Description

ctrueden

opened on May 16, 2024

Issue body actions

In [97]: from net.imglib2.outofbounds import OutOfBoundsFactory

In [98]: c = env.op("filter.gauss").inType(Img, Double.TYPE, OutOfBoundsFactory).outType(Img).computer()

In [99]: c
Out[99]: <java object 'org.scijava.ops.engine.matcher.impl.OpWrappers$Computer3OpWrapper$1GenericTypedComputer3'>

In [100]: c.infoTree().toString()
Out[100]: 'org.scijava.ops.image.filter.gauss.Gaussians.gaussRAISingleSigma(net.imglib2.RandomAccessibleInterval<I>,double,net.imglib2.outofbounds.OutOfBoundsFactory<I, net.imglib2.RandomAccessibleInterval<I>>,net.imglib2.RandomAccessibleInterval<O>)'

In [101]: sig = Ops.signature(c)

In [102]: sig
Out[102]: '|Info:org.scijava.ops.image.filter.gauss.Gaussians.gaussRAISingleSigma(net.imglib2.RandomAccessibleInterval<I>,double,net.imglib2.outofbounds.OutOfBoundsFactory<I, net.imglib2.RandomAccessibleInterval<I>>,net.imglib2.RandomAccessibleInterval<O>)@0-SNAPSHOT{}'

In [104]: try:
     ...:     o = env.opFromSignature(sig, Nil.of(Computers.Arity3))
     ...: except Exception as e:
     ...:     from scyjava import jstacktrace
     ...:     print(jstacktrace(e))
     ...:
java.lang.IllegalArgumentException: TODO
	at org.scijava.ops.engine.impl.DefaultOpEnvironment.opFromInfoChain(DefaultOpEnvironment.java:225)
	at org.scijava.ops.api.OpEnvironment.opFromInfoChain(OpEnvironment.java:165)
	at org.scijava.ops.api.OpEnvironment.opFromSignature(OpEnvironment.java:182)

Activity

1. ctrueden

   added

   bug

   on May 16, 2024
2. ctrueden

   added this to the unscheduled milestone on May 16, 2024
3. gselzer commented on May 28, 2024

   

   gselzer

   on May 28, 2024

   Member

   More actions

   @ctrueden thanks for raising this issue. I've thought about it a fair bit, and did a little prototyping to try and fix this issue, and I can't come up with a solution that I like - you get drawn into casting issues so easily. Here's what I've been thinking about:

   public <T> T opFromSignature(final String signature, final Class<T> rawOpType)

   This option is basically what you asked for above, but a little simpler without the Nil. While this is very easy to implement, it starts to reintroduce output casting in a way we tried to avoid with SciJava Ops.

   // some signature for a BiFunction<Img, double, Img>
   var sig = ...
   var img = ...
   var sigma = 5.0;
   var op = env.signature(sig, BiFunction.class); // This would return a raw BiFunction
   // -- EITHER -- //
   Img out1 = (Img) op.apply(img, sigma); // Requires cast
   // -- OR -- //
   var out2 = op.apply(img, sigma); // autocomplete only recognizes `Object` members on `out`.

   Then, I figured that if we're going for simplicity, maybe we should just do this:

   public Object opFromSignature(final String signature)

   But, while this look slick in a scripting environment, it'd make the casting even worse in languages where it's needed:

   // some signature for a BiFunction<Img, double, Img>
   var sig = ...
   var img = ...
   var sigma = 5.0;
   var op = (BiFunction<Img, double, Img>) env.signature(sig); // This would return an Object, cast required for following line
   Img out1 = op.apply(img, sigma);

   Since we worked so hard on the OpEnvironment methods to remove these casts from the original ImageJ Ops code, it feels like a step back, no?

   What I'd propose instead, if you really want this type of behavior, is a static function that could live...anywhere, really, although I'd prefer to put it closer to the scripts if possible (i.e. some future scijava-ops-python component down the line, that also contains the gateway?). It would look more or less like:

   public static Object reconstruct(final OpEnvironment env, final String signature) {
     InfoTree tree = env.treeFromSignature(signature);
     Nil<?> type = Nil.of(tree.info().opType());
     return env.opFromInfoTree(tree, type);
   }

   Then, in e.g. Python you could write:

   sig = ...
   img = ...
   sigma = 5.0
   
   out = Ops.reconstruct(env, sig).apply(img, sigma)

Issue actions