trainNetwork requiring 1x2 cell array from a Transformed image datastore

I am trying to train a network using the deep network designer. I started by creating a datastore and then transforming the images to grayscale. These images are then sent to the deep network designer. However, I keep getting an error saying that I need a 1x2 cell array despite using an image datastore.

Below is the code used before:

ds = imageDatastore("Training","IncludeSubfolders",true, "LabelSource","foldernames");

dsnew = transform(ds,@(x) rgb2gray(x));

The images are all 512 x 512, so in ImageInputLayer I have the InputSize set to 512 x 512 x 1.

Answer 1

[Sourabh]

Hi Rayan,

The error message you are seeing might be because the output of transform is not returning the data in the expected format for your neural network.

The “transform” function outputs a “TransformedDatastore” object. With the Deep Network Designer, you can import and train any datastore object that works with the trainNetwork function.

To pre-process images according to your own pipeline, try using the transform and combine functions.

For more information on pre-processing images for deep learning applications, see Preprocess Images for Deep Learning.

As a last resort solution for this issue, you could consider using the “Image Batch Processor” app from the Image Processing Toolbox to perform any pre-processing operations on your image dataset and save them to your local disk.

However, I would only recommend this option if your dataset were small, and you would not mind having a second copy of the images on your disk.