Add reshape_infer procedure (#646)

Unlike numpy, `reshape` does not support having dimensions with value -1 to infer their value. To do so a new `reshape_infer` is added.

This is added as a separate procedure to avoid the (small) cost this adds on top of the usual reshape (which could be called relatively frequently).

Author: AngelEzquerra

src/arraymancer/tensor/private/p_shapeshifting.nim

@@ -18,6 +18,8 @@ import  ../../laser/tensor/initialization,
         ./p_checks,
         nimblas
 
+import std / sequtils
+
 proc contiguousImpl*[T](t: Tensor[T], layout: OrderType, result: var Tensor[T]) =
   if layout == rowMajor:
     result = t.map_inline(x)
@@ -28,16 +30,39 @@ proc contiguousImpl*[T](t: Tensor[T], layout: OrderType, result: var Tensor[T])
     apply2_inline(result, t):
       y
 
-proc reshape_with_copy*[T](t: Tensor[T], new_shape: varargs[int]|Metadata, result: var Tensor[T]) =
+proc reshape_with_copy*[T](t: Tensor[T], new_shape: varargs[int]|Metadata|seq[int], result: var Tensor[T]) =
   result = newTensorUninit[T](new_shape)
   result.apply2_inline(t,y)
 
-proc reshape_no_copy*(t: AnyTensor, new_shape: varargs[int]|Metadata, result: var AnyTensor, layout: OrderType) {.noSideEffect.}=
+proc reshape_no_copy*(t: AnyTensor, new_shape: varargs[int]|Metadata|seq[int], result: var AnyTensor, layout: OrderType) {.noSideEffect.}=
   result.shape.copyFrom(new_shape)
   shape_to_strides(result.shape, layout, result.strides)
   result.offset = t.offset
 
-proc reshapeImpl*(t: AnyTensor, new_shape: varargs[int]|Metadata, result: var AnyTensor) =
+proc infer_shape*(t: Tensor, new_shape: varargs[int]): seq[int] {.noinit.} =
+  ## Replace the single -1 value on `new_shape` with the value that
+  ## makes the size the same as that of the input tensor
+  result = new_shape.toSeq
+  var auto_axis = -1
+  var auto_axis_count = 0
+  for n in 0 .. result.high:
+    if result[n] == -1:
+      auto_axis_count += 1
+      auto_axis = n
+      break
+  if auto_axis_count > 1:
+    raise newException(ValueError, "Only one dimension can be inferred by inferShape")
+  elif auto_axis_count == 0:
+    when compileOption("boundChecks"):
+      raise newException(ValueError, "At least one dimension must be inferred by inferShape")
+  else:
+    result[auto_axis] = t.size div result.filterIt(it != -1).prod
+
+proc reshapeImpl*(t: AnyTensor, new_shape: varargs[int]|Metadata|seq[int],
+    result: var AnyTensor, infer: static bool) =
+  when infer:
+    let new_shape = t.infer_shape(new_shape)
+
   when compileOption("boundChecks"):
     check_reshape(t, new_shape)
 

src/arraymancer/tensor/shapeshifting.nim

@@ -63,10 +63,12 @@ proc reshape*(t: Tensor, new_shape: varargs[int]): Tensor {.noinit.} =
   ##
   ## Input:
   ##   - a tensor
-  ##   - a new shape. Number of elements must be the same
+  ##   - a new shape. Number of elements must be the same. Unlike numpy,
+  ##     dimensions cannot be -1 to infer their value. If that is what you need
+  ##     you must use the alternative `reshape_infer` proc.
   ## Returns:
   ##   - a tensor with the same data but reshaped.
-  reshapeImpl(t, new_shape, result)
+  reshapeImpl(t, new_shape, result, infer = false)
 
 proc reshape*(t: Tensor, new_shape: Metadata): Tensor {.noinit.} =
   ## Reshape a tensor. If possible no data copy is done and the returned tensor
@@ -78,7 +80,21 @@ proc reshape*(t: Tensor, new_shape: Metadata): Tensor {.noinit.} =
   ##   - a new shape. Number of elements must be the same
   ## Returns:
   ##   - a tensor with the same data but reshaped.
-  reshapeImpl(t, new_shape, result)
+  reshapeImpl(t, new_shape, result, infer = false)
+
+proc reshape_infer*(t: Tensor, new_shape: varargs[int]):
+    Tensor {.noinit.} =
+  ## Reshape a tensor. If possible no data copy is done and the returned tensor
+  ## shares data with the input. If input is not contiguous, this is not possible
+  ## and a copy will be made.
+  ##
+  ## Input:
+  ##   - a tensor
+  ##   - a new shape. Number of elements must be the same. The new shape can
+  ##     contain -1 to infer the size of one (and only one) dimension
+  ## Returns:
+  ##   - a tensor with the same data but reshaped.
+  reshapeImpl(t, new_shape, result, infer = true)
 
 proc flatten*(t: Tensor): Tensor {.noinit,inline.} =
   ## Flatten a tensor, returning a rank-1 tensor with the same data as the input.

tests/tensor/test_shapeshifting.nim

@@ -62,9 +62,13 @@ proc main() =
         check: a == b
 
     test "Reshape":
-      let a = toSeq(1..4).toTensor().reshape(2,2)
+      let a = toSeq(1..4).toTensor().reshape(2, 2)
+      let b = toSeq(1..4).toTensor().reshape_infer(-1, 2)
+      let c = toSeq(1..4).toTensor().reshape_infer(2, -1)
       check: a == [[1,2],
                   [3,4]].toTensor()
+      check: a == b
+      check: a == c
 
     test "Unsafe reshape":
       block: