Miscellaneous fix for several numpy issues (apache#16664)

* fix behavior of np.array when given official numpy ndarray

* bool for expand_dims and cast

* recover original Makefile

* address comments

* add boolean support for cumsum

* add gpu cast boolean support

* add error message

python/mxnet/numpy/multiarray.py

@@ -577,7 +577,7 @@ def __setitem__(self, key, value):
             if not isinstance(key, tuple) or len(key) != 0:
                 raise IndexError('scalar tensor can only accept `()` as index')
             if isinstance(value, numeric_types):
-                self.full(value)
+                self._full(value)
             elif isinstance(value, ndarray) and value.size == 1:
                 if value.shape != self.shape:
                     value = value.reshape(self.shape)
@@ -1993,15 +1993,20 @@ def array(object, dtype=None, ctx=None):
     """
     if ctx is None:
         ctx = current_context()
-    if isinstance(object, ndarray):
+    if isinstance(object, (ndarray, _np.ndarray)):
         dtype = object.dtype if dtype is None else dtype
+    elif isinstance(object, NDArray):
+        raise ValueError("If you're trying to create a mxnet.numpy.ndarray "
+                         "from mx.nd.NDArray, please use the zero-copy as_np_ndarray function.")
     else:
-        dtype = _np.float32 if dtype is None else dtype
-        if not isinstance(object, (ndarray, _np.ndarray)):
-            try:
-                object = _np.array(object, dtype=dtype)
-            except Exception as e:
-                raise TypeError('{}'.format(str(e)))
+        if dtype is None:
+            dtype = object.dtype if hasattr(object, "dtype") else _np.float32
+        try:
+            object = _np.array(object, dtype=dtype)
+        except Exception as e:
+            # printing out the error raised by official NumPy's array function
+            # for transparency on users' side
+            raise TypeError('{}'.format(str(e)))
     ret = empty(object.shape, dtype=dtype, ctx=ctx)
     if len(object.shape) == 0:
         ret[()] = object

src/ndarray/ndarray_function.cu

@@ -76,7 +76,7 @@ void Copy<gpu, gpu>(const TBlob &from, TBlob *to,
                       from.FlatTo1D<gpu, DType>(s),
                       s);
       } else {
-        MSHADOW_TYPE_SWITCH(from.type_flag_, SrcDType, {
+        MSHADOW_TYPE_SWITCH_WITH_BOOL(from.type_flag_, SrcDType, {
           to->FlatTo1D<gpu, DType>(s) =
             mshadow::expr::tcast<DType>(from.FlatTo1D<gpu, SrcDType>(s));
         })

src/operator/mxnet_op.h

@@ -671,7 +671,7 @@ template <typename xpu>
 MSHADOW_CINLINE void copy(mshadow::Stream<xpu> *s, const TBlob& to, const TBlob& from) {
   CHECK_EQ(from.Size(), to.Size());
   CHECK_EQ(from.dev_mask(), to.dev_mask());
-  MSHADOW_TYPE_SWITCH(to.type_flag_, DType, {
+  MSHADOW_TYPE_SWITCH_WITH_BOOL(to.type_flag_, DType, {
     if (to.type_flag_ == from.type_flag_) {
       mshadow::Copy(to.FlatTo1D<xpu, DType>(s), from.FlatTo1D<xpu, DType>(s), s);
     } else {

src/operator/numpy/np_cumsum-inl.h

@@ -98,7 +98,7 @@ void CumsumForwardImpl(const OpContext& ctx,
   }
 
   Stream<xpu> *s = ctx.get_stream<xpu>();
-  MSHADOW_TYPE_SWITCH(in.type_flag_, IType, {
+  MSHADOW_TYPE_SWITCH_WITH_BOOL(in.type_flag_, IType, {
     MSHADOW_TYPE_SWITCH(out.type_flag_, OType, {
       Kernel<cumsum_forward, xpu>::Launch(
         s, out.Size() / middle, out.dptr<OType>(),
@@ -157,7 +157,7 @@ void CumsumBackwardImpl(const OpContext& ctx,
     }
   }
   Stream<xpu> *s = ctx.get_stream<xpu>();
-  MSHADOW_TYPE_SWITCH(igrad.type_flag_, IType, {
+  MSHADOW_TYPE_SWITCH_WITH_BOOL(igrad.type_flag_, IType, {
     MSHADOW_TYPE_SWITCH(ograd.type_flag_, OType, {
       Kernel<cumsum_backward, xpu>::Launch(
         s, igrad.Size() / middle, igrad.dptr<IType>(),

src/operator/numpy/np_cumsum.cc

@@ -55,6 +55,9 @@ inline bool CumsumType(const nnvm::NodeAttrs& attrs,
   } else {
     TYPE_ASSIGN_CHECK(*out_attrs, 0, in_attrs->at(0));
     TYPE_ASSIGN_CHECK(*in_attrs, 0, out_attrs->at(0));
+    if (out_attrs->at(0) == mshadow::kBool) {
+      (*out_attrs)[0] = mshadow::kInt64;
+    }
   }
 
   return out_attrs->at(0) != -1 && in_attrs->at(0) != -1;

src/operator/numpy/np_init_op.h

@@ -205,7 +205,7 @@ void IdentityCompute(const nnvm::NodeAttrs& attrs,
   Stream<xpu> *s = ctx.get_stream<xpu>();
   const TBlob& out_data = outputs[0];
   int n = out_data.shape_[0];
-  MSHADOW_TYPE_SWITCH(out_data.type_flag_, DType, {
+  MSHADOW_TYPE_SWITCH_WITH_BOOL(out_data.type_flag_, DType, {
     MXNET_ASSIGN_REQ_SWITCH(req[0], req_type, {
       Kernel<identity<req_type>, xpu>::Launch(
           s, out_data.Size(), out_data.dptr<DType>(), n);

src/operator/tensor/elemwise_unary_op.h

@@ -451,7 +451,7 @@ void CastCompute(const nnvm::NodeAttrs& attrs,
   Stream<xpu> *s = ctx.get_stream<xpu>();
   MSHADOW_TYPE_SWITCH_WITH_BOOL(outputs[0].type_flag_, DstDType, {
     Tensor<xpu, 1, DstDType> out = outputs[0].FlatTo1D<xpu, DstDType>(s);
-    MSHADOW_TYPE_SWITCH(inputs[0].type_flag_, SrcDType, {
+    MSHADOW_TYPE_SWITCH_WITH_BOOL(inputs[0].type_flag_, SrcDType, {
       Tensor<xpu, 1, SrcDType> data = inputs[0].FlatTo1D<xpu, SrcDType>(s);
       if (outputs[0].type_flag_ != inputs[0].type_flag_ ||
           req[0] != kWriteInplace) {

tests/python/unittest/test_numpy_interoperability.py

@@ -93,6 +93,7 @@ def _add_workload_copy():
 
 def _add_workload_expand_dims():
     OpArgMngr.add_workload('expand_dims', np.random.uniform(size=(4, 1)), -1)
+    OpArgMngr.add_workload('expand_dims', np.random.uniform(size=(4, 1)) > 0.5, -1)
     for axis in range(-5, 4):
         OpArgMngr.add_workload('expand_dims', np.empty((2, 3, 4, 5)), axis)
 
@@ -852,8 +853,8 @@ def _signs(dt):
     # test_float_remainder_corner_cases
     # Check remainder magnitude.
     for ct in _FLOAT_DTYPES:
-        b = _np.array(1.0)
-        a = np.array(_np.nextafter(_np.array(0.0), -b), dtype=ct)
+        b = _np.array(1.0, dtype=ct)
+        a = np.array(_np.nextafter(_np.array(0.0, dtype=ct), -b), dtype=ct)
         b = np.array(b, dtype=ct)
         OpArgMngr.add_workload('remainder', a, b)
         OpArgMngr.add_workload('remainder', -a, -b)

tests/python/unittest/test_numpy_ndarray.py

@@ -18,6 +18,7 @@
 # pylint: skip-file
 from __future__ import absolute_import
 from __future__ import division
+import itertools
 import os
 import unittest
 import numpy as _np
@@ -87,17 +88,20 @@ def test_np_array_creation():
         [],
         (),
         [[1, 2], [3, 4]],
+        _np.random.randint(-10, 10, size=rand_shape_nd(3)),
         _np.random.uniform(size=rand_shape_nd(3)),
         _np.random.uniform(size=(3, 0, 4))
     ]
     for dtype in dtypes:
         for src in objects:
             mx_arr = np.array(src, dtype=dtype)
             assert mx_arr.ctx == mx.current_context()
+            if dtype is None:
+                dtype = src.dtype if isinstance(src, _np.ndarray) else _np.float32
             if isinstance(src, mx.nd.NDArray):
-                np_arr = _np.array(src.asnumpy(), dtype=dtype if dtype is not None else _np.float32)
+                np_arr = _np.array(src.asnumpy(), dtype=dtype)
             else:
-                np_arr = _np.array(src, dtype=dtype if dtype is not None else _np.float32)
+                np_arr = _np.array(src, dtype=dtype)
             assert mx_arr.dtype == np_arr.dtype
             assert same(mx_arr.asnumpy(), np_arr)
 
@@ -471,9 +475,6 @@ def test_np_grad_ndarray_type():
 @with_seed()
 @use_np
 def test_np_ndarray_astype():
-    mx_data = np.array([2, 3, 4, 5], dtype=_np.int32)
-    np_data = mx_data.asnumpy()
-
     class TestAstype(HybridBlock):
         def __init__(self, dtype, copy):
             super(TestAstype, self).__init__()
@@ -483,24 +484,29 @@ def __init__(self, dtype, copy):
         def hybrid_forward(self, F, x):
             return x.astype(dtype=self._dtype, copy=self._copy)
 
-    def check_astype_equal(dtype, copy, expect_zero_copy=False, hybridize=False):
-        test_astype = TestAstype(dtype, copy)
+    def check_astype_equal(itype, otype, copy, expect_zero_copy=False, hybridize=False):
+        expect_zero_copy = copy is False and itype == otype
+        mx_data = np.array([2, 3, 4, 5], dtype=itype)
+        np_data = mx_data.asnumpy()
+        test_astype = TestAstype(otype, copy)
         if hybridize:
             test_astype.hybridize()
         mx_ret = test_astype(mx_data)
         assert type(mx_ret) is np.ndarray
-        np_ret = np_data.astype(dtype=dtype, copy=copy)
+        np_ret = np_data.astype(dtype=otype, copy=copy)
         assert mx_ret.dtype == np_ret.dtype
         assert same(mx_ret.asnumpy(), np_ret)
         if expect_zero_copy and not hybridize:
             assert id(mx_ret) == id(mx_data)
             assert id(np_ret) == id(np_data)
 
-    for dtype in [np.int8, np.uint8, np.int32, np.float16, np.float32, np.float64, np.bool, np.bool_,
-                  'int8', 'uint8', 'int32', 'float16', 'float32', 'float64', 'bool']:
+    dtypes = [np.int8, np.uint8, np.int32, np.float16, np.float32, np.float64, np.bool, np.bool_,
+              'int8', 'uint8', 'int32', 'float16', 'float32', 'float64', 'bool']
+
+    for itype, otype in itertools.product(dtypes, dtypes):
         for copy in [True, False]:
             for hybridize in [True, False]:
-                check_astype_equal(dtype, copy, copy is False and mx_data.dtype == dtype, hybridize)
+                check_astype_equal(itype, otype, copy, hybridize)
 
 
 @with_seed()

tests/python/unittest/test_numpy_op.py

@@ -442,7 +442,7 @@ def is_int(dtype):
             for axis in ([i for i in range(in_data_dim)] + [(), None]):
                 for itype in ['float16', 'float32', 'float64', 'int8', 'int32', 'int64', 'bool']:
                     for dtype in ['float16', 'float32', 'float64', 'int8', 'int32', 'int64']:
-                        if (is_int(dtype) and not is_int(itype))\
+                        if (is_int(dtype) and not is_int(itype)) or (is_windows and is_int(itype))\
                                 or (itype == 'bool' and\
                                     (dtype not in ('float32', 'float64', 'int32', 'int64') or is_windows)):
                             continue
@@ -2390,6 +2390,16 @@ def hybrid_forward(self, F, a):
                         np_out = _np.cumsum(x.asnumpy(), axis=axis, dtype=otype)
                         assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
 
+    for shape in shapes:
+        for axis in [None] + [i for i in range(0, len(shape))]:
+            for otype in [None, _np.int32, _np.int64]:
+                for itype in [_np.bool, _np.int8, _np.int32, _np.int64]:
+                    x = rand_ndarray(shape).astype(itype).as_np_ndarray()
+                    np_out = _np.cumsum(x.asnumpy(), axis=axis, dtype=otype)
+                    mx_out = np.cumsum(x, axis=axis, dtype=otype)
+                    assert mx_out.shape == np_out.shape
+                    assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+
 
 @with_seed()
 @use_np