how to use multiple input layers in DAG net as shown in the figure

I have DAG graph with two paths of layers inside it.

I am planning to feed this DAG with two types of data (D1, D2) but I can't do it as the DAG in matlab accept just one input layer.

I need to form a layer such as:

I noticed that there is a custom network that can provide a network with multiple inputs but how can I connect between this network and DAG graph? or how could I use DAG with two inputs?

[Ben Hur]

Guys I really confused with how to use two inputs data with Dag graph, Any idea?

[Ben Hur]

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

[Ville Laukkanen]

This would be nice to have an answer to.

I have a similar situation with three image + three float-variables regression case. We're trying to estimate the output of an industrial process with images of material flows in from three different lines and their respective line-speed (float). I get the true output result much later. Would like to train the whole image regression thing together.

Maybe some modified version of LSTM would work or perhaps some funny layer which would decompose the input to six different layergraph-lines, but I can't find a way to do this in MatLab.

On Python Tensorflow there is the node structure and inputs given in dictionary (matlabs' struct). Would there be a way to do this in Matlab? - Input to several points in an layer graph.

[Kenta]

As of 2019b, you can use custom training loop which allows you to do multi-input CNN.

https://jp.mathworks.com/matlabcentral/fileexchange/74760-image-classification-using-cnn-with-multi-input

This shows a demo to classify images with two-path sequence layers using two kinds of input images.

Answer 1

[Mahmoud Afifi]

One idea is to feed the network with concatenated inputs (e.g., image1;image2) then create splitter layers that split each input. The problem here is that you have to feed the network with .mat files, not image paths. Another idea is to store your images as tiff files which can hold 4 channels. In this case, you can store a colored image (3 channel) and a grayscale one. Have a look at this example https://www.mathworks.com/matlabcentral/fileexchange/65065-two-stream-cnn-for-gender-recognition-using-hand-images?s_tid=FX_rc1_behav .. see twoStream.m file.

[Koshy]

Can you give the code of splitter layer? Or how it can be implemented?

[Mahmoud Afifi]

Here is an example

%assume your input has 6 channels (i.e., two images stacked together

channels = 6; 

inputSize = [224,224,channels]; %change it as you want

lgraph = layerGraph;

input_layer = imageInputLayer(inputSize,'Name','Input_Layer');
lgraph = addLayers(lgraph,input_layer);
for i = 1 : channels
    eval(sprintf('ch_%d_splitter = convolution2dLayer(1,1,''Name'',''channel_%d_splitter'',''WeightLearnRateFactor'',0,''BiasLearnRateFactor'',0,''WeightL2Factor'',0,''BiasL2Factor'',0);',i,i));
    eval(sprintf('ch_%d_splitter.Weights = zeros(1,1,channels,1);',i));
    eval(sprintf('ch_%d_splitter.Weights(1,1,%d,1) = 1;',i,i));
    eval(sprintf('ch_%d_splitter.Bias = zeros(1,1,1,1);',i));
    eval(sprintf('lgraph = addLayers(lgraph,ch_%d_splitter);',i));
    eval(sprintf('lgraph = connectLayers(lgraph,''Input_Layer'',''channel_%d_splitter'');',i));
end

%Here, let's assume you want to use the first three layers as input for
%the first stream and the last three layers as input for second stream

input_stream_1 = depthConcatenationLayer(3,'Name','input_stream_1');
lgraph = addLayers(lgraph,input_stream_1);

input_stream_2 = depthConcatenationLayer(3,'Name','input_stream_2');
lgraph = addLayers(lgraph,input_stream_2);

lgraph = connectLayers(lgraph,'channel_1_splitter','input_stream_1/in1');
lgraph = connectLayers(lgraph,'channel_2_splitter','input_stream_1/in2');
lgraph = connectLayers(lgraph,'channel_3_splitter','input_stream_1/in3');


lgraph = connectLayers(lgraph,'channel_4_splitter','input_stream_2/in1');
lgraph = connectLayers(lgraph,'channel_5_splitter','input_stream_2/in2');
lgraph = connectLayers(lgraph,'channel_6_splitter','input_stream_2/in3');

%now you can continue building your two-stream network. The first input
%will come from input_stream_1 and the second input will come from
%input_stream_2

[Yuze Zou]

Thank you so much! Haha, maybe you can put more illustration on the codes. It's a nice trick to split the input layer using convolution2dLayer with preset weights and bias. The WeightLearnRateFactor=0 and etc. settings keep the added layers unchanged during the learning process.

[dinial utami]

Mr, you just make multiple input in convolution, but not in trainNetwork Mr?

I see in trainNetwork just rever to one image.

[miao li]

Hello, I have a question, how to separate each channel using 1X1 convolution? I think that each convolutional layer operates on 6 channels separately and then adds them, but the channels cannot be extracted.

Answer 2

[Mahmoud Afifi]

I just released an example Matlab code of how to implemenet multiple-input CNN in Matlab 2019b. You can find it here:

https://www.mathworks.com/matlabcentral/fileexchange/73138-multi-stream-cnn

please if it works for you, accept this answer.

[dinial utami]

thank you for your helping Mr.

in the code you have share, has multiple input in layer. not in trainNetwork.

Mr, can you help if we has 3 input in different image for training set, we set 3 input layers, but we can't set 3 training set. in the reality we need 3 input layers, and 3 training set.

thank you Mr. Mahmoud Afifi

Answer 3

[Shounak Mitra]

Hi Marcello and Arjun,

Support for multiple Input layers are not supported as of the 18b release. We are working on it and it should be available soon.

Thanks Shounak

[abir zendagui]

Hi,

Is the multiple input layers are really supported now (In 18b)? if it' is the case ,how this is done please?

[Bodo Rosenhahn]

Hi,

are multiple input /output layers for DAG networks supported in 19a ? Can you provide an example ?

Answer 4

[Bernhard Suhm]

Modeling DAG graphs with multiple inputs and/or outputs is currently not supported in our deep learning framework, but we are working on it. So hold your breath for one of the next releases.

[Mammo Image]

@ Bernhard Suhm, So any alternative suggestion for now?

[Sara Abdeldayem]

Is it in 2018a?

[Bernhard Suhm]

It's our policy to not officially commit to functionality prior to the release date, but more support for DAG is planned for 18a.

[Arjun Desai]

Is it supported now?

[Marcello Venzi]

Hello, can you please comment if multiple input layers are now supported (as of maltab 2018b)? I could not find this option in the documentation.

Answer 5

[Yanhui Guo]

In the DAGNetwork file, I found the property: InputLayerIndices. In the fasterrcnn, I also found two input for this network. I am wondering if matlab2018b has an indirect way to support multiple inputs? Thanks.

Answer 6

[sinan salim]

hi is there any update to manage multi-input layer >>because i want use different classes each 2 classes have to be assign for separate input layer