Hi everyone,
First, I would like to thank everybody involved in this project.
The library is simply amazing!
To my problem:
I am working with fastaiv2 for the first time.
So maybe I am missing something obvious.
I am working with images that have metadata available.
I would like to be able to optionally include the metadata.
The metadata should not be coupled too tightly with my other components, as the metadata will probably be used for very different tasks. (Influence sample weights, maybe included in the loss function, etc.)
I have sub-classed TensorImage
because the input requires extra initialization steps and a different show
method. (The input has more than 3 channels)
First, I thought that I could simply save the additional metadata during the type_tfm
to my tensor, but later I saw that the metadata gets overwritten during the collation step (or better, only the metadata of the first tensor is saved, if there are multiple attributes with the same name).
I tried to give a more detailed explanation in the following colab notebook:
But the question boils down to:
How would you include optional metadata, that will be used for different experiments, and still be able to reuse most of the pipeline?
Sorry if it seems too abstract. I hope that the colab notebook helps to show what my issue is.
Thanks!