Description
1 Regularizers
This time we will extend our framework to include common regularization strategies. As neural networks are extremely flexible statistical models they tend to show more variance and less bias. However this also means they are very prone to overfitting. Therefore, regularization is a very important part of deep learning. Throughout this exercise you will have to use baseclasses, which implement default members and methods accessible to each class derived from the respective base-class.
1.1 Refactoring
Since the behaviour of some layers changes depending on whether the network is currently training or testing, we need to extend our base-layer and refactor our neural network a little bit. Moreover, we choose to introduce a “base-optimizer”, which provides some basic functionality, in order to enable the use of regularizers.
Task:
• Add a boolean member testing phase to the BaseLayer. Set it to False by default.
• Implement a property phase in the NeuralNetwork class setting the phase of each of
its layers accordingly. Use this method to set the phase in the train and test methods.
• Create a base-class Optimizer for optimizers in the “Optimizers.py” file. Make all optimizers inherit from this “base-optimizer”.
• The class Optimizer should have a method add regularizer(regularizer) and a member storing the regularizer.
1.2 Optimization Constraints
Equality and inequality constraints on the norm of the weights are well known in the field of Machine Learning. They enforce the prior assumption on the model of small weights in case of L2-regularization and sparsity in case of L1-regularization.
Task:
Implement classes L2 Regularizer and L1 Regularizer in the file “Constraints.py” in the folder “Optimization”. Both have to provide the methods calculate gradient(weights) that calculates a (sub-)gradient on the weights needed for the optimizer. Additionally they have to
provide a method norm(weights), which is used to calculate the norm enhanced loss.
• Implement the two schemes. The constructor of each receives an argument alpha repre-
senting the regularization weight.
• Refactor the optimizers to apply the new regularizer if it is set using the calcu-
late gradient(weights) method.
• Refactor the NeuralNetwork class to add the regularization loss to the data loss. Use the method norm(weights) to get the regularization loss inside all layers (Fully Connected, Convolution and RNN) and sum it up.
You can verify your implementation using the provided testsuite by providing the commandline parameter TestConstraints.
1.3 Dropout
Dropout is a typical regularizer method for Deep Learning. It’s most often used to regularize fully connected layers. It enforces independent weights, reducing the effect of co-adaptation.
Task:
Implement a class Dropout in the file “Dropout.py” in folder “Layers”. This class has to provide the methods forward(input tensor) and backward(error tensor). This layer has no adjustable parameters. We choose to implement inverted dropout.
• Implement the constructor for this layer receiving the argument probability determining
the fraction units to keep.
• Implement the Dropout methods forward(input tensor) and backward(error tensor) for the training phase.
• Modify the Dropout method forward(input tensor) for the testing phase.
You can verify your implementation using the provided testsuite by providing the commandline parameter TestDropout.
1.4 Batch Normalization
Batch Normalization is a regularization technique which is conceptually very well known in Machine Learning but specially adapted to Deep Learning.
Task:
Implement a class BatchNormalization in the file “BatchNormalization.py” in folder “Layers”. This class has to provide the methods forward(input tensor) and backward(error tensor).
• Implement the constructor for this layer which receives the argument channels. chan-
nels denotes the number of channels of the input tensor in both, the vector and the image-case. Initialize the bias β and the weights γ according to the channels-size using the method initialize. This layer has trainable parameters, so remember to set the inherited member trainable accordingly.
• Implement the Batch Normalization methods forward(input tensor) and backward(error tensor) with independent activations for the training phase. Make sure you use an smaller than 1e-10.
• Hint: In Helpers.py we provide a function compute bn gradients(error tensor, input tensor, weights, mean, var) for the computation of the gradient w.r.t. inputs. Note that this function does not compute the gradient w.r.t. the weights.
• Implement the moving average estimation of training set mean and variance.
• Modify the Batch Normalization method forward(input tensor) for the testing phase. Use an online estimation of the mean and variance. Initialize mean and variance with the batch mean and the batch standard deviation of the first batch used for training.
• Implement the convolutional variant. The layer should change behaviour depending on the shape of the input tensor.
• Implement a method reformat(tensor) which receives the tensor that must be reshaped. Depending on the shape of the tensor, the method reformats the image-like tensor (with
4 dimension) into its vector-like variant (with 2 dimensions), and the same method reformats the vector-like tensor into its image-like tensor variant. Use this in the forward and the backward pass.
You can verify your implementation using the provided testsuite by providing the commandline parameter TestBatchNorm.
1.5 LeNet (optional)
By now our framework contains all the essential parts to build a convolutional neural network. However not every architecture can be trained to high performance. Therefore, successful architectures also play an important role. We implement a variant of a well known architecture called LeNet.
Task:
We implement a modification of a classic architecture called LeNet in file “LeNet.py” in folder “Models”.
• Add two functions (This means they should not be part of the class.) save(filename, net) and load(filename, data layer) to the NeuralNetwork file. Those two methods should use python’s pickle to save and load networks. After loading the network the
data layer should be set again.
• Exclude the data layer from saving by implementing the getstate () and setstate (state) methods. The setstate (state) method should initialize the
dropped members with None. This needs to be done, since the data layer is a generator-object, which cannot be processed by pickle.
• Implement the LeNet architecture as a function build() returning a network. Use a SoftMax layer and ReLU activation functions. Also ignore the fact, that the original architecture did not use padding for the convolution operation.
• Not typical for Deep Learning but useful, we also store the optimizer, initializer and their settings. Use the ADAM optimizer with a learning rate of 5 · 10−4 and a L2 regularizer of weight 4 · 10−4.
You can train this LeNet-variant on MNIST using the script “TrainLeNet.py”.
2 Recurrent Layers
Many machine learning problems require memorization capabilities along a particular dimension. Typical examples are time series prediction problems. A common method to address those are Recurrent Neural Networks (RNNs). To implement RNNs we can reuse most algorithms of our framework. The only necessary new components are new layers implementing the specific RNN cells.
2.1 Activation functions
Two common activation functions which we didn’t implement so far will come in handy: The hyperbolic tangent (Tangens Hyperbolicus) and the Sigmoid.
Task:
Implement two classes TanH and Sigmoid in files: “TanH.py” and “Sigmoid.py” in the folder “Layers”.
• Implement the operations forward(input tensor), backward(error tensor) for the TanH and the Sigmoid activation function.
• Specifically store activations for the dynamic programming component, instead of the input tensor. This is possible because the gradient involves only activations instead of the input (see slides).
2.2 Elman Recurrent Neural Network (RNN)
The type of recursive neural networks known as Elman network consists of the simplest RNN cells. They can be modularly implemented as layers.
Task:
Implement a class RNN in the file: “RNN.py” in folder “Layers”. This class has to provide all the functionalities required by a trainable layer for our framework.
• Write a constructor, receiving the arguments (input size, hidden size, output size).
Here input size denotes the dimension of the input vector while hidden size denotes the dimension of the hidden state. Initialize the hidden state with all zeros.
• Add a property memorize which sets the boolean state representing whether the RNN
regards subsequent sequences as a belonging to the same long sequence. In the constructor, initialize this member with False. Hint: think about how to transfer the ’memory’ from this sequence to the next sequence.
• Implement a method forward(input tensor) which returns a tensor that serves as the input tensor for the next layer. Consider the “batch” dimension as the “time” dimension of a sequence over which the recurrence is performed. The first hidden state for this iteration is all zero if the boolean member variable is False, otherwise restore the hidden state from the last iteration. You can choose to compose parts of the RNN
from other layers you already implemented.
• Implement a method backward(error tensor) which updates the parameters and returns the error tensor for the next layer. Remember that optimizers are decoupled from our layers. For the gradient calculation of some layers we need the input of the last forward pass at the respective point in time. Be sure to save those values in the forward pass and set them when backpropagating through time.
• Implement the accessor property gradient weights. Here the weights are defined as
the weights which are involved in calculating the hidden state as a stacked tensor. E.g. if the hidden state is computed with a single Fully Connected layer, which receives a stack of the hidden state and the input tensor, the weights of this particular Fully Connected Layer, are the weights considered to be weights for the whole class. In order to provide access to the weights of the RNN layer, implement a getter and a setter with a property for the weights member.
• To be able to reuse all regularizers, add the property to add an optimizer as
optimizer and to calculate the loss caused by regularization as calculate regularization loss()
as introduced in the regularization exercise. Finally add the method initialize(weights initializer, bias initializer) to use our initializers.
You can verify your implementation using the provided testsuite by providing the commandline parameter TestRNN.
2.3 Long Short-Term Memory (LSTM) (optional)
Elman networks severely suffer from the vanishing gradient problem. A common method to remedy this is to use more complicated RNN cells. The classical example is the LSTM cell. The unit tests for this class will only run when you create a file LSTM.py in your Layers folder.
Task:
Implement a class LSTM in the file: “LSTM.py” in folder “Layers”. This class has to provide all the functionalities required by a trainable layer for our framework.
• Write a constructor, receiving the arguments (input size, hidden size, output size).
Here input size denotes the dimension of the input vector while hidden size denotes the dimension of the hidden state. Initialize the hidden state with all zeros.
• Add a property memorize which sets the boolean state representing whether the RNN
regards subsequent sequences as a belonging to the same long sequence. In the constructor, initialize this member with False. That means if this state is True, the hidden state of one sequence processed in one forward call is carried over to next forward call as an initial state for the hidden state. Otherwise the hidden state is initialized with zeros every time a new sequence is processed. This is required to switch form BPTT to TBPTT.
• Implement a method forward(input tensor) which returns a tensor that serves as the input tensor for the next layer. Consider the batch dimension as the time dimension of a sequence over which the recurrence is performed. The first hidden state for this iteration is all zero if the boolean member variable is False, otherwise restore the hidden state from the last iteration. You can choose to compose parts of the LSTM from other
layers you already implemented.
• Implement a method backward(error tensor) which updates the parameters and returns a tensor which serves as error tensor for the next layer. Remember that optimizers are decoupled from our layers.
• Implement the accessor property gradient weights. Here the weights are defined as
the weights which are involved in calculating the hidden state as a stacked tensor. E.g. if the hidden state is computed with a single Fully Connected layer, which receives a stack of the hidden state and the input tensor, the weights of this particular Fully Connected Layer, are the weights considered to be weights for the whole class. In order to provide access to the weights of the LSTM layer, implement a getter and a setter with a property decorator for the weights member.
• To be able to reuse all regularizers, add the property to add an optimizer as
optimizer and to calculate the loss caused by regularization as calculate regularization loss() as introduced in the regularization exercise. Finally add the method initialize(weights initializer, bias initializer) to use our initializers.
You can verify your implementation using the provided testsuite by providing the commandline parameter TestLSTM.
3 Test, Debug and Finish
Now we implemented everything.
Task:
Debug your implementation until every test in the suite passes. You can run all tests by providing no commandline parameter. To run the unittests you can either execute them with python in the terminal or with the dedicated unittest environment of PyCharm. We recommend the latter one, as it provides a better overview of all tests. For the automated computation of the bonus points achieved in one exercise, run the unittests with the bonus flag in a terminal, with python3 NeuralNetworkTests.py Bonus
or set in PyCharm a new “Python” configuration with Bonus as “Parameters”. Notice, in some cases you need to set your src folder as “Working Directory”. More information about PyCharm configurations can be found here .
Make sure you don’t forget to upload your submission to StudOn. Use the dispatch tool, which checks all files for completeness and zips the files you need for the upload. Try python3 dispatch.py –help to check out the manual. For dispatching your folder run e.g.
python3 dispatch.py -i ./src -o submission.zip and upload the .zip file to StudOn.
