Add output lengths for layers #491
Conversation
SebastianM-C
force-pushed
the
io_length
branch
from
7ea3e36 to
6f112c4
Compare
|
Can you also add them to the LuxCore API docs? |
|
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #491 +/- ##
==========================================
- Coverage 88.61% 87.85% -0.77%
==========================================
Files 29 29
Lines 1625 1630 +5
==========================================
- Hits 1440 1432 -8
- Misses 185 198 +13 ☔ View full report in Codecov by Sentry. |
test/layers/basic_tests.jl
Outdated
| @test LuxCore.inputsize(layer) == (10, 5) | ||
| @test LuxCore.outputsize(layer) == (10, 5) |
There was a problem hiding this comment.
@avik-pal is this ok? It seems a bit inconsistent with the size checks above.
There was a problem hiding this comment.
I think outputsize is fine. Do we really need to be able to define inputsize for Symbolics? I feel we should return a set of possible sizes for the input.
There was a problem hiding this comment.
No. I defined it only for consistency, but I'm not sure where it would be useful.
Would that be a tuple of tuples?
| @@ -9,6 +9,7 @@ | |||
| x = randn(rng, 6, 3) |> aType | |||
|
|
|||
| @test size(layer(x, ps, st)[1]) == (2, 3, 3) | |||
| @test Lux.outputsize(layer) == (2, 3) | |||
There was a problem hiding this comment.
I'm not entirely sure if the definition of outputsize makes sense for ReshapeLayer.
There was a problem hiding this comment.
I'm a bit confused about the fact that the size of the output in the test above (and also in general) has a different size compared to outputsize because the input also matters and we don't know it from the layer definition.
|
Should I just implement |
|
Yes let's restrict the current PR to |
SebastianM-C
force-pushed
the
io_length
branch
from
9b269f0 to
b73c174
Compare
You mean |
SebastianM-C
changed the title
|
@avik-pal can you approve ci? |
SebastianM-C
force-pushed
the
io_length
branch
from
b73c174 to
7279bdd
Compare
|
I rebased on master and fixed formatting. I'm not sure why codecov failed. |
|
@avik-pal Looks like CI passes, the codecov part at the end of the tests is unrelated and seems to randomly fail. |
This PR implements the
inputsizeandoutputsizeinterface defined in LuxCore in LuxDL/LuxCore.jl#19I'm not sure if defined this correctly for
ReshapeLayerandScale.Also, I'm not sure if the functions can be defined for some of the layers.