Datablock for multiple inputs

[anandms (Anand)]

Noob here. Pardon if this is already covered elsewhere. I am trying to build a certain model:
Inputs: one Image, two float values which are metadata
Output: one float value

Does the DataBlock API allow me to do this directly?

[anandms (Anand)]

I have a hunch the solution has got something to do with fastai - Vision learner

Not able to find examples or explanations. Hmm…

[Archaeologist (Jürgen Landauer)]

Jeremy has an example notebook for a similar task i.e. Multi target :

Multi-target: Road to the Top, Part 4

[anandms (Anand)]

Had seen that. It shows image input along with category input, not continuous input. I did a naive try with a RegressionBlock as input, but that did not work.
Thank you for the tip, though

[Archaeologist (Jürgen Landauer)]

I guess you just need a TransformBlock that loads a torch floatTensor. Does this discussion help?

Loading Torch Tensors into the DataBlock API

There’s likely a better way to do this but here’s the implementation I came up with to load some dummy tensors. from fastai.vision.all import * def get_mask(fname, masklabel, tensorlabel): path_pieces = fname.parts new_path = [] for i in range(len(path_pieces)): if path_pieces[i] == tensorlabel: new_path.append(masklabel) continue new_path.append(path_pieces[i]) mask_path = os.sep.join(new_path) return Path(mask_path) def MaskGette…

[anandms (Anand)]

I think TransformBlock is the way to go, thank you for that. The example you pointed to seems a little too advanced for my current level of experience with python/ai/fastai/etc

I am checking out
https://medium.com/mlearning-ai/constructing-a-datablock-using-fastai-7703752ae17a
and

Lesson 2 - Tabular Regression and Permutation Importance | walkwithfastai