Dependency of visualizers on backend leads to confusing error about trulens HOT 2 CLOSED

Makes sense, if anything, I would move the matplotlib assumption above the choosing any single dimension, since that one is more arbitrary and kind of a last ditch naive approach.

from trulens.

I would move the matplotlib assumption above the choosing any single dimension, since that one is more arbitrary and kind of a last ditch naive approach.

My thinking was that this ordering would lead to the greatest convenience. Basically, I like to use the Tiler class as a convenient for visualizing training data, even if I'm not using other TruLens features.

If we use the matplotlib assumption, then we will always pick (I think) the last dimension as the channel dimension (so we'll never reach any other cases in our "switch statement" suggested above). This is standard in tensorflow, but not pytorch. So if the user is using pytorch, or keras with 'channels_first', they may have some data that they would like to visualize that is naturally in NCHW format. In that case, they would have to transpose the dimensions before sending it to the tiler. On the other hand, this data would probably have a shape of something like (n, 3, x, y), in which case we could infer that dimension 1 is the channel dimension, as long as y doesn't happen to be 3. This would make it so that NCHW users could use the tiler with their preferred format without having to "know" that matplotlib wants NHWC format, and also without having to specify the backend to TruLens.

If it happens to be ambiguous for some reason we could warn them that we're using the default of NHWC.

from trulens.