How could we convert reshaped `FloatArray` Back to...

# datascience

[Ananiya]

How could we convert reshaped FloatArray
Back to BufferedImage or save it as .jpg

[zaleslaw]

Do you keep the shape? Is it known before converting to the buffered image? Is it a reshaped float array of that? That is the one float in the float array?

[Ananiya]

It's usually 3024 by 3024 BGR 3 mb file
I was trying to convert the Image -> floatarray -> reshape -> image

mk.ndarray(ImageConverter.toNormalizedFloatArray(File("shoe.jpg")),28,28,1).toList().toFloatArray()

[zaleslaw]

So, as I understand the initial shape is 3024*3024*3 (BGR), but what is the desired shape? Looks like you want to resize, not reshape to 28*28*1 to create a greyscale image with size 28*28, isn't it?
Correct me if I'm wrong. What is the goal? Use the image in the LeNet trained on fashionMnist with KotlinDL?
Why did you use Multik? For reshape only?

[zaleslaw]

In this example for KotlinDL you could use ImagePreprocessing to load an image, resize it and save the image to the directory https://github.com/JetBrains/KotlinDL/blob/release_0.3/examples/src/main/kotlin/examples/dataset/PreprocessOnlyOneImage.kt
or I changed its example to be closer your needs

val imageResource = ImagePreprocessing::class.java.getResource("/datasets/vgg/image2.jpg")
    val image = File(imageResource!!.toURI())
    val preprocessedImagesDirectory = File("processedImages")

    val preprocessing: Preprocessing = preprocess {
        load {
            pathToData = image
            imageShape = ImageShape(3024, 3024, 3)
            colorMode = ColorOrder.BGR
        }
        transformImage {
           
            resize {
                outputWidth = 28
                outputHeight = 28
                interpolation = InterpolationType.NEAREST
                save {
                    dirLocation = preprocessedImagesDirectory
                }
            }
        }
        transformTensor {
            rescale {
                scalingCoefficient = 255f
            }
        }
    }

    val rawImage = preprocessing().first

[zaleslaw]

This is an example how to use ImagePreprocessing together with the Neural Network https://github.com/JetBrains/KotlinDL/blob/release_0.3/examples/src/main/kotlin/examples/dataset/VGGDogCatOnFly.kt

[Ananiya]

Sorry I am kinda new for DS
U somewhat got the idea
Idk if its possible but I was looking for a way to reshape the given image as the same way I used to train the model(fashionMinst) then predict it using TensorFlowInferenceModel

Thnx a lot for the help

[zaleslaw]

It's ok to ask the help with the Python or Kotlin part of your example here or in #kotlindl channel. I'll try to help. Usually, if you want to put the large image into the input of a neural net like LeNet with 28*28 input, you need to resize it first and if it has a plain input (784 instead of 28*28) you could reshape it to plain 1D FloatArray.

[zaleslaw]

Unfortunately, currently the KotlinDL doesn't support conversion of BGR images to greyscale

[Ananiya]

Oh I never thought the existence of #kotlindl channel I will ask there for next time
Btw thnx