def pad_op_transform(graph: Graph, match: dict):
op = match['op']
pad_op = match['pad_op']
input_data = pad_op.in_node(0)
pads = pad_op.in_node(1).value if len(
pad_op.in_nodes()) == 2 else pad_op.pads
if pad_op.mode != 'constant':
log.info(
'The pad node "{}" with pad mode "{}" cannot be fused.'.format(
pad_op.soft_get('name'), pad_op.mode))
return
if pad_op.mode == 'constant' and pad_op.fill_value != 0.0:
log.info('The pad node "{}" with non-zero fill value cannot be fused.'.
format(pad_op.soft_get('name')))
return
input_tensor_dims = len(match['pad_output'].shape)
if np.any(pads[get_features_dim(op.graph.graph['layout'], input_tensor_dims)] != 0) or \
np.any(pads[get_batch_dim(op.graph.graph['layout'], input_tensor_dims)] != 0):
log.info(
'The pad node "{}" with padding over feature/batch dimension cannot be fused.'
.format(pad_op.soft_get('name')))
return
op.pad += pads
op.pad_spatial_shape = op.pad[op.spatial_dims]
op['auto_pad'] = None
if op.type == 'Pooling':
op['exclude_pad'] = False
assert (graph[match['pad_output'].node][match['op'].node][0]['in'] == 0)
edge_attrs = graph.get_edge_data(match['pad_output'].id, match['op'].id)[0]
graph.remove_edge(match['pad_output'].id, match['op'].id)
graph.add_edge(input_data.id, match['op'].id, **{'in': 0, **edge_attrs})
def infer(node: Node):
in_shape = node.in_node().shape
if in_shape.size != 4:
raise Error(
'TensorFlow SpaceToDepth operation is supported for 4D \'NHWC\' input layout only. '
'Current input shape is \'{}\''.format(in_shape))
layout = node.graph.graph['layout']
N = in_shape[get_batch_dim(layout, 4)]
H = in_shape[get_height_dim(layout, 4)]
W = in_shape[get_width_dim(layout, 4)]
C = in_shape[get_features_dim(layout, 4)]
block_size = node['block_size']
if H % block_size or W % block_size:
raise Error(
'Spatial dimensions of input tensor of SpaceToDepth operation have to be divisible by '
'SpaceToDepth \'block_size\' parameter. Input tensor shape = {}. Spatial dimensions = {},{}. '
'block_size = {}'.format(in_shape, H, W, block_size))
out_shape = shape_for_layout(layout,
batch=N,
features=int(C * (block_size**2)),
height=int(H / block_size),
width=int(W / block_size))
assert np.prod(in_shape) == np.prod(out_shape)
node.out_node().shape = int64_array(out_shape)
def infer(node: Node):
layout = node.graph.graph['layout']
assert len(layout) == 4
assert len(
[p for p in node.in_ports().values() if not p.disconnected()])
assert node.has_valid('mode')
assert node.has_valid('axes')
src_shape = node.in_port(0).data.get_shape()
assert src_shape is not None
dst_shape = node.in_port(1).data.get_value()
assert dst_shape is not None
out_height = dst_shape[0]
out_width = dst_shape[1]
node.out_node().shape = shape_for_layout(
layout,
batch=src_shape[get_batch_dim(layout, 4)],
features=src_shape[get_features_dim(layout, 4)],
height=out_height,
width=out_width)
PermuteAttrs.create_permute_attrs(node, attrs=[('axes', 'input:0')])
def infer(node: Node):
in_shape = node.in_node().shape
if in_shape.size != 4:
raise Error(
'TensorFlow DepthToSpace operation is supported for 4D \'NHWC\' input layout only. '
'Current input shape is \'{}\''.format(in_shape))
layout = node.graph.graph['layout']
N = in_shape[get_batch_dim(layout, 4)]
H = in_shape[get_height_dim(layout, 4)]
W = in_shape[get_width_dim(layout, 4)]
C = in_shape[get_features_dim(layout, 4)]
block_size = node['block_size']
if C % (block_size**2):
raise Error(
'Feature dimensions of input tensor of DepthToSpace operation have to be divisible by square '
'of DepthToSpace \'block_size\' parameter. Input tensor shape = {}. Feature dimension = {}. '
'block_size = {}'.format(in_shape, C, block_size))
out_shape = shape_for_layout(layout,
batch=N,
features=int(C / (block_size**2)),
height=int(H * block_size),
width=int(W * block_size))
assert np.prod(in_shape) == np.prod(out_shape)
node.out_node().shape = int64_array(out_shape)
def upsample_infer(node: Node):
layout = node.graph.graph['layout']
assert len(layout) == 4
input_shape = node.in_node(0).shape
if input_shape is None:
return
if len(node.in_nodes()) == 1:
in_height = input_shape[get_height_dim(layout, 4)]
in_width = input_shape[get_width_dim(layout, 4)]
assert node.has('width_scale') is not None and node.has(
'height_scale') is not None
out_height = math.floor(in_height * node.height_scale)
out_width = math.floor(in_width * node.width_scale)
node.out_node().shape = shape_for_layout(
layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=input_shape[get_features_dim(layout, 4)],
height=out_height,
width=out_width)
else:
assert node.in_node(1).value is not None
eps = 1e-5 # This is to make rounding in case of very close number to round to closest instead of down
# generic output shape calculation to support 5D input shape case
node.out_node().shape = np.array(
(input_shape + eps) * node.in_node(1).value).astype(np.int64)
def infer(node: Node):
assert [port.idx for port in node.in_ports().values() if not port.disconnected()] == [0], \
'Wrong input nodes number for node {} with type ExtractImagePatches'.format(node.soft_get('name', node.id))
input_shape = node.in_port(0).data.get_shape()
name = node.soft_get('name', node.id)
assert input_shape is not None, 'Input shape is not set for node {} with type ExtractImagePatches'.format(name)
assert len(input_shape) == 4, 'ExtractImagePatches operation supports only 4D tensors'
layout = node.graph.graph['layout']
N = input_shape[get_batch_dim(layout, 4)]
C = input_shape[get_features_dim(layout, 4)]
size_spatial = int64_array(node.sizes)[node.spatial_dims]
input_spatial_shape = input_shape[node.spatial_dims]
stride_spatial_shape = node.strides[node.spatial_dims]
size_extent = node.rates[node.spatial_dims] * (size_spatial - 1) + 1
pad_spatial_shape, output_spatial_shape = tf_window_op_pad_infer(input_spatial_shape,
size_extent,
stride_spatial_shape,
node.auto_pad,
False)
out_shape = shape_for_layout(layout,
batch=N,
features=C * np.prod(size_spatial),
height=output_spatial_shape[0],
width=output_spatial_shape[1])
node.out_port(0).data.set_shape(int64_array(out_shape))
def upsample_infer(node: Node):
layout = node.graph.graph['layout']
assert len(layout) == 4
input_shape = node.in_node(0).shape
if input_shape is None:
return
if len(node.in_nodes()) == 1:
in_height = input_shape[get_height_dim(layout, 4)]
in_width = input_shape[get_width_dim(layout, 4)]
assert node.has('width_scale') is not None and node.has(
'height_scale') is not None
out_height = math.floor(in_height * node.height_scale)
out_width = math.floor(in_width * node.width_scale)
node.out_node().shape = shape_for_layout(
layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=input_shape[get_features_dim(layout, 4)],
height=out_height,
width=out_width)
else:
assert node.in_node(1).value is not None
# generic output shape calculation to support 5D input shape case
node.out_node().shape = input_shape * node.in_node(1).value
def infer(node: Node):
in_shape = node.in_node().shape
if in_shape.size != 4:
raise Error(
'TensorFlow SpaceToDepth operation is supported for 4D \'NHWC\' input layout only. '
'Current input shape is \'{}\''.format(in_shape))
layout = node.graph.graph['layout']
N = in_shape[get_batch_dim(layout, 4)]
H = in_shape[get_height_dim(layout, 4)]
W = in_shape[get_width_dim(layout, 4)]
C = in_shape[get_features_dim(layout, 4)]
block_size = node['block_size']
if (H is not dynamic_dimension
and H % block_size) or (W is not dynamic_dimension
and W % block_size):
raise Error(
'Spatial dimensions of input tensor of SpaceToDepth operation have to be divisible by '
'SpaceToDepth \'block_size\' parameter. Input tensor shape = {}. Spatial dimensions = {},{}. '
'block_size = {}'.format(in_shape, H, W, block_size))
out_shape = shape_for_layout(layout,
batch=N,
features=C * (block_size**2),
height=H // block_size,
width=W // block_size)
node.out_port(0).data.set_shape(out_shape)
def roipooling_infer(node: Node):
"""
Sets shape of output node according specified parameters input blobs and node
Sets number from the first input blob, channels from the second one, height and width are specified
Parameters
----------
node
"""
shapes = [node.in_node(i).shape for i in range(len(node.in_nodes()))]
if any(s is None for s in shapes):
return
if len(node.in_nodes()) == 4: # TensorFlow case of CropAndResize operation
crop_size = node.in_node(3).value
if crop_size is None:
log.error('The ROIPooling size is not known for node {}'.format(
node.soft_get('name')))
return
if not isinstance(crop_size, np.ndarray) or len(crop_size) != 2:
log.error(
'The ROIPooling size is should have 2 elements for node {}'.
format(node.soft_get('name')))
node.pooled_h = crop_size[0]
node.pooled_w = crop_size[1]
node.graph.remove_edge(node.in_node(3).id, node.id)
node.graph.remove_edge(node.in_node(2).id, node.id)
layout = node.graph.graph['layout']
assert len(layout) == 4
node.out_node().shape = shape_for_layout(
layout,
batch=shapes[1][get_batch_dim(layout, 4)],
features=shapes[0][get_features_dim(layout, 4)],
height=node.pooled_h,
width=node.pooled_w)
def infer(node: Node):
in_shape = node.in_port(0).data.get_shape()
if in_shape.size != 4:
raise Error(
'TensorFlow DepthToSpace operation is supported for 4D \'NHWC\' input layout only. '
'Current input shape is \'{}\''.format(in_shape))
layout = node.graph.graph['layout']
N = in_shape[get_batch_dim(layout, 4)]
H = in_shape[get_height_dim(layout, 4)]
W = in_shape[get_width_dim(layout, 4)]
C = in_shape[get_features_dim(layout, 4)]
block_size = node['block_size']
if C is not dynamic_dimension and C % (block_size**2):
raise Error(
'Feature dimensions of input tensor of DepthToSpace operation have to be divisible by square '
'of DepthToSpace \'block_size\' parameter. Input tensor shape = {}. Feature dimension = {}. '
'block_size = {}'.format(in_shape, C, block_size))
out_shape = shape_for_layout(layout,
batch=N,
features=C // (block_size * block_size),
height=H * block_size,
width=W * block_size)
if is_fully_defined(in_shape) and is_fully_defined(
out_shape) and np.prod(in_shape) != np.prod(out_shape):
raise Error(
'Number of input elements "{}" is not equal to number of output elements "" for node "{}"'
''.format(in_shape, out_shape, node.soft_get('name', node.id)))
node.out_port(0).data.set_shape(out_shape)
def resample_infer(node: Node):
layout = node.graph.graph['layout']
assert len(layout) == 4
input_shape = node.in_node(0).shape
if input_shape is None:
return
in_height = input_shape[get_height_dim(layout, 4)]
in_width = input_shape[get_width_dim(layout, 4)]
if node.has('fw') and node.fw == 'tf':
dst_shape = node.in_node(1).value
if dst_shape is None or len(input_shape) != 4 or len(
dst_shape) != 2:
log.error(
'Node {} with op {} cannot be converted to Resample layer because there is no enough info about '
'src/dst shapes: src_shape = {}, dst_shape = {}'.format(
node.name, node.op, input_shape, dst_shape))
node.type = None # prevent translation to a valid IE layer
return
out_height = dst_shape[0]
out_width = dst_shape[1]
node.graph.remove_edge(node.in_node(1).id, node.id)
else:
if len(node.in_nodes()) == 1:
if node.has('width') and node.has('height'):
out_height = node.height
out_width = node.width
else:
out_height = node.factor * in_height
out_width = node.factor * in_width
else:
out_height = node.in_node(1).shape[get_height_dim(layout, 4)]
out_width = node.in_node(1).shape[get_width_dim(layout, 4)]
node.factor = factor_update(
node.factor,
[float(out_height) / in_height,
float(out_width) / in_width], [in_height, in_width],
[out_height, out_width], node.soft_get('name'))
node.out_node().shape = shape_for_layout(
layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=input_shape[get_features_dim(layout, 4)],
height=out_height,
width=out_width)
def upsample_infer(node: Node):
node_name = node.soft_get('name', node.id)
layout = node.graph.graph['layout']
assert len(
layout
) == 4, 'Input tensor rank must be equal to 4 for node "{}"'.format(
node_name)
input_shape = node.in_port(0).data.get_shape()
if len(node.in_nodes()) == 1:
in_height = input_shape[get_height_dim(layout, 4)]
in_width = input_shape[get_width_dim(layout, 4)]
assert node.has('width_scale') is not None and node.has(
'height_scale') is not None
if in_height is not dynamic_dimension:
out_height = math.floor(in_height * node.height_scale)
else:
out_height = dynamic_dimension
if in_width is not dynamic_dimension:
out_width = math.floor(in_width * node.width_scale)
else:
out_width = dynamic_dimension
node.out_port(0).data.set_shape(
shape_for_layout(layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=input_shape[get_features_dim(
layout, 4)],
height=out_height,
width=out_width))
else:
scales = node.in_port(1).data.get_value()
assert scales is not None, 'The input with scales for node "{}" is not constant'.format(
node_name)
eps = 1e-5 # This is to make rounding in case of very close number to round to closest instead of down
# generic output shape calculation to support 5D input shape case
output_shape = shape_array(
[dynamic_dimension for _ in range(len(input_shape))])
for idx in range(len(output_shape)):
if input_shape[idx] is not dynamic_dimension:
output_shape[idx] = int(
(input_shape[idx] + eps) * scales[idx])
else:
output_shape[idx] = dynamic_dimension_value
node.out_port(0).data.set_shape(output_shape)
def pad_op_transform(graph: Graph, match: dict):
op = match['op']
pad_op = match['pad_op']
input_data = pad_op.in_node(0)
if pad_op.mode != 'constant':
log.info(
'The pad node "{}" with pad mode "{}" cannot be fused.'.format(
pad_op.soft_get('name'), pad_op.mode))
return
if op.type == 'Pooling' and op.pool_method == 'max':
return
if pad_op.mode == 'constant':
fill_value = pad_op.in_port(3).data.get_value()
if fill_value is None or fill_value != 0.0:
log.info(
'The pad node "{}" with non-zero fill value cannot be fused.'.
format(pad_op.soft_get('name')))
return
input_tensor_dims = len(match['pad_output'].shape)
for in_port in [1, 2]:
pads = pad_op.in_port(in_port).data.get_value()
if pads[get_features_dim(op.graph.graph['layout'], input_tensor_dims)] != 0 or \
pads[get_batch_dim(op.graph.graph['layout'], input_tensor_dims)] != 0:
log.info(
'The pad node "{}" with padding over feature/batch dimension cannot be fused.'
.format(pad_op.soft_get('name')))
return
op.pad += np.concatenate([
pad_op.in_port(1).data.get_value().reshape([-1, 1]),
pad_op.in_port(2).data.get_value().reshape([-1, 1])
],
axis=1)
op.pad_spatial_shape = op.pad[op.spatial_dims]
op['auto_pad'] = None
if op.type == 'Pooling':
op['exclude_pad'] = False
assert (graph[match['pad_output'].node][match['op'].node][0]['in'] == 0)
edge_attrs = graph.get_edge_data(match['pad_output'].id, match['op'].id)[0]
graph.remove_edge(match['pad_output'].id, match['op'].id)
graph.add_edge(input_data.id, match['op'].id, **{'in': 0, **edge_attrs})
def psroipooling_infer(node: Node):
"""
Sets shape of output node according specified parameters input blobs and node
Sets number from the first input blob, channels from the second one, height and width are specified
Parameters
----------
node
"""
shapes = [node.in_node(i).shape for i in range(len(node.in_nodes()))]
if any(s is None for s in shapes):
return
layout = node.graph.graph['layout']
assert len(layout) == 4
node.out_node().shape = shape_for_layout(layout,
batch=shapes[1][get_batch_dim(layout, 4)],
features=node.output_dim,
height=node.group_size,
width=node.group_size)
def regionyolo_infer(node: Node):
input_shape = node.in_node(0).shape
if input_shape is None:
return
axis = get_canonical_axis_index(input_shape, node.axis)
end_axis = get_canonical_axis_index(input_shape, node.end_axis)
node.axis = axis
node.end_axis = end_axis
if node.do_softmax:
flat_dim = np.prod(input_shape[axis: end_axis + 1])
node.out_node().shape = np.array([*input_shape[:axis], flat_dim, *input_shape[end_axis + 1:]])
else:
layout = node.graph.graph['layout']
assert len(layout) == 4
node.out_node().shape = shape_for_layout(layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=(node.classes + node.coords + 1) * len(node.mask),
height=input_shape[get_height_dim(layout, 4)],
width=input_shape[get_width_dim(layout, 4)])
def regionyolo_infer(node: Node):
input_shape = node.in_port(0).data.get_shape()
axis = get_canonical_axis_index(input_shape, node.axis)
end_axis = get_canonical_axis_index(input_shape, node.end_axis)
node.axis = axis
node.end_axis = end_axis
if node.do_softmax:
dims_to_flatten = input_shape[axis: end_axis + 1]
if is_fully_defined(dims_to_flatten):
flat_dim = np.ma.prod(dims_to_flatten)
else:
flat_dim = dynamic_dimension_value
node.out_port(0).data.set_shape([*input_shape[:axis], flat_dim, *input_shape[end_axis + 1:]])
else:
layout = node.graph.graph['layout']
assert len(layout) == 4
node.out_port(0).data.set_shape(shape_for_layout(layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=(node.classes + node.coords + 1) * len(node.mask),
height=input_shape[get_height_dim(layout, 4)],
width=input_shape[get_width_dim(layout, 4)]))
def test_get_batch_dim_NCDHW(self):
self.assertEqual(get_batch_dim('NCHW', 5), 0)
def interp_infer(node: Node):
layout = node.graph.graph['layout']
assert len(layout) == 4
if len(node.in_nodes()) == 2:
src_shape = node.in_node(0).shape
dst_shape = node.in_node(1).shape
# in Caffe can be 2 inputs too, but shape should be got from shape of the second input
if node.parse_2nd_input == 'shape':
dst_shape = [
dst_shape[get_height_dim(layout, 4)],
dst_shape[get_width_dim(layout, 4)]
]
else:
# it is TF case
dst_shape = node.in_node(1).value
if src_shape is None or dst_shape is None or len(
src_shape) != 4 or len(dst_shape) != 2:
log.error(
'Node {} with op {} cannot be converted to Resample layer because there is no enough info about '
'src/dst shapes: src_shape = {}, dst_shape = {}'.format(
node.name, node.op, src_shape, dst_shape))
node.type = None # prevent translation to a valid IE layer
return
in_height = src_shape[get_height_dim(layout, 4)]
in_width = src_shape[get_width_dim(layout, 4)]
out_height = dst_shape[0]
out_width = dst_shape[1]
node.factor = factor_update(
node.factor,
[float(out_height) / in_height,
float(out_width) / in_width], [in_height, in_width],
[out_height, out_width], node.soft_get('name'))
if node.factor is None:
node['width'] = out_width
node['height'] = out_height
node.out_node().shape = shape_for_layout(
layout,
batch=src_shape[get_batch_dim(layout, 4)],
features=src_shape[get_features_dim(layout, 4)],
height=out_height,
width=out_width)
node.graph.remove_edge(node.in_node(1).id, node.id)
else:
outn = node.out_node(0)
in_shape = node.in_node(0)
num_ = in_shape.shape[get_batch_dim(layout, 4)]
channels_ = in_shape.shape[get_features_dim(layout, 4)]
height_in_ = in_shape.shape[get_height_dim(layout, 4)]
width_in_ = in_shape.shape[get_width_dim(layout, 4)]
height_out_ = height_in_ + node.pad_beg + node.pad_end
width_out_ = width_in_ + node.pad_beg + node.pad_end
if node.shrink_factor != 1 and node.zoom_factor == 1:
shrink_factor = node.shrink_factor
if shrink_factor < 1:
log.error(
'Shrink factor should be positive in node {}'.format(
node.id))
return None
height_out_ = (height_out_ - 1) / shrink_factor + 1
width_out_ = (width_out_ - 1) / shrink_factor + 1
elif node.shrink_factor == 1 and node.zoom_factor != 1:
zoom_factor = node.zoom_factor
if zoom_factor < 1:
log.error(
'Zoom factor should be positive in node {}'.format(
node.id))
return None
node['debug_message'] = 'Interp layer shape inference function may be wrong, please, try to update ' \
'layer shape inference function in the file (extensions/ops/interp.op at the ' \
'line {}).'.format(inspect.currentframe().f_lineno) + refer_to_faq_msg(100)
# Reshape methods can be different in some cases
# Commented out section represents reshape that used in deeplab-caffe
# Uncomment the following lines, if your model was trained with deeplab-caffe
# or have the same reshape method
# height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor - 1)
# width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1)
# Comment out the following lines if you use the reshape method from previous section
height_out_ = height_out_ * zoom_factor
width_out_ = width_out_ * zoom_factor
elif node.width != 0 and node.height != 0:
height_out_ = node.height
width_out_ = node.width
elif node.shrink_factor != 1 and node.zoom_factor != 1:
shrink_factor = node.shrink_factor
zoom_factor = node.zoom_factor
if shrink_factor < 1:
log.error(
'Shrink factor should be positive in node {}'.format(
node.id))
return None
if zoom_factor < 1:
log.error(
'Zoom factor should be positive in node {}'.format(
node.id))
return None
height_out_ = (height_out_ - 1) / shrink_factor + 1
width_out_ = (width_out_ - 1) / shrink_factor + 1
height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor -
1)
width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1)
outn.shape = shape_for_layout(layout,
batch=num_,
features=channels_,
height=height_out_,
width=width_out_)
def test_get_batch_dim_NDHWC(self):
self.assertEqual(get_batch_dim('NHWC', 5), 0)
def replace_pattern(self, graph: nx.MultiDiGraph, match: dict):
node = match['reduce']
if not node.has_valid('reduce_type') or node.reduce_type.lower() not in self.supported_reduce_types:
log.error("Reduce type {} is not supported for node {}".format(node.soft_get('reduce_type'), node.id))
return
reduce_type = node.reduce_type.lower()
if reduce_type not in self.pool_method_map:
log.error("Reduce type {} is not included in pool_method_map. Please update pool_method_map with new key "
"{}".format(reduce_type, reduce_type))
return
input_data = node.in_node()
output_data = node.out_node()
input_shape = node.in_node().shape
output_shape = node.out_node().shape
# normalize node.axis to exclude negative indices
node.axis = [get_canonical_axis_index(input_shape, a) for a in node.axis]
axis = node.axis
# Check that values in axis list are consecutive
for idx in range(1, len(axis)):
if axis[idx] != (axis[idx - 1] + 1):
log.error("Reduce with not consecutive axes {} is not supported ".format(axis))
return
layout = graph.graph['layout']
# So now we are sure that we can convert Reduce to appropriate operation
# 1. Calculate shape that will be used in reduction
reduction_dim = np.prod([input_shape[idx] for idx in axis])
begin_dims = np.array([input_shape[idx] for idx in range(axis[0])])
end_dim = np.prod([input_shape[idx] for idx in range(axis[-1] + 1, len(input_shape))])
# 2. Create reshape with appropriate shape
if layout == 'NCHW':
if len(begin_dims) > 2:
begin_dims = np.array([np.prod(begin_dims[0:-1]), begin_dims[-1]], dtype=np.int64)
else:
# Expand begin_dims to 2
begin_dims = np.array(np.append(begin_dims, [1] * (2 - len(begin_dims))), dtype=np.int64)
reshape_shape = np.array([*begin_dims, reduction_dim, end_dim], dtype=np.int64)
pool_window = np.array([1, 1, reduction_dim, 1], dtype=np.int64)
elif layout == 'NHWC':
begin_dims = np.prod(begin_dims)
reshape_shape = np.array([begin_dims, reduction_dim, 1, end_dim], dtype=np.int64)
pool_window = np.array([1, reduction_dim, 1, 1], dtype=np.int64)
else:
log.error('{} layout currently is not supported'.format(layout))
return
# 3. Reduce => Reshape->Pooling->Reshape
reshape_op = Reshape(graph, {'name': node.id + '/Reshape', 'dim': reshape_shape})
final_reshape_op = Reshape(graph, {'name': node.id + '/FinalReshape', 'dim': output_shape})
pooling_op = Pooling(graph,
dict(name=node.id + '/Pool',
window=pool_window,
output_spatial_shape=None,
batch_dims=np.array([get_batch_dim(layout, 4)], dtype=np.int64),
channel_dims=np.array([get_features_dim(layout, 4)], dtype=np.int64),
exclude_pad='false', pool_method=self.pool_method_map[reduce_type]))
graph.remove_edge(input_data.id, node.id)
graph.remove_edge(node.id, output_data.id)
final_reshape_op.create_node_with_data(
inputs=[pooling_op.create_node_with_data(
inputs=[reshape_op.create_node_with_data(
inputs=[input_data]
)]
)],
data_nodes=output_data)
# 4. If it is reduction with summation, we need to multiply by size of the reduction slice with Mul op
if reduce_type == 'sum':
output_data.in_node().insert_node_with_data_after(
output_data,
Power,
{'name': node.name + '/Mul', 'scale': float(reduction_dim)}
)
