I have a semantic segmentation deep learning model which i want to deploy on kubeflow using TFX.
As I am moving the standalone DL code to TFX components I was having some questions
- The input images and masks will be stored in a tf-record. Would it be good practice to do the pre-processing like cropping, resizing, combining mask and image to ground truth, before the TFX pipeline starts (i.e ExampleGen)?
- Alternatively, would it be better practice to store the raw images and maskes in tf-record and then do pre-processing in Transform component of TFX?
- I also have some code for data augmentation during training. Would it be better to apply the augmentations in the trainer component or the transform component of TFX?
I would highly appreciate any pro tips or cautions to look out for in general!