How to use prune for CNNs?
7 Ansichten (letzte 30 Tage)
Ältere Kommentare anzeigen
I want to prune pretrained CNNs by certain rates (0.1-0.9). But when I try to use the prune function I get the following error:
Check for missing argument or incorrect argument data type in call to function 'prune'.
How can I use this function for CNNs (in this case resnet18) and is there a way to specify the pruning rate?
Thank you,
1 Kommentar
Antworten (2)
Nitin Kapgate
am 15 Jan. 2021
Currently the "prune" function does not provide the functionality to prune the network at specified pruninng rate.
The prune function removes zero-sized inputs, layers, and outputs from a network.
This leaves a network which may have fewer inputs and outputs, but which implements the same operations, as zero-sized inputs and outputs do not convey any information.
Maksym Tymchenko
am 2 Nov. 2023
In MATLAB, you can compress deep neural networks by using the Deep Learning Model Quantization Library, which is a free Add-On to the Deep Learning Toolbox.
This library contains a set of different functions that allow you to compress your deep learning network.
You mentioned pruning, this is a type of structural compression where the least important weights in the network are discarded making the network smaller. The "prune" function that you mentioned only accepts "shallow neural networks" as input. That is the reason why it did not work on resnet18. If you want to perform structural compression on a deep neural network like resnet18, you can use either:
- The taylorPrunableNetwork function which enables you to prune convolutional filters in network.
- The compressNetworkUsingProjection function which uses a MathWorks patented algorithm to analyze the activations of each layer using principal component analysis and reduce the size of the weights using projection.
Both of these functions work only on dlnetwork objects, so you might need to convert your network to that form in order to use them. For an example of how to prune convolutional filters in a CNN see:
0 Kommentare
Siehe auch
Kategorien
Mehr zu Recognition, Object Detection, and Semantic Segmentation finden Sie in Help Center und File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!