Can we do away with acquireImageFxn? Or generalize it.

[adeldaoud]

@cjerzak I am fiddling around with the acquireImageFxn to adapt it to my simulated image data, where the bands and dimensions are different from how acquireImageFxn is defined in your simulated data.

This is the code in your simulated data

  # Define acquireImageFxn
  acquireImageFxn <- function(keys) {
    m_ <- FullImageArray[match(keys, KeysOfImages),,,]
    if(length(keys) == 1) {
      m_ <- array(m_, dim = c(1L, dim(m_)[1], dim(m_)[2], dim(m_)[3]))
    }
    return(m_)
  }

These function that feeds WriteTFRecord.

if(reSaveTFRecord) {
  causalimages::WriteTfRecord(
    file = TFRecordName_im,
    uniqueImageKeys = unique(KeysOfObservations),
    acquireImageFxn = acquireImageFxn
  )
}

But how do we handle varying image classes dynamically?

I will see if I can help improving this part of the code if you do not have a readily available piece of general code already,

[cjerzak]

WriteTfRecord should handle images of varying channel number. i agree that the function does add complexity as a pre-analysis step. in practice, there are large speed gains from using some kind of sequentialized data format (e.g., in the markus pipeline the reading of images into memory takes up >50% of all clocktime; sequentialized formats like tfrecord bring that down (in causalimages, to less than 1%). in short, it is an extra step and can introduce potential issues (like needing to re-write the record if changes are made to the data structure), but the gains are considerable speedwise.