Show tensors shaped [H, W, 1, 1] as images (and more!) #2075
Conversation
… an image. Fixes #1871 Quite a bit of nuance to support single channel 1x1 images & line-like images.
Wumpf
force-pushed
the
andreas/handle-trailing-dimensions
branch
from
10205f3 to
2efeb12
Compare
| .enumerate() | ||
| .rev() | ||
| .find(|(_, dim)| dim.size != 1) | ||
| .map_or(&self.shape[0..1], |(i, _)| &self.shape[..(i + 1)]) |
There was a problem hiding this comment.
I'm starting to think we should have a struct Shape(Vec<TensorDimension>). It would also solve the formatting of shapes for which we currently use the ugly Debug formatting in a bunch of places (one more in this PR).
Oh well, out of scope for this PR
There was a problem hiding this comment.
I added a comment about this to #1992
| match shape_short.len() { | ||
| 1 => { |
There was a problem hiding this comment.
if shape_short would return just the integers (not the names) we could make this match a lot nicer;
| match shape_short.len() { | |
| 1 => { | |
| match shape_short { | |
| [h] => {…}, | |
| [h, w] => {…}, | |
| [h, w, c] => {…}, |
There was a problem hiding this comment.
…but it would require an allocation, or that we change how shape is stored (see #1992 (comment)), so let's not right now
|
(almost done addressing everything, but logging off now; finishing tomorrow) |
|
nuancing away! This one might warrant a short re-review. |
| dim => self.image_height_width_channels().and_then(|_| { | ||
| self.get( | ||
| &[x, y, channel] | ||
| .into_iter() | ||
| .chain(std::iter::repeat(0).take(dim - 3)) | ||
| .collect::<Vec<u64>>(), | ||
| ) | ||
| }), |
There was a problem hiding this comment.
I think we should return None if using get_with_image_coords on a tensor that is non-image like, e.g. 8x7x6x5x4x3
So maybe just check if self.short_shape().len() <= 3 in this branch
There was a problem hiding this comment.
the check in here implied by self.image_height_width_channels().and_then is actually stronger and would also reject e.g. 3x3x5
So I think this is alright this way (albeit more expensive)
* Ignore trailing tensor dimension when determining whether a tensor is an image. Fixes #1871 Quite a bit of nuance to support single channel 1x1 images & line-like images. * fix image preview for images other than M x N x C and M x N * comment fix * better shape_short comment * handle empty tensors * is_shaped_like_an_image is now defined via image_height_width_channels, improve comment on both * any Nx1x... image now now treated as image * rename to get_with_image_coords, make it more restrictive * tensor_to_gpu height_width_depth utility uses now tensor.image_height_width_channels * change behavior of is_vector
Ignore trailing tensor dimension when determining whether a tensor is an image.
Fixes #1871
There's a surprising amount of nuance in order to support single channel 1x1 images & line-like images.
Checklist
PR Build Summary: https://build.rerun.io/pr/2075