Can we change the input size of a pretrained network for transfer learning

Asked by Arjun Desai on 26 May 2018

Latest activity Commented on by Lenin Falconi on 16 Apr 2019

57 views (last 30 days)

I want to use transfer learning on the Resnet-50 architecture trained on Imagenet. I noticed that the input size into the Resnet-50 architecture is [224 224 3]. However my images are [150 150 3]. I was wondering if there were a way to change the input size of the input layer rather than resizing my images.

2 Comments

Jake on 20 Sep 2018

I am also interested in this answer.

Lenin Falconi on 14 Apr 2019

Hello I think you could try with the following code:

%% Load Data
% unzip('v_200x400.zip');  
imds = imageDatastore('v_200x400', ...
    'IncludeSubfolders',true, ...
    'LabelSource','foldernames'); 
%% Resizing images to fin googleNet
imds.ReadSize = numpartitions(imds);
imds.ReadFcn = @(loc)imresize(imread(loc),[224,224]);

This code I used to resize images from bigger than 224 to 224. However since you are upsampling. This routine might not be of much help since we are not controlling the interpolation needed to fill the missing pixels when you go from 150 to 224. I would you suggest to check

https://la.mathworks.com/help/images/ref/imresize.html Pay attention to Interpolation Methods some are better than others to upsampling. Others better for downsampling

I leave there this code that I used although it is python, it could be written in Matlab code:

# Rescaling:

# We have two problems if image is smaller than desired or bigger than desired

http://tanbakuchi.com/posts/comparison-of-openv-interpolation-algorithms/

# Upsampling:

# image width < new width, and image Height < new height

if (width<=new_w) and (height<=new_h):

# USE CUBIC INTERPOL

img_scaled = cv2.resize(img_seg, (new_w, new_h), interpolation=cv2.INTER_CUBIC)

else:

# Downdampling

img_scaled = cv2.resize(img_seg, (new_w, new_h), interpolation=cv2.INTER_AREA)

Products

Deep Learning Toolbox

Release

R2018a

1 Answer

Link

Answer by BERGHOUT Tarek on 16 Apr 2019

yes this methode is cold:

1- if you are changing the neumber of neurons from N to n where N>n: this is called :'constructive' learning

2- if you are changing the neumber of neurons from N to n where N<n: this is called :'distructive' learning

but retraining alwayse requeired , not from the begining but from the final weights that u have in final net

if you want to found some papaers in this area use this key words :

neural networks with additive hidden nodes ; distructive neural nets .....etc. good luck.

1 Comment

Lenin Falconi on 16 Apr 2019

I am having a hard time understanding your answer. Sure that is something new to me about constructive and distructive learning.

However, it seems like Arjun Desai wants to change the input layer of Resnet 50 so that he can test transfer learning on his/her images. As far as I understand, since Resnet 50 is trained with specific natural image dataset with the dimensions 224x224x3, I don't think the input layer could be changed because this would affect all internal dimensions of the ConvNet arquitecture.

Because of that I aided with a code to simply change image size.

About retraining, not from the beginning I agree totally with you. That's what should happen with transfer, but in order to that happen The image of interest must go forward through the ConvNet until it reaches the las Full Connecting Layers to start retraining or a pertinent node in case of fine tuning. But again image should be in the dimensions expressed by the architecture.

Time ago I found a paper that proposed an arquitecture to deal with images from different sizes. If I found it I will share the name....

Products

Deep Learning Toolbox

Release

R2018a

You are now following this question

Arjun Desai (view profile)