I first wanted to post this on stackoverflow, but since it is not thát program-language related, but more on a conceptual level of data processing.. I thought it would fit better here. If things are unclear, please let me know.
I have a batch of 3d matrices (each of size 30 x 128 x 128), due to my problem definition I slice/reshape one such matrix to an array of length 3840 (=30 * 128) where each element is an array of size (1 x 1 x 128).
In 'normal' language: I transform the 3D matrix to an array of 1D vectors.
All of these 1D vectors are then fed into my Encoding Model, which has as input an 1D vector, and as output (a transformed version) of this 1D vector. Similar in dimensions.
After all the 1D vectors went through this Encoding Model, I concatenate them and pass them on to my Decoding Model. There it is compared to my target image, and a loss is calculated. I can't calculate a loss value from the output of the Encoding Model.
Having those 3840 Input variables (per 3D matrix) becomes a huge burden... Since I am also explaining the problem to myself now.. I want to reemphasize that I have the following requirements
- From the 3D matrix, I need to encode each 1D vector individually
- All the 1D vectors need to be assembled in a later stage for further processing
Either I need to drop a requirement, or maybe one of you has a smart solution to process multiple Input-variables/encoding-models in an efficient fashion.