Hello, where can we see the changes for the 0.5.0 ...

# kotlindl

[Alexandre Brown]

Hello, where can we see the changes for the 0.5.0 alpha1 ? W'ere upgrading from KotlinDL 0.4.0.
https://search.maven.org/artifact/org.jetbrains.kotlinx/kotlin-deeplearning-dataset/0.5.0-alpha-1/jar

Cheers!

[Alexandre Brown]

Also, is KotlinDL 0.5.0-alpha1 multiplatform ready?

[Alexandre Brown]

FYI I'm mainly interested in the changes regarding the preprocessing API, we have a new use case where images are now tiff images (usually with values 0-65536) and we cannot let KotlinDL load the image since it doesn't support this image type so we'd like to load the image ourselves and pass the image data to KotlinDL for it to apply the preprocessing functions to it.

Eg: We load the image into Multik array, pass that to KotlinDL and get a float array or multik array as output.
(If multik integration is not there then I guess any collection would do, as long as we can pass the data from the image we loaded ourselves).

Currently taking a look at https://github.com/Kotlin/kotlindl/blob/master/examples/src/main/kotlin/examples/dataset/PreprocessOnlyOneImage.kt

[Alexandre Brown]

Also, where did the normalize preprocesing moved to ? (We used to have this in 0.4.0)

transformTensor {
				normalize {
						mean = normalizeOptions.channelsMean.toFloatArray()
						std = normalizeOptions.channelsStd.toFloatArray()
					}
			}

[Julia Beliaeva]

Hi, sorry, changelog for the upcoming 0.5.0 is not done yet. All the preprocessing was moved to the kotlin-deeplearning-impl artifact. Yes, it is multiplatform (with the jvm and android variants).

[Julia Beliaeva]

We have preprocessing operations that use either BufferedImage or Pair<FloatArray, TensorShape> . You can use create your own Operation which converts you custom type to one of these types (depending of which KotlinDL operations you want to use). You can write something like this:

val preprocessing = pipeline<CustomImageType>()
        .toFloatArray()
        .normalize {
            // std =
            // mean =
        }
    preprocessing.apply(loadImageAsCustomImageType("/path/to/image"))

(you'll have to implement toFloatArray and loadImageAsCustomImageType is your function that loads images.

[Alexandre Brown]

Awesome! I will look into this, thank you so much this helps a lot

[Alexandre Brown]

Is preprocessing.apply(loadImageAsCustomImageType("/path/to/image")) needed if my data is already a float array?

[Julia Beliaeva]

I assumed you want to apply normalize preprocessing to it? apply applies preprocessing defined above to the data.

[Alexandre Brown]

Yes I want to apply normalize , nevermind I didnt read the method name properly, I get it , loadImageAsCustomImageType is just your way of saying whatever function that returns the custom type my bad

[Alexandre Brown]

Will try that, looks pretty straightforward from your explanation, I'll keep you posted. Thanks a lot, the new API looks pretty good btw, decoupling the loading from preprocessing is awesome

[Alexandre Brown]

Hello @Julia Beliaeva, here is what I have so far is that correct?
Any guidance is appreciated, thanks!

[Alexandre Brown]

@Julia Beliaeva The following works but I'm wondering how to do it using Operation

[Julia Beliaeva]

You can either implement your own dsl function similar to normalize , or use call to call your operation, like so:

val image: Image<Int> = //
val preprocessing = pipeline<Image<Int>>().call(ImageOperation()).normalize {
        //
}
preprocessing.apply(image)

[Julia Beliaeva]

Here is the updated CHANGELOG: https://github.com/Kotlin/kotlindl/blob/master/CHANGELOG.md