numpy operator ravel, derive from reshape (#16016)

* it is the same as reshape(x, -1)

* register reshape with prefix _npi_

* fix format error

* edit examples in doc

* fix error in review

* add out in wrapper

* remove out

* test data type and add order

* change order check

* remove redundant test and add docstring about order

python/mxnet/ndarray/numpy/_op.py

@@ -33,7 +33,8 @@
            'rint', 'radians', 'reciprocal', 'square', 'negative', 'fix', 'ceil', 'floor',
            'trunc', 'logical_not', 'arcsinh', 'arccosh', 'arctanh', 'tensordot',
            'linspace', 'expand_dims', 'tile', 'arange', 'split', 'concatenate', 'stack', 'mean',
-           'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign']
+           'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign',
+           'ravel']
 
 
 @set_module('mxnet.ndarray.numpy')
@@ -2483,3 +2484,56 @@ def copysign(x1, x2, out=None):
     array([-1.,  0.,  1.])
     """
     return _ufunc_helper(x1, x2, _npi.copysign, _np.copysign, _npi.copysign_scalar, _npi.rcopysign_scalar, out)
+
+
+@set_module('mxnet.ndarray.numpy')
+def ravel(x, order='C'):
+    r"""
+    ravel(x)
+
+    Return a contiguous flattened array.
+    A 1-D array, containing the elements of the input, is returned.  A copy is
+    made only if needed.
+
+    Parameters
+    ----------
+    x : ndarray
+        Input array.  The elements in `x` are read in row-major, C-style order and
+        packed as a 1-D array.
+    order : `C`, optional
+        Only support row-major, C-style order.
+
+    Returns
+    -------
+    y : ndarray
+        y is an array of the same subtype as `x`, with shape ``(x.size,)``.
+        Note that matrices are special cased for backward compatibility, if `x`
+        is a matrix, then y is a 1-D ndarray.
+
+    Notes
+    -----
+    This function differs from the original numpy.arange in the following aspects:
+        - Only support row-major, C-style order.
+
+    Examples
+    --------
+    It is equivalent to ``reshape(x, -1)``.
+
+    >>> x = np.array([[1, 2, 3], [4, 5, 6]])
+    >>> print(np.ravel(x))
+    [1. 2. 3. 4. 5. 6.]
+
+    >>> print(x.reshape(-1))
+    [1. 2. 3. 4. 5. 6.]
+
+    >>> print(np.ravel(x.T))
+    [1. 4. 2. 5. 3. 6.]
+    """
+    if order != 'C':
+        raise NotImplementedError('order {} is not supported'.format(order))
+    if isinstance(x, numeric_types):
+        return _np.reshape(x, -1)
+    elif isinstance(x, NDArray):
+        return _npi.reshape(x, -1)
+    else:
+        raise TypeError('type {} not supported'.format(str(type(x))))

python/mxnet/numpy/multiarray.py

@@ -52,7 +52,8 @@
            'degrees', 'log2', 'log1p', 'rint', 'radians', 'reciprocal', 'square', 'negative',
            'fix', 'ceil', 'floor', 'trunc', 'logical_not', 'arcsinh', 'arccosh', 'arctanh',
            'tensordot', 'linspace', 'expand_dims', 'tile', 'arange', 'split', 'concatenate',
-           'stack', 'mean', 'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign']
+           'stack', 'mean', 'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign',
+           'ravel']
 
 # Return code for dispatching indexing function call
 _NDARRAY_UNSUPPORTED_INDEXING = -1
@@ -3986,3 +3987,49 @@ def copysign(x1, x2, out=None):
     array([-1.,  0.,  1.])
     """
     return _mx_nd_np.copysign(x1, x2, out=out)
+
+
+@set_module('mxnet.numpy')
+def ravel(x, order='C'):
+    r"""
+    ravel(x)
+
+    Return a contiguous flattened array.
+    A 1-D array, containing the elements of the input, is returned.  A copy is
+    made only if needed.
+
+    Parameters
+    ----------
+    x : ndarray
+        Input array.  The elements in `x` are read in row-major, C-style order and
+        packed as a 1-D array.
+    order : `C`, optional
+        Only support row-major, C-style order.
+
+    Returns
+    -------
+    y : ndarray
+        y is an array of the same subtype as `x`, with shape ``(x.size,)``.
+        Note that matrices are special cased for backward compatibility, if `x`
+        is a matrix, then y is a 1-D ndarray.
+
+    Notes
+    -----
+    This function differs from the original numpy.arange in the following aspects:
+        - Only support row-major, C-style order.
+
+    Examples
+    --------
+    It is equivalent to ``reshape(x, -1)``.
+
+    >>> x = np.array([[1, 2, 3], [4, 5, 6]])
+    >>> print(np.ravel(x))
+    [1. 2. 3. 4. 5. 6.]
+
+    >>> print(x.reshape(-1))
+    [1. 2. 3. 4. 5. 6.]
+
+    >>> print(np.ravel(x.T))
+    [1. 4. 2. 5. 3. 6.]
+    """
+    return _mx_nd_np.ravel(x, order)

python/mxnet/symbol/numpy/_symbol.py

@@ -35,7 +35,8 @@
            'rint', 'radians', 'reciprocal', 'square', 'negative', 'fix', 'ceil', 'floor',
            'trunc', 'logical_not', 'arcsinh', 'arccosh', 'arctanh', 'tensordot',
            'linspace', 'expand_dims', 'tile', 'arange', 'split', 'concatenate', 'stack', 'mean',
-           'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign']
+           'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign',
+           'ravel']
 
 
 def _num_outputs(sym):
@@ -2778,4 +2779,43 @@ def copysign(x1, x2, out=None):
     return _ufunc_helper(x1, x2, _npi.copysign, _np.copysign, _npi.copysign_scalar, _npi.rcopysign_scalar, out)
 
 
+@set_module('mxnet.symbol.numpy')
+def ravel(x, order='C'):
+    r"""
+    ravel(x)
+
+    Return a contiguous flattened array.
+    A 1-D array, containing the elements of the input, is returned.  A copy is
+    made only if needed.
+
+    Parameters
+    ----------
+    x : ndarray
+        Input array.  The elements in `x` are read in row-major, C-style order and
+        packed as a 1-D array.
+    order : `C`, optional
+        Only support row-major, C-style order.
+
+    Returns
+    -------
+    y : ndarray
+        y is an array of the same subtype as `x`, with shape ``(x.size,)``.
+        Note that matrices are special cased for backward compatibility, if `x`
+        is a matrix, then y is a 1-D ndarray.
+
+    Notes
+    -----
+    This function differs from the original numpy.arange in the following aspects:
+        - Only support row-major, C-style order.
+    """
+    if order != 'C':
+        raise NotImplementedError('order {} is not supported'.format(order))
+    if isinstance(x, numeric_types):
+        return _np.reshape(x, -1)
+    elif isinstance(x, _Symbol):
+        return _npi.reshape(x, -1)
+    else:
+        raise TypeError('type {} not supported'.format(str(type(x))))
+
+
 _set_np_symbol_class(_Symbol)

src/operator/numpy/np_matrix_op.cc

@@ -163,6 +163,7 @@ bool NumpyReshapeShape(const nnvm::NodeAttrs& attrs,
 
 NNVM_REGISTER_OP(_np_reshape)
 .describe(R"code()code" ADD_FILELINE)
+.add_alias("_npi_reshape")
 .set_num_inputs(1)
 .set_num_outputs(1)
 .set_attr_parser(ParamParser<NumpyReshapeParam>)

tests/python/unittest/test_numpy_op.py

@@ -1320,6 +1320,39 @@ def hybrid_forward(self, F, a, *args):
                 assert same(mx_out.asnumpy(), np_out)
 
 
+@with_seed()
+@use_np
+def test_np_ravel():
+    class TestRavel(HybridBlock):
+        def __init__(self):
+            super(TestRavel, self).__init__()
+
+        def hybrid_forward(self, F, a):
+            return F.np.ravel(a)
+
+    types = ['float64', 'float32', 'float16', 'int64', 'int32', 'int8']
+    for oneType in types:
+        for hybridize in [True, False]:
+            for shape in [(), (2,), (2, 2), (1, 2, 3), (3, 0), (1, 0, 2)]:
+                test_ravel = TestRavel()
+                if hybridize:
+                    test_ravel.hybridize()
+                x = rand_ndarray(shape, dtype=oneType).as_np_ndarray()
+                x.attach_grad()
+                np_out = _np.ravel(x.asnumpy())
+                with mx.autograd.record():
+                    mx_out = test_ravel(x)
+                assert mx_out.shape == np_out.shape
+                assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+                mx_out.backward()
+                np_backward = _np.ones(shape)
+                assert_almost_equal(x.grad.asnumpy(), np_backward, rtol=1e-3, atol=1e-5)
+
+                mx_out = np.ravel(x)
+                np_out = _np.ravel(x.asnumpy())
+                assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+
+
 @with_seed()
 @use_np
 def test_np_randint():