how can i make alexnet accept 277x277x1 images

i got error in example TransferLearningUsingAlexNetExample Error using trainNetwork (line 140) The training images are of size 227x227x1 but the input layer expects images of size 227x227x3. my dataset 277X277x1

[SAQIB SHEIKH]

you can use Software Faststone Image resizer

[arun anoop m]

a=imread('                          ');
l=alexnet;
a=imresize(a,[227 227]);
s=classify(l,a);
imshow(a);

try this.

[Eline Geerits]

How can I apply this to an DataSet. I have abaout 4000 Images I have to resize so it would take way to long to do this for every Image indevidually.

Answer 1

[Brendan Hamm]

You can resize an image with the imresize function. Now since your images are of size 277x277x1 I will assume they are grayscale, but AlexNet was trained with RGB values and are thus 227x227x 3. It is not possible for you to recover color information from a grayscale image. You may be required to retrain the entire network with grayscale images.

I suggest for more questions on getting started with Deep Learning you check out the FREE Deep Learning OnRamp: https://matlabacademy.mathworks.com/

[abdo elsayed]

can i change the code to accept image of size 227 x 277 x 1 ??

[Brendan Hamm]

Sure. But this would require retraining the entire network. This may take weeks to months depending on your hardware.

[abdo elsayed]

ya i need to retraining how can i do that ?

any fast tutorial can i follow it i use 2017b matlab

[Hridya PI]

I completed the Alexnet Training with my dataset. How I can input an image and make the network predict it? I tried the code below.Butb it takes tha classnames of the pretrained network.Not the classnames that I newly created.

        function img = imageread(file)
  img=imread('D:\as.jpg');
            % Resize to match AlexNet input
            img = imresize(img,[227 227]);
            % Convert grayscale to color (RGB)
            img = repmat(img,[1 1 3]);
            imds1 = imageDatastore(fullfile(matlabroot,'toolbox','matlab','images','New Folder'),...
        'IncludeSubfolders',true,'ReadFcn',@imageread)
            img=read(imds1)
            imshow(img)
        end

but it shows error in @imageread. is this code right?

[abdo elsayed]

 Brendan Hamm 
how can retraining  ???

[Walter Roberson]

I suggest you consider converting your grayscale to RGB by taking

YourImage(:,:,[1 1 1])

or equivalently

repmat(YourImage, 1, 1, 3)

[Brendan Hamm]

@abdo You can retrain by following the process in the doc:

transfer-learning-using-alexnet

You can also learn more through the both free and hands on training course Deep Learning Onramp. I highly recommend this approach if you are just getting started.

[arun anoop m]

Thank you Brendan Hamm sir for recommending useful training course Deep Learning Onramp.

[Walter Roberson]

These days you should use an augmented image datastore as that can automatically resize and automatically convert to gray or rgb.

Answer 2

[Atika Aftab]

how can i make [224 224 1] img size ?

[arun anoop m]

I hope

a=imread('                          ');
a=imresize(a,[224 224]);

Answer 3

[Ayshath Afra]

The training images are of size 256x256x3 but the input layer expects images of size 224x224x3.

pls help me to solve this error

[Michael Keeling]

You can make an augmented datastore of the images that resizes them auomatically, while leaving your raw images the same

Answer 4

[Zaheer Hussain Fida]

can we extract the binary images features using alexnet.

[Walter Roberson]

Use https://www.mathworks.com/help/matlab/ref/matlab.io.datastore.transform.html to apply double() to the binary images so that they become grayscale images containing 0 (false) and 1 (true).

Use augmentedImageDatastore to automatically apply any rescaling needed to get the size needed by alexnet, 227 x 227 x 3 -- there is a color conversion option that can be used in the function.

And then hope that alexnet just happens to have been trained on a number of binary images with useful class names, so that instead of classifying as (say) 'Mathworks Screwdriver', it classifies as (say) 'Solidity 0.8 Eccentricity 0.14 MajorAxes 85' .

... But I must admit, I think the chances are low.

I think you would have more success if you used Transfer Learning rather than hoping that alexnet is already able to identify the exact set of binary features that you want.