np compatible vstack (apache#15850)

python/mxnet/ndarray/numpy/_op.py

@@ -32,7 +32,7 @@
            'exp', 'expm1', 'arcsin', 'arccos', 'arctan', 'sign', 'log', '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',
+           'linspace', 'expand_dims', 'tile', 'arange', 'split', 'concatenate', 'stack', 'vstack', 'mean',
            'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign',
            'ravel']
 
@@ -1990,6 +1990,57 @@ def get_list(arrays):
     return _npi.stack(*arrays, axis=axis, out=out)
 
 
+@set_module('mxnet.ndarray.numpy')
+def vstack(arrays, out=None):
+    r"""Stack arrays in sequence vertically (row wise).
+
+    This is equivalent to concatenation along the first axis after 1-D arrays
+    of shape `(N,)` have been reshaped to `(1,N)`. Rebuilds arrays divided by
+    `vsplit`.
+
+    This function makes most sense for arrays with up to 3 dimensions. For
+    instance, for pixel-data with a height (first axis), width (second axis),
+    and r/g/b channels (third axis). The functions `concatenate` and `stack`
+    provide more general stacking and concatenation operations.
+
+    Parameters
+    ----------
+    tup : sequence of ndarrays
+        The arrays must have the same shape along all but the first axis.
+        1-D arrays must have the same length.
+
+    Returns
+    -------
+    stacked : ndarray
+        The array formed by stacking the given arrays, will be at least 2-D.
+
+    Examples
+    --------
+    >>> a = np.array([1, 2, 3])
+    >>> b = np.array([2, 3, 4])
+    >>> np.vstack((a, b))
+    array([[1., 2., 3.],
+            [2., 3., 4.]])
+
+    >>> a = np.array([[1], [2], [3]])
+    >>> b = np.array([[2], [3], [4]])
+    >>> np.vstack((a, b))
+    array([[1.],
+            [2.],
+            [3.],
+            [2.],
+            [3.],
+            [4.]])
+    """
+    def get_list(arrays):
+        if not hasattr(arrays, '__getitem__') and hasattr(arrays, '__iter__'):
+            raise ValueError("expected iterable for arrays but got {}".format(type(arrays)))
+        return [arr for arr in arrays]
+
+    arrays = get_list(arrays)
+    return _npi.vstack(*arrays)
+
+
 @set_module('mxnet.ndarray.numpy')
 def maximum(x1, x2, out=None):
     """Returns element-wise maximum of the input arrays with broadcasting.

python/mxnet/numpy/multiarray.py

@@ -52,8 +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',
-           'ravel']
+           'stack', 'vstack', 'mean', 'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices',
+           'copysign', 'ravel']
 
 # Return code for dispatching indexing function call
 _NDARRAY_UNSUPPORTED_INDEXING = -1
@@ -3560,6 +3560,51 @@ def stack(arrays, axis=0, out=None):
     return _mx_nd_np.stack(arrays, axis=axis, out=out)
 
 
+@set_module('mxnet.numpy')
+def vstack(arrays, out=None):
+    r"""Stack arrays in sequence vertically (row wise).
+
+    This is equivalent to concatenation along the first axis after 1-D arrays
+    of shape `(N,)` have been reshaped to `(1,N)`. Rebuilds arrays divided by
+    `vsplit`.
+
+    This function makes most sense for arrays with up to 3 dimensions. For
+    instance, for pixel-data with a height (first axis), width (second axis),
+    and r/g/b channels (third axis). The functions `concatenate` and `stack`
+    provide more general stacking and concatenation operations.
+
+    Parameters
+    ----------
+    tup : sequence of ndarrays
+        The arrays must have the same shape along all but the first axis.
+        1-D arrays must have the same length.
+
+    Returns
+    -------
+    stacked : ndarray
+        The array formed by stacking the given arrays, will be at least 2-D.
+
+    Examples
+    --------
+    >>> a = np.array([1, 2, 3])
+    >>> b = np.array([2, 3, 4])
+    >>> np.vstack((a, b))
+    array([[1., 2., 3.],
+           [2., 3., 4.]])
+
+    >>> a = np.array([[1], [2], [3]])
+    >>> b = np.array([[2], [3], [4]])
+    >>> np.vstack((a, b))
+    array([[1.],
+           [2.],
+           [3.],
+           [2.],
+           [3.],
+           [4.]])
+    """
+    return _mx_nd_np.vstack(arrays)
+
+
 @set_module('mxnet.numpy')
 def maximum(x1, x2, out=None):
     """Returns element-wise maximum of the input arrays with broadcasting.

python/mxnet/symbol/numpy/_symbol.py

@@ -34,7 +34,7 @@
            'expm1', 'arcsin', 'arccos', 'arctan', 'sign', 'log', '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',
+           'linspace', 'expand_dims', 'tile', 'arange', 'split', 'concatenate', 'stack', 'vstack', 'mean',
            'maximum', 'minimum', 'swapaxes', 'clip', 'argmax', 'std', 'var', 'indices', 'copysign',
            'ravel']
 
@@ -2396,6 +2396,39 @@ def get_list(arrays):
     return _npi.stack(*arrays, axis=axis, out=out)
 
 
+@set_module('mxnet.symbol.numpy')
+def vstack(arrays, out=None):
+    r"""Stack arrays in sequence vertically (row wise).
+
+    This is equivalent to concatenation along the first axis after 1-D arrays
+    of shape `(N,)` have been reshaped to `(1,N)`. Rebuilds arrays divided by
+    `vsplit`.
+
+    This function makes most sense for arrays with up to 3 dimensions. For
+    instance, for pixel-data with a height (first axis), width (second axis),
+    and r/g/b channels (third axis). The functions `concatenate` and `stack`
+    provide more general stacking and concatenation operations.
+
+    Parameters
+    ----------
+    tup : sequence of _Symbol
+        The arrays must have the same shape along all but the first axis.
+        1-D arrays must have the same length.
+
+    Returns
+    -------
+    stacked : _Symbol
+        The array formed by stacking the given arrays, will be at least 2-D.
+    """
+    def get_list(arrays):
+        if not hasattr(arrays, '__getitem__') and hasattr(arrays, '__iter__'):
+            raise ValueError("expected iterable for arrays but got {}".format(type(arrays)))
+        return [arr for arr in arrays]
+
+    arrays = get_list(arrays)
+    return _npi.vstack(*arrays)
+
+
 @set_module('mxnet.symbol.numpy')
 def maximum(x1, x2, out=None):
     return _ufunc_helper(x1, x2, _npi.maximum, _np.maximum, _npi.maximum_scalar, None, out)

src/operator/numpy/np_matrix_op-inl.h

@@ -41,6 +41,14 @@ struct NumpyTransposeParam : public dmlc::Parameter<NumpyTransposeParam> {
   }
 };
 
+struct NumpyVstackParam : public dmlc::Parameter<NumpyVstackParam> {
+  int num_args;
+  DMLC_DECLARE_PARAMETER(NumpyVstackParam) {
+    DMLC_DECLARE_FIELD(num_args).set_lower_bound(1)
+    .describe("Number of inputs to be vstacked.");
+  }
+};
+
 template<typename xpu>
 void NumpyTranspose(const nnvm::NodeAttrs& attrs,
                     const OpContext& ctx,
@@ -60,6 +68,78 @@ void NumpyTranspose(const nnvm::NodeAttrs& attrs,
   }
 }
 
+template<typename xpu>
+void NumpyVstackForward(const nnvm::NodeAttrs& attrs,
+                        const OpContext& ctx,
+                        const std::vector<TBlob>& inputs,
+                        const std::vector<OpReqType>& req,
+                        const std::vector<TBlob>& outputs) {
+  using namespace mshadow;
+  using namespace mshadow_op;
+
+  const NumpyVstackParam& param = nnvm::get<NumpyVstackParam>(attrs.parsed);
+  CHECK_EQ(inputs.size(), param.num_args);
+  CHECK_EQ(outputs.size(), 1);
+  CHECK_EQ(req.size(), 1);
+
+  // reshape if necessary
+  std::vector<TBlob> data(param.num_args);
+  for (int i = 0; i < param.num_args; i++) {
+    if (inputs[i].shape_.ndim() == 0 || inputs[i].shape_.ndim() == 1) {
+      TShape shape = Shape2(1, inputs[i].shape_.Size());
+      data[i] = inputs[i].reshape(shape);
+    } else {
+      data[i] = inputs[i];
+    }
+  }
+
+  // initialize ConcatOp
+  ConcatParam cparam;
+  cparam.num_args = param.num_args;
+  cparam.dim = 0;
+  MSHADOW_TYPE_SWITCH(inputs[0].type_flag_, DType, {
+    ConcatOp<xpu, DType> op;
+    op.Init(cparam);
+    op.Forward(ctx, data, req, outputs);
+  });
+}
+
+template<typename xpu>
+void NumpyVstackBackward(const nnvm::NodeAttrs& attrs,
+                         const OpContext& ctx,
+                         const std::vector<TBlob>& inputs,
+                         const std::vector<OpReqType>& req,
+                         const std::vector<TBlob>& outputs) {
+  using namespace mshadow;
+  using namespace mshadow_op;
+
+  const NumpyVstackParam& param = nnvm::get<NumpyVstackParam>(attrs.parsed);
+  CHECK_EQ(inputs.size(), 1);
+  CHECK_EQ(outputs.size(), param.num_args);
+  CHECK_EQ(req.size(), param.num_args);
+
+  // reshape if necessary
+  std::vector<TBlob> data(param.num_args);
+  for (int i = 0; i < param.num_args; i++) {
+    if (outputs[i].shape_.ndim() == 0 || outputs[i].shape_.ndim() == 1) {
+      TShape shape = Shape2(1, outputs[i].shape_.Size());
+      data[i] = outputs[i].reshape(shape);
+    } else {
+      data[i] = outputs[i];
+    }
+  }
+
+  // initialize ConcatOp
+  ConcatParam cparam;
+  cparam.num_args = param.num_args;
+  cparam.dim = 0;
+  MSHADOW_TYPE_SWITCH(inputs[0].type_flag_, DType, {
+    ConcatOp<xpu, DType> op;
+    op.Init(cparam);
+    op.Backward(ctx, inputs[0], req, data);
+  });
+}
+
 }  // namespace op
 }  // namespace mxnet