I'm having some problem :

I have this matrix array :

The dimension is : [1] 27455 784

I want to transform the 2nd Dimension into 28 x 28 array

Basically I want to change it into Convolution2D Input Shape (n_samples, height, width, channels) for TensorFlow The channels number is 1 because the image is in grayscale How can I transform the 784 into 28x28 array and add another array as channels = 1 ?

Thanks a lot!

I try this :

 # reshape 1-D data into 3-D data
train_x_img <- train_x
dim(train_x_img) <- c(27455, 28,28)

test_x_img <- test_x
dim(test_x_img) <- c(7172, 28,28,)

# check the dimensions of the 3-D data

dim(train_x_img)

But I guess it is wrong, because the value is not inserted in a correct order.

CodePudding user response：

When R handles indexing for arrays or matrices, it assumes that the ordering as primarily on the columns. You, however, appear to want to create a 25 x 25 matrix-slice based on successive rows of that larger matrix, so the first thing to do is transpose so the row values are in columns columns:

train_x_img <- t( train_x )  # now  784 x 27455

Fold the previous row values, now in columns by re-dimensioning:

dim(train_x_img) <- c(28,28, 27455)

And now, because you want the third dimension to be recast as the first index, use base R's aperm function:

train_x_img <- aperm(train_x_img, c(3, 1,2)

Demonstration with a somewhat smaller object:

x  <- matrix(1:24, 4, 6)  # 4 x 6 example
 x
     [,1] [,2] [,3] [,4] [,5] [,6]
[1,]    1    5    9   13   17   21   # desire rows to be "folded"
[2,]    2    6   10   14   18   22
[3,]    3    7   11   15   19   23
[4,]    4    8   12   16   20   24

a <- t(x)  # now 6 x 4
dim(a) <- c(2,3,4)  # fold the first dimension in two smaller dimensions
a2 <- aperm(a, c(3,1,2)) # reorder the indices

Examine the first entries indexed by the first index
 a2[1,,]
     [,1] [,2] [,3]
[1,]    1    9   17
[2,]    5   13   21   # success, that matrix is composed of the first row of x