Load dlnetwork Object
Load a trained dlnetwork object and the corresponding classes.
Inspect the layers of the dlnetwork object. The network has three convolution layers at locations 2, 5, and 8 of the Layer array.
layers_1 =
12x1 Layer array with layers:
1 'input' Image Input 28x28x1 images with 'zerocenter' normalization
2 'conv1' 2-D Convolution 20 5x5x1 convolutions with stride [1 1] and padding [0 0 0 0]
3 'bn1' Batch Normalization Batch normalization with 20 channels
4 'relu1' ReLU ReLU
5 'conv2' 2-D Convolution 20 3x3x20 convolutions with stride [1 1] and padding [1 1 1 1]
6 'bn2' Batch Normalization Batch normalization with 20 channels
7 'relu2' ReLU ReLU
8 'conv3' 2-D Convolution 20 3x3x20 convolutions with stride [1 1] and padding [1 1 1 1]
9 'bn3' Batch Normalization Batch normalization with 20 channels
10 'relu3' ReLU ReLU
11 'fc' Fully Connected 10 fully connected layer
12 'softmax' Softmax softmax
Load Data for Prediction
Load the digits data for prediction.
Partition the data into pruning and validation sets. Set aside 10% of the data for validation using the splitEachLabel function.
The network used in this example requires input images of size 28-by-28-by-1. To automatically resize the images, use augmented image datastores.
Prune dlnetwork Object
Convert the dlnetwork object to a representation that is suitable for pruning by using the taylorPrunableNetwork function. This function returns a TaylorPrunableNetwork object that has the NumPrunables property set to 48. This indicates that 48 filters in the original model are suitable for pruning by using the Taylor pruning algorithm.
prunableNet_1 =
TaylorPrunableNetwork with properties:
Learnables: [14x3 table]
State: [6x3 table]
InputNames: {'input'}
OutputNames: {'softmax'}
NumPrunables: 48
Create a minibatchqueue object that processes and manages mini-batches of images during pruning. For each mini-batch:
Use the custom mini-batch preprocessing function preprocessMiniBatch (defined at the end of this example) to convert the labels to one-hot encoded variables.
Format the image data with the dimension labels "SSCB" (spatial, spatial, channel, batch). By default, the minibatchqueue object converts the data to dlarray objects with underlying type single. Do not format the class labels.
Train on a GPU if one is available. By default, the minibatchqueue object converts each output to a gpuArray if a GPU is available. Using a GPU requires Parallel Computing Toolbox™ and a supported GPU device. For information on supported devices, see GPU Computing Requirements (Parallel Computing Toolbox).
Calculate Taylor-based importance scores of the prunable filters in the network by looping over the mini-batches of data. For each mini-batch:
Finally, remove filters with the lowest importance scores to create a new TaylorPrunableNetwork object by using the updatePrunables function. By default, a single call to this function removes 8 filters. Observe that the new network prunableNet_2 has 40 prunable filters remaining.
prunableNet_2 =
TaylorPrunableNetwork with properties:
Learnables: [14x3 table]
State: [6x3 table]
InputNames: {'input'}
OutputNames: {'softmax'}
NumPrunables: 40
To further compress the model, run the custom pruning loop and update prunables again.
Extract Pruned dlnetwork Object
Use the dlnetwork function to extract the pruned dlnetwork object from the pruned TaylorPrunableNetwork object. You can now use this compressed dlnetwork object to perform inference.
Compare the convolution layers of the original and the pruned dlnetwork objects. Observe that the three convolution layers in the pruned network have fewer filters. These counts agree with the fact that, by default, a single call to the updatePrunables function removes 8 filters from the network.
conv_layers_1 =
3x1 Convolution2DLayer array with layers:
1 'conv1' 2-D Convolution 20 5x5x1 convolutions with stride [1 1] and padding [0 0 0 0]
2 'conv2' 2-D Convolution 20 3x3x20 convolutions with stride [1 1] and padding [1 1 1 1]
3 'conv3' 2-D Convolution 20 3x3x20 convolutions with stride [1 1] and padding [1 1 1 1]
conv_layers_2 =
3x1 Convolution2DLayer array with layers:
1 'conv1' 2-D Convolution 17 5x5x1 convolutions with stride [1 1] and padding [0 0 0 0]
2 'conv2' 2-D Convolution 18 3x3x17 convolutions with stride [1 1] and padding [1 1 1 1]
3 'conv3' 2-D Convolution 17 3x3x18 convolutions with stride [1 1] and padding [1 1 1 1]
Supporting Functions
Model Loss Function
The modelLoss function takes a TaylorPrunableNetwork object net, a mini-batch of input data X with corresponding targets T and returns activations in net and the gradients of the loss with respect to the activations in net. To compute the gradients automatically, this function uses the dlgradient function.
Mini Batch Preprocessing Function
The preprocessMiniBatch function preprocesses a mini-batch of predictors and labels using the following steps:
Preprocess the images using the preprocessMiniBatchPredictors function.
Extract the label data from the incoming cell array and concatenate into a categorical array along the second dimension.
One-hot encode the categorical labels into numeric arrays. Encoding into the first dimension produces an encoded array that matches the shape of the network output.
Mini-Batch Predictors Preprocessing Function
The preprocessMiniBatchPredictors function preprocesses a mini-batch of predictors by extracting the image data from the input cell array and concatenating into a numeric array. For grayscale input, concatenating over the fourth dimension adds a third dimension to each image, to use as a singleton channel dimension.
