[bug] javacpp -pytorch UpsampleImpl have bad bug ,can not work

[mullerhai]

Hi, @saudet
UpsampleImpl javacpp have bad bug, the same input parameter cannot work in javacpp ,please fix up

in python example
···
input_3x3 = torch.zeros(3, 3).view(1, 1, 3, 3)
print(input_3x3.shape)
m = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
# Notice that values in top left corner are now changed
output = m(input_3x3)
print(output.shape)

···

python console

input_3x3  torch.Size([1, 1, 3, 3])
output size torch.Size([1, 1, 6, 6])

in javacpp

    val input3 = torch.zeros(Seq(3, 3)).view(1, 1, 3, 3)
    val options: UpsampleOptions = UpsampleOptions()
    options.scale_factor().put(DoubleVector(2))
    options.align_corners().put(true)
    options.mode().put(new kBilinear())
//    options.size().put(LongPointer(3,3))
    val model = UpsampleImpl(options)

    println(s"  align ${model.options().align_corners().get()}  scala factor ${model.options().scale_factor().get().get(0)} mode ${model.options().mode()}")
    val output = fromNative(model.forward(input3.native))
    println(s"output ${output.shape}")

console log

  align true  scala factor 0.0 mode org.bytedeco.pytorch.UpsampleMode[address=0x25be4257950,position=0,limit=0,capacity=0,deallocator=null]

java.lang.RuntimeException: Input and output sizes should be greater than 0, but got input (H: 3, W: 3) output (H: 0, W: 0)
Exception raised from upsample_2d_common_check at D:\a\javacpp-presets\javacpp-presets\pytorch\cppbuild\windows-x86_64-gpu\pytorch\aten\src\ATen/native/UpSample.h:165 (most recent call first):
00007FF8894F83C9 <unknown symbol address> c10.dll!<unknown symbol> [<unknown file> @ <unknown line number>]
00007FF8894F6C5A <unknown symbol address> c10.dll!<unknown symbol> [<unknown file> @ <unknown line number>]

[saudet]

You'll need to set the values of scale_factor for this to work

[mullerhai]

scale_factor
Hi, @saudet
I promise I have set the scale_factor field , options.scale_factor().put(DoubleVector(2))， does you not found it? , I think if we want to fixup the bug maybe could run my test code ,to find really bug or give me the really code content ,thanks

[saudet]

Do like #1594

[mullerhai]

Do like #1594

for DoubleVector put value maybe is complex than DoublePointer

I try to use options.scale_factor().put(DoubleVector(1).put(3)), then scale_factor is set ,but forward() meet error
also tell me error [java.lang.RuntimeException: invalid vector subscript]

Testing started at 17:24 ...
  align true  scala factor 3.0   mode org.bytedeco.pytorch.UpsampleMode[address=0x1cb5a739640,position=0,limit=0,capacity=0,deallocator=null]

java.lang.RuntimeException: invalid vector subscript

[mullerhai]

HI ,@saudet pure javacpp code, unsample really can not work

   val input = torch.arange(1, 5, dtype = torch.float32).view(1, 1, 2, 2)
    val options: UpsampleOptions = UpsampleOptions()
    options.scale_factor().put(DoubleVector(1).put(2))
    options.mode().put(new kNearest())
    val model = UpsampleImpl(options)
    val output = fromNative(model.forward(input.native))
    println(s"output ${output.shape}")

console log

input: tensor dtype=float32, shape=[1, 1, 2, 2], device=CPU 
[[[[1.0000, 2.0000],
   [3.0000, 4.0000]]]]
java.lang.RuntimeException: invalid vector subscript

	at org.bytedeco.pytorch.UpsampleImpl.forward(Native Method)
	at torch.nn.modules.UpsampleRawSuite.$init$$$anonfun$10(otherSuite.scala:187)

same python code

>>> input = torch.arange(1, 5, dtype=torch.float32).view(1, 1, 2, 2)
>>> input
tensor([[[[1., 2.],
          [3., 4.]]]])

>>> m = nn.Upsample(scale_factor=2, mode='nearest')
>>> m(input)

tensor([[[[1., 1., 2., 2.],
          [1., 1., 2., 2.],
          [3., 3., 4., 4.],
          [3., 3., 4., 4.]]]])

[saudet]

Please try to set the "org.bytedeco.javacpp.nopointergc" system property to "true".

[mullerhai]

Please try to set the "org.bytedeco.javacpp.nopointergc" system property to "true".

·```
class AdapativeMaxPool2dRawSuite3 extends munit.FunSuite {
test("AdapativeMaxPool2d output shapes") {

  System.setProperty("org.bytedeco.javacpp.nopointergc", "true")
  val nativeOutputSize = LongPointer(Array(5l, 7l) *)
  val options: AdaptiveMaxPool2dOptions = AdaptiveMaxPool2dOptions(nativeOutputSize)
  val nativeModule: AdaptiveMaxPool2dImpl = AdaptiveMaxPool2dImpl(options)
  println(s"AdaptiveMaxPool2d raw options 1: ${options.output_size().get} options 2: ${options.output_size().get} ")
  val pi = options.output_size()
  println(s"pi : ${pi}")
  val input = torch.randn(Seq(1, 64, 8, 9))
  val output = fromNative(nativeModule.forward(input.native))
  println(s"AdaptiveMaxPool2d nativeModule options 1: ${nativeModule.options.output_size().get} options 2: ${nativeModule.options.output_size()} ")
  val shape = output.shape
  println(s"output shape: ${output.shape}")
  assertEquals(shape, Seq(1, 64, 5, 7))
}

}

console log

Testing started at 19:43 ...
AdaptiveMaxPool2d raw options 1: 5 options 2: 5
pi : org.bytedeco.pytorch.LongOptional[address=0x2b2d88a8620,position=0,limit=0,capacity=0,deallocator=null]
AdaptiveMaxPool2d nativeModule options 1: 5 options 2: org.bytedeco.pytorch.LongOptional[address=0x2b2d92c1970,position=0,limit=0,capacity=0,deallocator=null]
output shape: ArraySeq(1, 64, 5, 0)

munit.ComparisonFailException: D:\data\storch_demo\src\main\scala\torch\pooltest.scala:422
421: println(s"output shape: ${output.shape}")
422: assertEquals(shape, Seq(1, 64, 5, 7))
423: // val m1 = new AdaptiveAvgPool2d((5, 7))
values are not the same
=> Obtained
ArraySeq(
1,
64,
5,
0
)
=> Diff (- obtained, + expected)
-ArraySeq(
+List(
1,
5,

• 0

• 7
  )
  Expected :List(1, 64, 5, 7)
  Actual :ArraySeq(1, 64, 5, 0)

[mullerhai]

·```
class UpsampleRawSuite extends munit.FunSuite {
test("UpsampleRawSuite output shapes") {
System.setProperty("org.bytedeco.javacpp.nopointergc", "true")
val input = torch.arange(1, 5, dtype = torch.float32).view(1, 1, 2, 2)
val input3 = torch.zeros(Seq(3, 3)).view(1, 1, 3, 3)
val options: UpsampleOptions = UpsampleOptions()
options.scale_factor().put(DoubleVector(1).put(2))
options.mode().put(new kNearest())
println(s"input: ${input}")
val model = UpsampleImpl(options)

val output = fromNative(model.forward(input.native.to(ScalarType.Float)))
println(s"output ${output.shape}")

console log

Testing started at 19:46 ...
input: tensor dtype=float32, shape=[1, 1, 2, 2], device=CPU
[[[[1.0000, 2.0000],
[3.0000, 4.0000]]]]

java.lang.RuntimeException: invalid vector subscript

at org.bytedeco.pytorch.UpsampleImpl.forward(Native Method)
at torch.nn.modules.UpsampleRawSuite.$init$$$anonfun$10(otherSuite.scala:190)

·```