New predict command: how to pass input_block_shape?

[constantinpape]

It's unclear how to pass the value for input_block_shape in order to specify a fixed tiling shape.
The code

import bioimageio.core
import imageio.v3 as imageio
from xarray import DataArray

# For a reproducible example you can use the model test inputs.
# (I cannot easily figure out how to load them from the model ...)
image = imageio.imread("/home/pape/.cache/micro_sam/sample_data/livecell-2d-image.png")
model = bioimageio.core.load_description("hiding-tiger")

inputs = {model.inputs[0].name: DataArray(image[None, None], dims=("batch", "channel", "y", "x"))}

tile_shape = {"y": 512, "x": 512}
pred = bioimageio.core.predict(model=model, inputs=inputs, input_block_shape=tile_shape)

fails with

Traceback (most recent call last):
  File "/home/pape/Work/playground/bioimageio/new-pred/check_new_pred.py", line 13, in <module>
    pred = bioimageio.core.predict(model=model, inputs=inputs, input_block_shape=tile_shape)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/pape/Work/bioimageio/core-bioimage-io-python/bioimageio/core/prediction.py", line 100, in predict
    output = pp.predict_sample_with_fixed_blocking(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/pape/Work/bioimageio/core-bioimage-io-python/bioimageio/core/_prediction_pipeline.py", line 193, in predict_sample_with_fixed_blocking
    n_blocks, input_blocks = sample.split_into_blocks(
                             ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/pape/Work/bioimageio/core-bioimage-io-python/bioimageio/core/sample.py", line 68, in split_into_blocks
    assert not (

[FynnBe]

input_block_shape is annotated as Mapping[MemberId, Mapping[AxisId, int]]
so it expects {'input0': {AxisId("x"): 512, AxisId("y"): 512}}

I think we should improve handing of this argument in two ways:

1. accept AxisLike axis ids (such that "x" would be a valid axis identifier as well)
2. consider resolving unambiguous associations of axis ids to their input tensor ids, such that you attempted input would become valid

[constantinpape]

I see. It works with {'input0': {AxisId("x"): 512, AxisId("y"): 512}}. (But then runs into other issues, I am gonna open another issue).

We can move ahead with this for now, but longer term I think we need to either:

• simplify this (and other arguments) so that elementary python types are accepted
• provide an extremely good documentation

otherwise I don't see much hope of any outside users picking this up ...

[FynnBe]

yes, we should make the API a bit more forgiving to improve user experience, as well as continue to expand the documentation 👍

Let's keep this issue open until at least the input arguments to predict allow for simple python types.