_(0, 1 / math.sqrt(param.shape[1]))
        else:
            param.data.normal_(0, math.sqrt(2) / math.sqrt(param.shape[1]))


model = BaseNetwork(act_fn=nn.ReLU()).to(device)
kaiming_init(model)
visualize_gradients(model, print_variance=True)
visualize_activations(model, print_variance=True)
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_41_0.svg
layers.0.weight - Variance: 3.730092794285156e-05
layers.2.weight - Variance: 4.7479999921051785e-05
layers.4.weight - Variance: 6.661842780886218e-05
layers.6.weight - Variance: 0.00014274715795181692
layers.8.weight - Variance: 0.002348906360566616
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_41_2.svg
Layer 0 - Variance: 1.011019229888916
Layer 2 - Variance: 1.0046828985214233
Layer 4 - Variance: 1.060686469078064
Layer 6 - Variance: 1.1315085887908936
Layer 8 - Variance: 0.5648466348648071

The variance stays stable across layers. We can conclude that the Kaiming initialization indeed works well for ReLU-based networks. Note that for Leaky-ReLU etc., we have to slightly adjust the factor of 2 in the variance as half of the values are not set to zero anymore. PyTorch provides a function to calculate this factor for many activation function, see torch.nn.init.calculate_gain (link).

Optimization

Besides initialization, selecting a suitable optimization algorithm can be an important choice for deep neural networks. Before taking a closer look at them, we should define code for training the models. Most of the following code is copied from the previous tutorial, and only slightly altered to fit our needs.

[20]:
def _get_config_file(model_path, model_name):
    return os.path.join(model_path, model_name + ".config")


def _get_model_file(model_path, model_name):
    return os.path.join(model_path, model_name + ".tar")


def _get_result_file(model_path, model_name):
    return os.path.join(model_path, model_name + "_results.json")


def load_model(model_path, model_name, net=None):
    config_file = _get_config_file(model_path, model_name)
    model_file = _get_model_file(model_path, model_name)
    assert os.path.isfile(
        config_file
    ), f'Could not find the config file "{config_file}". Are you sure this is the correct path and you have your model config stored here?'
    assert os.path.isfile(
        model_file
    ), f'Could not find the model file "{model_file}". Are you sure this is the correct path and you have your model stored here?'
    with open(config_file) as f:
        config_dict = json.load(f)
    if net is None:
        act_fn_name = config_dict["act_fn"].pop("name").lower()
        assert (
            act_fn_name in act_fn_by_name
        ), f'Unknown activation function "{act_fn_name}". Please add it to the "act_fn_by_name" dict.'
        act_fn = act_fn_by_name[act_fn_name]()
        net = BaseNetwork(act_fn=act_fn, **config_dict)
    net.load_state_dict(torch.load(model_file))
    return net


def save_model(model, model_path, model_name):
    config_dict = model.config
    os.makedirs(model_path, exist_ok=True)
    config_file = _get_config_file(model_path, model_name)
    model_file = _get_model_file(model_path, model_name)
    with open(config_file, "w") as f:
        json.dump(config_dict, f)
    torch.save(model.state_dict(), model_file)


def train_model(net, model_name, optim_func, max_epochs=50, batch_size=256, overwrite=False):
    """Train a model on the training set of FashionMNIST.

    Args:
        net: Object of BaseNetwork
        model_name: (str) Name of the model, used for creating the checkpoint names
        max_epochs: Number of epochs we want to (maximally) train for
        patience: If the performance on the validation set has not improved for #patience epochs, we stop training early
        batch_size: Size of batches used in training
        overwrite: Determines how to handle the case when there already exists a checkpoint. If True, it will be overwritten. Otherwise, we skip training.
    """
    file_exists = os.path.isfile(_get_model_file(CHECKPOINT_PATH, model_name))
    if file_exists and not overwrite:
        print(f'Model file of "{model_name}" already exists. Skipping training...')
        with open(_get_result_file(CHECKPOINT_PATH, model_name)) as f:
            results = json.load(f)
    else:
        if file_exists:
            print("Model file exists, but will be overwritten...")

        # Defining optimizer, loss and data loader
        optimizer = optim_func(net.parameters())
        loss_module = nn.CrossEntropyLoss()
        train_loader_local = data.DataLoader(
            train_set, batch_size=batch_size, shuffle=True, drop_last=True, pin_memory=True
        )

        results = None
        val_scores = []
        train_losses, train_scores = [], []
        best_val_epoch = -1
        for epoch in range(max_epochs):
            train_acc, val_acc, epoch_losses = epoch_iteration(
                net, loss_module, optimizer, train_loader_local, val_loader, epoch
            )
            train_scores.append(train_acc)
            val_scores.append(val_acc)
            train_losses += epoch_losses

            if len(val_scores) == 1 or val_acc > val_scores[best_val_epoch]:
                print("\t   (New best performance, saving model...)")
                save_model(net, CHECKPOINT_PATH, model_name)
                best_val_epoch = epoch

    if results is None:
        load_model(CHECKPOINT_PATH, model_name, net=net)
        test_acc = test_model(net, test_loader)
        results = {
            "test_acc": test_acc,
            "val_scores": val_scores,
            "train_losses": train_losses,
            "train_scores": train_scores,
        }
        with open(_get_result_file(CHECKPOINT_PATH, model_name), "w") as f:
            json.dump(results, f)

    # Plot a curve of the validation accuracy
    sns.set()
    plt.plot([i for i in range(1, len(results["train_scores"]) + 1)], results["train_scores"], label="Train")
    plt.plot([i for i in range(1, len(results["val_scores"]) + 1)], results["val_scores"], label="Val")
    plt.xlabel("Epochs")
    plt.ylabel("Validation accuracy")
    plt.ylim(min(results["val_scores"]), max(results["train_scores"]) * 1.01)
    plt.title(f"Validation performance of {model_name}")
    plt.legend()
    plt.show()
    plt.close()

    print((f" Test accuracy: {results['test_acc']*100.0:4.2f}% ").center(50, "=") + "\n")
    return results


def epoch_iteration(net, loss_module, optimizer, train_loader_local, val_loader, epoch):
    ############
    # Training #
    ############
    net.train()
    true_preds, count = 0.0, 0
    epoch_losses = []
    t = tqdm(train_loader_local, leave=False)
    for imgs, labels in t:
        imgs, labels = imgs.to(device), labels.to(device)
        optimizer.zero_grad()
        preds = net(imgs)
        loss = loss_module(preds, labels)
        loss.backward()
        optimizer.step()
        # Record statistics during training
        true_preds += (preds.argmax(dim=-1) == labels).sum().item()
        count += labels.shape[0]
        t.set_description(f"Epoch {epoch+1}: loss={loss.item():4.2f}")
        epoch_losses.append(loss.item())
    train_acc = true_preds / count

    ##############
    # Validation #
    ##############
    val_acc = test_model(net, val_loader)
    print(
        f"[Epoch {epoch+1:2i}] Training accuracy: {train_acc*100.0:05.2f}%, Validation accuracy: {val_acc*100.0:05.2f}%"
    )
    return train_acc, val_acc, epoch_losses


def test_model(net, data_loader):
    """Test a model on a specified dataset.

    Args:
        net: Trained model of type BaseNetwork
        data_loader: DataLoader object of the dataset to test on (validation or test)
    """
    net.eval()
    true_preds, count = 0.0, 0
    for imgs, labels in data_loader:
        imgs, labels = imgs.to(device), labels.to(device)
        with torch.no_grad():
            preds = net(imgs).argmax(dim=-1)
            true_preds += (preds == labels).sum().item()
            count += labels.shape[0]
    test_acc = true_preds / count
    return test_acc

First, we need to understand what an optimizer actually does. The optimizer is responsible to update the network’s parameters given the gradients. Hence, we effectively implement a function w^{t} = f(w^{t-1}, g^{t}, ...) with w being the parameters, and g^{t} = \nabla_{w^{(t-1)}} \mathcal{L}^{(t)} the gradients at time step t. A common, additional parameter to this function is the learning rate, here denoted by \eta. Usually, the learning rate can be seen as the “step size” of the update. A higher learning rate means that we change the weights more in the direction of the gradients, a smaller means we take shorter steps.

As most optimizers only differ in the implementation of f, we can define a template for an optimizer in PyTorch below. We take as input the parameters of a model and a learning rate. The function zero_grad sets the gradients of all parameters to zero, which we have to do before calling loss.backward(). Finally, the step() function tells the optimizer to update all weights based on their gradients. The template is setup below:

[21]:
class OptimizerTemplate:
    def __init__(self, params, lr):
        self.params = list(params)
        self.lr = lr

    def zero_grad(self):
        # Set gradients of all parameters to zero
        for p in self.params:
            if p.grad is not None:
                p.grad.detach_()  # For second-order optimizers important
                p.grad.zero_()

    @torch.no_grad()
    def step(self):
        # Apply update step to all parameters
        for p in self.params:
            if p.grad is None:  # We skip parameters without any gradients
                continue
            self.update_param(p)

    def update_param(self, p):
        # To be implemented in optimizer-specific classes
        raise NotImplementedError

The first optimizer we are going to implement is the standard Stochastic Gradient Descent (SGD). SGD updates the parameters using the following equation:

\begin{split}
    w^{(t)} & = w^{(t-1)} - \eta \cdot g^{(t)}
\end{split}

As simple as the equation is also our implementation of SGD:

[22]:
class SGD(OptimizerTemplate):
    def __init__(self, params, lr):
        super().__init__(params, lr)

    def update_param(self, p):
        p_update = -self.lr * p.grad
        p.add_(p_update)  # In-place update => saves memory and does not create computation graph

In the lecture, we also have discussed the concept of momentum which replaces the gradient in the update by an exponential average of all past gradients including the current one:

\begin{split}
    m^{(t)} & = \beta_1 m^{(t-1)} + (1 - \beta_1)\cdot g^{(t)}\\
    w^{(t)} & = w^{(t-1)} - \eta \cdot m^{(t)}\\
\end{split}

Let’s also implement it below:

[23]:
class SGDMomentum(OptimizerTemplate):
    def __init__(self, params, lr, momentum=0.0):
        super().__init__(params, lr)
        self.momentum = momentum  # Corresponds to beta_1 in the equation above
        self.param_momentum = {p: torch.zeros_like(p.data) for p in self.params}  # Dict to store m_t

    def update_param(self, p):
        self.param_momentum[p] = (1 - self.momentum) * p.grad + self.momentum * self.param_momentum[p]
        p_update = -self.lr * self.param_momentum[p]
        p.add_(p_update)

Finally, we arrive at Adam. Adam combines the idea of momentum with an adaptive learning rate, which is based on an exponential average of the squared gradients, i.e. the gradients norm. Furthermore, we add a bias correction for the momentum and adaptive learning rate for the first iterations:

\begin{split}
    m^{(t)} & = \beta_1 m^{(t-1)} + (1 - \beta_1)\cdot g^{(t)}\\
    v^{(t)} & = \beta_2 v^{(t-1)} + (1 - \beta_2)\cdot \left(g^{(t)}\right)^2\\
    \hat{m}^{(t)} & = \frac{m^{(t)}}{1-\beta^{t}_1}, \hat{v}^{(t)} = \frac{v^{(t)}}{1-\beta^{t}_2}\\
    w^{(t)} & = w^{(t-1)} - \frac{\eta}{\sqrt{v^{(t)}} + \epsilon}\circ \hat{m}^{(t)}\\
\end{split}

Epsilon is a small constant used to improve numerical stability for very small gradient norms. Remember that the adaptive learning rate does not replace the learning rate hyperparameter \eta, but rather acts as an extra factor and ensures that the gradients of various parameters have a similar norm.

[24]:
class Adam(OptimizerTemplate):
    def __init__(self, params, lr, beta1=0.9, beta2=0.999, eps=1e-8):
        super().__init__(params, lr)
        self.beta1 = beta1
        self.beta2 = beta2
        self.eps = eps
        self.param_step = {p: 0 for p in self.params}  # Remembers "t" for each parameter for bias correction
        self.param_momentum = {p: torch.zeros_like(p.data) for p in self.params}
        self.param_2nd_momentum = {p: torch.zeros_like(p.data) for p in self.params}

    def update_param(self, p):
        self.param_step[p] += 1

        self.param_momentum[p] = (1 - self.beta1) * p.grad + self.beta1 * self.param_momentum[p]
        self.param_2nd_momentum[p] = (1 - self.beta2) * (p.grad) ** 2 + self.beta2 * self.param_2nd_momentum[p]

        bias_correction_1 = 1 - self.beta1 ** self.param_step[p]
        bias_correction_2 = 1 - self.beta2 ** self.param_step[p]

        p_2nd_mom = self.param_2nd_momentum[p] / bias_correction_2
        p_mom = self.param_momentum[p] / bias_correction_1
        p_lr = self.lr / (torch.sqrt(p_2nd_mom) + self.eps)
        p_update = -p_lr * p_mom

        p.add_(p_update)

Comparing optimizers on model training

After we have implemented three optimizers (SGD, SGD with momentum, and Adam), we can start to analyze and compare them. First, we test them on how well they can optimize a neural network on the FashionMNIST dataset. We use again our linear network, this time with a ReLU activation and the kaiming initialization, which we have found before to work well for ReLU-based networks. Note that the model is over-parameterized for this task, and we can achieve similar performance with a much smaller network (for example 100,100,100). However, our main interest is in how well the optimizer can train deep neural networks, hence the over-parameterization.

[25]:
base_model = BaseNetwork(act_fn=nn.ReLU(), hidden_sizes=[512, 256, 256, 128])
kaiming_init(base_model)

For a fair comparison, we train the exact same model with the same seed with the three optimizers below. Feel free to change the hyperparameters if you want (however, you have to train your own model then).

[26]:
SGD_model = copy.deepcopy(base_model).to(device)
SGD_results = train_model(
    SGD_model, "FashionMNIST_SGD", lambda params: SGD(params, lr=1e-1), max_epochs=40, batch_size=256
)
Model file of "FashionMNIST_SGD" already exists. Skipping training...
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_56_1.svg
============= Test accuracy: 89.09% ==============

[27]:
SGDMom_model = copy.deepcopy(base_model).to(device)
SGDMom_results = train_model(
    SGDMom_model,
    "FashionMNIST_SGDMom",
    lambda params: SGDMomentum(params, lr=1e-1, momentum=0.9),
    max_epochs=40,
    batch_size=256,
)
Model file of "FashionMNIST_SGDMom" already exists. Skipping training...
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_57_1.svg
============= Test accuracy: 88.83% ==============

[28]:
Adam_model = copy.deepcopy(base_model).to(device)
Adam_results = train_model(
    Adam_model, "FashionMNIST_Adam", lambda params: Adam(params, lr=1e-3), max_epochs=40, batch_size=256
)
Model file of "FashionMNIST_Adam" already exists. Skipping training...
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_58_1.svg
============= Test accuracy: 89.46% ==============

The result is that all optimizers perform similarly well with the given model. The differences are too small to find any significant conclusion. However, keep in mind that this can also be attributed to the initialization we chose. When changing the initialization to worse (e.g. constant initialization), Adam usually shows to be more robust because of its adaptive learning rate. To show the specific benefits of the optimizers, we will continue to look at some possible loss surfaces in which momentum and adaptive learning rate are crucial.

Pathological curvatures

A pathological curvature is a type of surface that is similar to ravines and is particularly tricky for plain SGD optimization. In words, pathological curvatures typically have a steep gradient in one direction with an optimum at the center, while in a second direction we have a slower gradient towards a (global) optimum. Let’s first create an example surface of this and visualize it:

[29]:
def pathological_curve_loss(w1, w2):
    # Example of a pathological curvature. There are many more possible, feel free to experiment here!
    x1_loss = torch.tanh(w1) ** 2 + 0.01 * torch.abs(w1)
    x2_loss = torch.sigmoid(w2)
    return x1_loss + x2_loss
[30]:
def plot_curve(
    curve_fn, x_range=(-5, 5), y_range=(-5, 5), plot_3d=False, cmap=cm.viridis, title="Pathological curvature"
):
    fig = plt.figure()
    ax = fig.gca()
    if plot_3d:
        ax = fig.add_subplot(projection="3d")

    x = torch.arange(x_range[0], x_range[1], (x_range[1] - x_range[0]) / 100.0)
    y = torch.arange(y_range[0], y_range[1], (y_range[1] - y_range[0]) / 100.0)
    x, y = torch.meshgrid([x, y])
    z = curve_fn(x, y)
    x, y, z = x.numpy(), y.numpy(), z.numpy()

    if plot_3d:
        ax.plot_surface(x, y, z, cmap=cmap, linewidth=1, color="#000", antialiased=False)
        ax.set_zlabel("loss")
    else:
        ax.imshow(z.T[::-1], cmap=cmap, extent=(x_range[0], x_range[1], y_range[0], y_range[1]))
    plt.title(title)
    ax.set_xlabel(r"$w_1$")
    ax.set_ylabel(r"$w_2$")
    plt.tight_layout()
    return ax


sns.reset_orig()
_ = plot_curve(pathological_curve_loss, plot_3d=True)
plt.show()
/usr/local/lib/python3.9/dist-packages/torch/functional.py:504: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at ../aten/src/ATen/native/TensorShape.cpp:3190.)
  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_62_1.svg

In terms of optimization, you can image that w_1 and w_2 are weight parameters, and the curvature represents the loss surface over the space of w_1 and w_2. Note that in typical networks, we have many, many more parameters than two, and such curvatures can occur in multi-dimensional spaces as well.

Ideally, our optimization algorithm would find the center of the ravine and focuses on optimizing the parameters towards the direction of w_2. However, if we encounter a point along the ridges, the gradient is much greater in w_1 than w_2, and we might end up jumping from one side to the other. Due to the large gradients, we would have to reduce our learning rate slowing down learning significantly.

To test our algorithms, we can implement a simple function to train two parameters on such a surface:

[31]:
def train_curve(optimizer_func, curve_func=pathological_curve_loss, num_updates=100, init=[5, 5]):
    """
    Args:
        optimizer_func: Constructor of the optimizer to use. Should only take a parameter list
        curve_func: Loss function (e.g. pathological curvature)
        num_updates: Number of updates/steps to take when optimizing
        init: Initial values of parameters. Must be a list/tuple with two elements representing w_1 and w_2
    Returns:
        Numpy array of shape [num_updates, 3] with [t,:2] being the parameter values at step t, and [t,2] the loss at t.
    """
    weights = nn.Parameter(torch.FloatTensor(init), requires_grad=True)
    optim = optimizer_func([weights])

    list_points = []
    for _ in range(num_updates):
        loss = curve_func(weights[0], weights[1])
        list_points.append(torch.cat([weights.data.detach(), loss.unsqueeze(dim=0).detach()], dim=0))
        optim.zero_grad()
        loss.backward()
        optim.step()
    points = torch.stack(list_points, dim=0).numpy()
    return points

Next, let’s apply the different optimizers on our curvature. Note that we set a much higher learning rate for the optimization algorithms as you would in a standard neural network. This is because we only have 2 parameters instead of tens of thousands or even millions.

[32]:
SGD_points = train_curve(lambda params: SGD(params, lr=10))
SGDMom_points = train_curve(lambda params: SGDMomentum(params, lr=10, momentum=0.9))
Adam_points = train_curve(lambda params: Adam(params, lr=1))

To understand best how the different algorithms worked, we visualize the update step as a line plot through the loss surface. We will stick with a 2D representation for readability.

[33]:
all_points = np.concatenate([SGD_points, SGDMom_points, Adam_points], axis=0)
ax = plot_curve(
    pathological_curve_loss,
    x_range=(-np.absolute(all_points[:, 0]).max(), np.absolute(all_points[:, 0]).max()),
    y_range=(all_points[:, 1].min(), all_points[:, 1].max()),
    plot_3d=False,
)
ax.plot(SGD_points[:, 0], SGD_points[:, 1], color="red", marker="o", zorder=1, label="SGD")
ax.plot(SGDMom_points[:, 0], SGDMom_points[:, 1], color="blue", marker="o", zorder=2, label="SGDMom")
ax.plot(Adam_points[:, 0], Adam_points[:, 1], color="grey", marker="o", zorder=3, label="Adam")
plt.legend()
plt.show()
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_68_0.svg

We can clearly see that SGD is not able to find the center of the optimization curve and has a problem converging due to the steep gradients in w_1. In contrast, Adam and SGD with momentum nicely converge as the changing direction of w_1 is canceling itself out. On such surfaces, it is crucial to use momentum.

Steep optima

A second type of challenging loss surfaces are steep optima. In those, we have a larger part of the surface having very small gradients while around the optimum, we have very large gradients. For instance, take the following loss surfaces:

[34]:
def bivar_gaussian(w1, w2, x_mean=0.0, y_mean=0.0, x_sig=1.0, y_sig=1.0):
    norm = 1 / (2 * np.pi * x_sig * y_sig)
    x_exp = (-1 * (w1 - x_mean) ** 2) / (2 * x_sig**2)
    y_exp = (-1 * (w2 - y_mean) ** 2) / (2 * y_sig**2)
    return norm * torch.exp(x_exp + y_exp)


def comb_func(w1, w2):
    z = -bivar_gaussian(w1, w2, x_mean=1.0, y_mean=-0.5, x_sig=0.2, y_sig=0.2)
    z -= bivar_gaussian(w1, w2, x_mean=-1.0, y_mean=0.5, x_sig=0.2, y_sig=0.2)
    z -= bivar_gaussian(w1, w2, x_mean=-0.5, y_mean=-0.8, x_sig=0.2, y_sig=0.2)
    return z


_ = plot_curve(comb_func, x_range=(-2, 2), y_range=(-2, 2), plot_3d=True, title="Steep optima")
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_71_0.svg

Most of the loss surface has very little to no gradients. However, close to the optima, we have very steep gradients. To reach the minimum when starting in a region with lower gradients, we expect an adaptive learning rate to be crucial. To verify this hypothesis, we can run our three optimizers on the surface:

[35]:
SGD_points = train_curve(lambda params: SGD(params, lr=0.5), comb_func, init=[0, 0])
SGDMom_points = train_curve(lambda params: SGDMomentum(params, lr=1, momentum=0.9), comb_func, init=[0, 0])
Adam_points = train_curve(lambda params: Adam(params, lr=0.2), comb_func, init=[0, 0])

all_points = np.concatenate([SGD_points, SGDMom_points, Adam_points], axis=0)
ax = plot_curve(comb_func, x_range=(-2, 2), y_range=(-2, 2), plot_3d=False, title="Steep optima")
ax.plot(SGD_points[:, 0], SGD_points[:, 1], color="red", marker="o", zorder=3, label="SGD", alpha=0.7)
ax.plot(SGDMom_points[:, 0], SGDMom_points[:, 1], color="blue", marker="o", zorder=2, label="SGDMom", alpha=0.7)
ax.plot(Adam_points[:, 0], Adam_points[:, 1], color="grey", marker="o", zorder=1, label="Adam", alpha=0.7)
ax.set_xlim(-2, 2)
ax.set_ylim(-2, 2)
plt.legend()
plt.show()
../../_images/notebooks_course_UvA-DL_03-initialization-and-optimization_73_0.svg

SGD first takes very small steps until it touches the border of the optimum. First reaching a point around (-0.75,-0.5), the gradient direction has changed and pushes the parameters to (0.8,0.5) from which SGD cannot recover anymore (only with many, many steps). A similar problem has SGD with momentum, only that it continues the direction of the touch of the optimum. The gradients from this time step are so much larger than any other point that the momentum m_t is overpowered by it. Finally, Adam is able to converge in the optimum showing the importance of adaptive learning rates.

What optimizer to take

After seeing the results on optimization, what is our conclusion? Should we always use Adam and never look at SGD anymore? The short answer: no. There are many papers saying that in certain situations, SGD (with momentum) generalizes better where Adam often tends to overfit [5,6]. This is related to the idea of finding wider optima. For instance, see the illustration of different optima below (credit: Keskar et al., 2017):

704f16e3acc44aa78dc78358110d36ad

The black line represents the training loss surface, while the dotted red line is the test loss. Finding sharp, narrow minima can be helpful for finding the minimal training loss. However, this doesn’t mean that it also minimizes the test loss as especially flat minima have shown to generalize better. You can imagine that the test dataset has a slightly shifted loss surface due to the different examples than in the training set. A small change can have a significant influence for sharp minima, while flat minima are generally more robust to this change.

In the next tutorial, we will see that some network types can still be better optimized with SGD and learning rate scheduling than Adam. Nevertheless, Adam is the most commonly used optimizer in Deep Learning as it usually performs better than other optimizers, especially for deep networks.

Conclusion

In this tutorial, we have looked at initialization and optimization techniques for neural networks. We have seen that a good initialization has to balance the preservation of the gradient variance as well as the activation variance. This can be achieved with the Xavier initialization for tanh-based networks, and the Kaiming initialization for ReLU-based networks. In optimization, concepts like momentum and adaptive learning rate can help with challenging loss surfaces but don’t guarantee an increase in performance for neural networks.

References

[1] Glorot, Xavier, and Yoshua Bengio. “Understanding the difficulty of training deep feedforward neural networks.” Proceedings of the thirteenth international conference on artificial intelligence and statistics. 2010. link

[2] He, Kaiming, et al. “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification.” Proceedings of the IEEE international conference on computer vision. 2015. link

[3] Kingma, Diederik P. & Ba, Jimmy. “Adam: A Method for Stochastic Optimization.” Proceedings of the third international conference for learning representations (ICLR). 2015. link

[4] Keskar, Nitish Shirish, et al. “On large-batch training for deep learning: Generalization gap and sharp minima.” Proceedings of the fifth international conference for learning representations (ICLR). 2017. link

[5] Wilson, Ashia C., et al. “The Marginal Value of Adaptive Gradient Methods in Machine Learning.” Advances in neural information processing systems. 2017. link

[6] Ruder, Sebastian. “An overview of gradient descent optimization algorithms.” arXiv preprint. 2017. link

Congratulations - Time to Join the Community!

Congratulations on completing this notebook tutorial! If you enjoyed this and would like to join the Lightning movement, you can do so in the following ways!

Star Lightning on GitHub

The easiest way to help our community is just by starring the GitHub repos! This helps raise awareness of the cool tools we’re building.

Join our Slack!

The best way to keep up to date on the latest advancements is to join our community! Make sure to introduce yourself and share your interests in #general channel

Contributions !

The best way to contribute to our community is to become a code contributor! At any time you can go to Lightning or Bolt GitHub Issues page and filter for “good first issue”.

Great thanks from the entire Pytorch Lightning Team for your interest !

Pytorch Lightning