Semantic Segmentation - How many layers to replace in transfer learning?

Im doing semantic segmentation using Resnet-18 with Deeplab v3+ (https://www.mathworks.com/help/vision/examples/semantic-segmentation-using-deep-learning.html)

However, I want to retain on progressively harder tasks and want to use transfer learning. How many of the final layers should I be replacing? How do I figure out how many layers I have?

On the analyzeNetwork it says I have 101 but I am using Resnet - 18 which I thought had much less?

[Tran Thanh]

I can not find see helperDeeplabv3PlusResnet18.m file. Where can I find this file? Would you mind to send this file for me?

[Philip Li]

Hi, I also can not find this file anywhere. Were you able to get it?

[Tohru Kikawada]

Philip, you'll need to try this on R2019a since the example has been revised to use DeepLab v3 instead of SegNet in the latest vresion.

[Guy Reading]

I've been doing a bit of digging on the resnet18() <-> DeepLab V3+ connection, in this link MATLAB writes:

"This example creates the Deeplab v3+ network with weights initialized from a pre-trained Resnet-18 network"

But then we might ask, "how do we use the weights trained on one network to be used in another? Won't they be meaningless relative to Deeplab?

Then this link writes:

"The latest implementation of DeepLab supports multiple network backbones, like MobileNetv2, Xception, ResNet-v1, PNASNET and Auto-DeepLab."

So I guess we can treat Deeplab V3+ as some form of extension of resnet18 and thus can use the weights.

Answer 1

[Tohru Kikawada]

Did you see helperDeeplabv3PlusResnet18.m which is attached to the example as a supporting file? The supporting function might be helpful to create your own transfer learning network.

[awezmm]

Thank you for your response,

I looked at it and it not too helpful becasue I dont think it is adding or removing any layers after adding a decoder to the original network.

I tried using this function

[learnableLayer,classLayer] = findLayersToReplace(lgraph);

from the page https://www.mathworks.com/help/deeplearning/examples/train-deep-learning-network-to-classify-new-images.html

However, it is giving me an error: Layer graph must have a single classification layer. This is becasue that function seems to be setup only for image classification not semantic segmentation.

I need implement transfer learning for semantic segmentation. Any advice?

Thank you

[Tohru Kikawada]

Type the following command

openExample('deeplearning_shared/SemanticSegmentationUsingDeepLearningExample')

then the helperDeeplabv3PlusResnet18.m can be found in your current folder.

[awezmm]

I have already looked at the helperDeeplabv3PlusResnet18.m and said in my previous comment that is want not helpful...

[Tohru Kikawada]

Sorry for the confusion. Someone asked me where the file can be found.

Answer 2

[Guy Reading]

Did you make any progress on this, @awezmm? I'm looking to do the same thing as you, with Resnet-18, too, and I got stuck at the same point as you so Googled & found your question here!

So far I've followed the (adapted for resnet) instructions of this tutorial:

https://uk.mathworks.com/videos/deep-learning-with-matlab-transfer-learning-in-10-lines-of-matlab-code-1487714838381.html

So for resnet that'd be:

%% load a pre-trained CNN
pretrainedFolder = fullfile(tempdir,'pretrainedNetwork');
pretrainedNetwork = fullfile(pretrainedFolder,'deeplabv3plusResnet18CamVid.mat'); 
data = load(pretrainedNetwork); 
net = data.net;
layers = net.Layers

This shows me all 101 layers for resnet. For me, personally, I'd like to classify 2 things (background or object) so I've edited the final layer to give me 2 things, but I'm pretty sure I need to do more layers and unsure which ones:

%% Modify the network to use 2 categories
layers(101) = pixelClassificationLayer; % note, in the example he uses classificationLayer as it's not semantic seg

& now I'm stuck! I'll comment r.e. letting you know any progress I've made... I'm going to look into the structure of resnet more, now, to get a better understanding of what I need to change and how...

[Guy Reading]

Update:

I think these are the layers you need to change:

%% Modify the network to use 2 categories
numClasses = 2;
layers(97) = convolution2dLayer(1,numClasses);
layers(98) = transposedConv2dLayer(2, 8, 'Stride', 4, 'Cropping', 2);
layers(99) = crop2dLayer('Name', 'dec_crop2'); % crop layer requires connections
layers(100) = softmaxLayer();
layers(101) = pixelClassificationLayer;

If you stop here and then look at what we've made using the following function, we can see that we've broken a lot of connections of the CNN:

analyzeNetwork(layers);

So right now I'm looking to try and re-build those. I think we can do that with the following functions but I'm not sure...

%% creating connections
% Create a layerGraph. The first input of crop2dLayer is automatically
% connected to the first output of the image input layer.
lgraph = layerGraph(layers);

% Connect the second input to the image layer output.
lgraph = connectLayers(lgraph, 'image', 'crop/ref') 

Examples of the above functions taken from here: https://uk.mathworks.com/help/deeplearning/examples/train-deep-learning-network-to-classify-new-images.html

[Guy Reading]

Okay I've got a network with no errors showing when analysing it with analyzeNetwork(), I'll see if this has actually adapted the network correctly tomorrow, though, when I feed it some data. I couldn't use the function noted in the previous comment, connectLayers, as it doesn't work with a DAGNetwork:

%% load a pre-trained CNN
pretrainedFolder = fullfile(tempdir,'pretrainedNetwork');
pretrainedNetwork = fullfile(pretrainedFolder,'deeplabv3plusResnet18CamVid.mat'); 
data = load(pretrainedNetwork); 
net = data.net;
layers = net.Layers;
lgraph = layerGraph(net);

%% Modify the network to use 2 categories
numClasses = 2;
objectCategories = categorical({'Lump','Background'});

newScorerLayer = convolution2dLayer(1,numClasses, ...
    'Name', 'NewScorer',...
    'NumChannels', 256,...
    'WeightLearnRateFactor', 1,...
    'BiasLearnRateFactor', 1);

newdec_ups2Layer = transposedConv2dLayer(8, 2, ...
    'Stride', 4, ...
    'Cropping', 2, ...
    'Name', 'newdec_upsample2',...
    'NumChannels', 2);

newDec_crop2Layer = crop2dLayer('centercrop', ...
    'Name', 'dec_crop2'); % crop layer requires connections

newSoftmaxLayer = softmaxLayer('Name', 'newSoftmax-out');

newLabelsLayer = pixelClassificationLayer('Name', 'newLabels',...
    'Classes', objectCategories);

% taken from: https://uk.mathworks.com/help/deeplearning/examples/train-deep-learning-network-to-classify-new-images.html
lgraph = replaceLayer(lgraph, 'scorer', newScorerLayer);
lgraph = replaceLayer(lgraph, 'dec_upsample2', newdec_ups2Layer);
lgraph = replaceLayer(lgraph, 'dec_crop2', newDec_crop2Layer);
lgraph = replaceLayer(lgraph, 'softmax-out', newSoftmaxLayer);
lgraph = replaceLayer(lgraph, 'labels', newLabelsLayer);

% check we have something that looks okay
analyzeNetwork(lgraph);

[awezmm]

Hey Guy,

I think it's best to just replace the last layer only. Why are you replacing the other ones?

[Guy Reading]

Oh right! So I looked to replace the layers which are specific to the amount of classes we want to categorise into. So layers 97:101 refer to layers which have a dimension set to 11, which was the original amount of classes:

Which is why I chose those. But I'm not 100% certain! Have you got this working for you at all? / what's the intuition behind picking the last layer, only?

[Guy Reading]

To all that are reading: the above method worked for me & I'm starting to get labelled images from my model. I'm still not sure about which layers to freeze, if there are any suggestions with that I'd be interested to hear!