def make_interpolate_reshapeable(interpolate, concat):
assert interpolate.soft_get('type') == 'Interpolate'
assert concat.soft_get('type') == 'Concat'
output_shape = interpolate.out_port(0).data.get_shape()
interp_axes = [get_canonical_axis_index(output_shape, axis) for axis in Interpolate.get_axes(interpolate)]
concat_axis = get_canonical_axis_index(output_shape, concat.axis)
if concat_axis in interp_axes:
return
concat_srcs = [port.get_source() for port in concat.in_ports().values() if not port.disconnected()]
non_interp_concat_srcs = [src for src in concat_srcs if src.node.soft_get('type') != 'Interpolate']
if len(non_interp_concat_srcs) == 0:
return
graph = interpolate.graph
src = non_interp_concat_srcs[0]
shape = Shape(graph, {'name': src.node.soft_get('name', src.node.id) + '/Shape'}).create_node()
shape.in_port(0).connect(src)
gather = create_op_with_const_inputs(graph, Gather,
{1: np.array(interp_axes, dtype=np.int32), 2: int64_array(0)},
{'name': shape.name + '/Gathered'}, shape)
interpolate.in_port(1).get_connection().set_source(gather.out_port(0))
def infer(node: Node):
real_squeeze_dims = int64_array([])
input_shape = node.in_port(0).data.get_shape()
node_name = node.soft_get('name', node.id)
if input_shape is None:
raise Error(
'Input shape is not defined for node {}'.format(node_name))
output_shape = input_shape.copy()
assert len(node.in_nodes(
)) == 2, 'The Squeeze node {} must have 2 inputs'.format(node_name)
# TODO remove the following 'if' statement when IE start support 0D tensors
squeeze_dims = node.in_port(1).data.get_value()
if squeeze_dims.ndim == 0:
squeeze_dims = squeeze_dims.reshape([1])
for dim in squeeze_dims:
if output_shape[dim] == 1 or output_shape[dim] is dynamic_dimension:
real_squeeze_dims = np.ma.append(
real_squeeze_dims,
get_canonical_axis_index(output_shape, dim))
else:
raise Error(
'Trying to squeeze dimension not equal to 1 for node "{}"'.
format(node_name))
# if squeeze_dims empty then all 1s should be removed (tf specification of Squeeze op)
if squeeze_dims.size == 0:
for i in range(output_shape.size):
if output_shape[i] == 1:
real_squeeze_dims = np.ma.append(
real_squeeze_dims,
get_canonical_axis_index(output_shape, i))
assert is_fully_defined(
real_squeeze_dims
), 'Squeeze dimension(s) is not defined for op "{}"'.format(node_name)
output_shape = shape_delete(output_shape, real_squeeze_dims)
node.out_port(0).data.set_shape(output_shape)
# make dimensions positive to correctly translate from NHWC to NCHW layout
if node.in_port(1).get_source().node.op == 'Const':
node.in_port(1).data.set_value(real_squeeze_dims)
if node.in_port(0).data.get_value() is not None:
node.out_port(0).data.set_value(
node.in_port(0).data.get_value().reshape(output_shape))
# the squeeze_dim attribute will be converted to the second input in the end of the Middle phase
PermuteInputs().set_input_permutation(node.in_node(1), node, 'input:0',
'axis')
def mxnet_slice_axis_infer(node):
in_shape = node.in_node(0).shape
node.axis = get_canonical_axis_index(in_shape, node.axis)
slice_axis = node.axis
new_shape = np.array(in_shape, dtype=np.int64)
new_shape[slice_axis] = new_shape[slice_axis] / len(node.out_nodes())
axis_size = in_shape[slice_axis]
if node.offset < 0:
node.offset += axis_size
if not node.dim:
node.dim = axis_size
elif node.dim < 0:
node.dim += axis_size
input_dim = in_shape.size
node.dim = (node.dim - node.offset)
if node.dim > in_shape[slice_axis]:
raise Error(
'{0} node dimension value is bigger than the corresponding value in the input shape {1}. ' +
'\nIn particular {2} is bigger than {3}. The Model Optimizer does not support this case. ' +
'\nTo overcome, try to edit the original model "end" property of the {0} layer.',
node.name, ','.join(str(i) for i in in_shape), str(node.dim), str(in_shape[slice_axis])
)
for i in range(0, input_dim):
if i == slice_axis:
new_shape[i] = node.dim
else:
new_shape[i] = in_shape[i]
for i in range(0, len(node.out_nodes())):
node.out_node(i)['shape'] = new_shape
def replace_pattern(self, graph: Graph, match: dict):
matmul = match['matmul']
reshape = match['reshape']
other_input_port_idx = 0 if match['matmul'].in_port(0).get_source().node.id == match['other_input'].id else 1
shape_source = match['matmul'].in_port(other_input_port_idx).get_source()
initial_reshape_pattern = reshape.in_port(1).data.get_value()
if len(initial_reshape_pattern) != 2:
return
reshape_is_A_input = matmul.in_port(0).get_source().node.id == reshape.id
if reshape_is_A_input:
idx = -1 if matmul.transpose_b else -2
else:
idx = -2 if matmul.transpose_a else -1
idx = get_canonical_axis_index(initial_reshape_pattern, idx)
shape_name = shape_source.node.soft_get('name', shape_source.node.id)
shape = Shape(graph, {'name': shape_name + '/Shape'}).create_node()
shape.in_port(0).connect(shape_source)
C = node_to_get_shape_value_of_indices(shape, [idx])
N = Const(graph, {'name': shape_name + '/MinusOne', 'value': int64_array([-1])}).create_node()
if len(initial_reshape_pattern) == 2:
if reshape_is_A_input:
reshape_pattern = [C, N] if matmul.transpose_a else [N, C]
else:
reshape_pattern = [N, C] if matmul.transpose_b else [C, N]
new_reshape_pattern = new_shape_node_from_shape_nodes(reshape_pattern)
reshape.in_port(1).get_connection().set_source(new_reshape_pattern.out_port(0))
else:
return
def argmax_infer(node: Node):
shape = node.in_node(0).shape
if shape is None:
return
# there are two inputs in TensorFlow. The second input is the axis for ArgMax
if len(node.in_nodes()) == 2:
if node.in_node(1).value is None:
log.debug('The second argument to ArgMax is None')
return
node.axis = node.in_node(1).value.item()
# remove the unnecessary input
node.graph.remove_edge(node.in_node(1).id, node.id)
num_top_axes = shape.size
if num_top_axes < 3:
num_top_axes = 3
out_shape = np.ones(num_top_axes, dtype=int)
if node.has_valid('axis'):
axis = get_canonical_axis_index(shape, node.axis)
node.axis = axis
out_shape = np.array(shape)
out_shape[axis] = node.top_k
PermuteAttrs.create_permute_attrs(node,
attrs=[('axis', 'input:0')])
else:
out_shape[0] = shape[0]
out_shape[2] = node.top_k
if node.out_max_val:
out_shape[1] = 2
node.out_node().shape = out_shape
def tf_squeeze_infer(node):
if node.squeeze_dims is None:
# TODO: implement; there is no implementation now because no test
return
real_squeeze_dims = []
input_shape = node.in_node().shape
if input_shape is None:
return
# UGLY
output_shape = input_shape.copy()
for n in node.squeeze_dims:
if output_shape[n] == 1:
real_squeeze_dims.append(get_canonical_axis_index(output_shape, n))
else:
raise Error('Trying to squeeze dimension not equal to 1 for node "{}"'.format(node.soft_get('name')))
output_shape = np.delete(output_shape, real_squeeze_dims)
node.out_node().shape = output_shape
if is_spatial_squeeze(node.graph.graph['layout'], input_shape, output_shape):
output_shape = int64_array([0, -1])
node['dim'] = output_shape
if node.in_node().value is not None:
node.out_node().value = np.array(np.reshape(node.in_node().value, output_shape))
PermuteAttrs.create_permute_attrs(node, attrs=[('dim', 'output:0')])
def infer(node: None):
input_shape = node.in_node(0).shape
name = node.soft_get('name', node.id)
if node.axes is not None and node.across_channels is not None:
raise Error('Either axes or across_channels can be set for the MVN in node "{}".'.format(name))
if node.across_channels is None:
if node.axes is not None:
# normalizing (replacing -1 with actual index)
axes_data_value = node.axes
axes = [axes_data_value.item()] if axes_data_value.size == 1 else axes_data_value
axes = [get_canonical_axis_index(input_shape, a) for a in axes]
# deduce across_channels from the axes, e.g. if the first axis is included (assuming batch is zero axis)
feature_dim = get_features_dim(node.graph.graph['layout'], len(input_shape)) \
if (4 <= len(input_shape) <= 5) \
else 1
node.across_channels = int(feature_dim in axes)
if 0 in axes:
raise Error('Reduction over the batch dimension in node "{}" '
'is not supported by the backend.'.format(name))
for i in range(2, len(input_shape)):
if i not in axes:
raise Error(
'Reduction over spatial dimensions in node "{}" '
'is obligatory for the backend.'.format(name))
else:
node.across_channels = 0 # default
copy_shape_infer(node)
def infer(node: Node):
assert len(node.in_nodes()) == 2 or len(node.in_nodes()) == 3
# There may be three inputs in TensorFlow. The third input is axis
if len(node.in_nodes()) == 3:
if node.in_node(2).value is None:
log.error("Gather is supported only with constant axis value")
return
node.axis = node.in_node(2).value.item()
node.graph.remove_edge(node.in_node(2).id, node.id)
axis = node.axis
data = node.in_node(0)
indices = node.in_node(1)
# both inputs are constant
if data.value is not None and indices.value is not None:
indices.value = np.array(indices.value, dtype=np.int64)
node.out_node(0).value = np.array(np.take(data.value,
indices.value, axis),
dtype=data.value.dtype)
node.out_node(0).shape = np.array(node.out_node(0).value.shape,
dtype=np.int64)
return
# Convert negative axis
axis = get_canonical_axis_index(data.shape, axis)
node.axis = axis
shape = np.concatenate((data.shape[:axis], indices.shape))
if axis < len(data.shape) - 1:
shape = np.concatenate((shape, data.shape[axis + 1:]))
node.out_node(0).shape = np.array(shape, dtype=np.int64)
def crop_infer(node):
"""
Crops the shape of the output blob according to input ones be specified params.
Node should have 2 input blobs - 1st blob is getting cropped by specified axis according
to the the 2nd (reference) blob.
The result blob is written to output node shape, and reference blob is removed from graph.
In order to save the reference dims, it is written to dims parameter.
Parameters
----------
node
"""
N = len(node.in_nodes())
if N < 2:
log.debug('Wrong number of bottom blobs in ' + node.node)
return
shapes = [node.in_node(i).shape for i in range(N)]
if any(s is None for s in shapes):
return
input_shape = np.array(shapes[0])
start_axis = get_canonical_axis_index(input_shape, node.axis)
node.axis = start_axis
reference_shape = np.array(shapes[1])
input_dim = input_shape.size
# set new shape to current shape
new_shape = input_shape.copy()
ir_axis = []
ir_offset = []
dim = []
for i in range(0, input_dim):
if i < start_axis:
new_shape[i] = input_shape[i]
continue
crop_offset = 0
if len(node.offset) == 1:
crop_offset = node.offset[0]
elif len(node.offset) > 1:
crop_offset = node.offset[i - start_axis]
if input_shape[i] - crop_offset < reference_shape[i]:
log.error('The crop for dimension is out of bounds in ' + node.node)
return
dim.append(reference_shape[i])
ir_axis.append(i)
ir_offset.append(crop_offset)
new_shape[i] = reference_shape[i]
node.axis = ir_axis
node.offset = ir_offset
node.dim = dim
node.out_node().shape = new_shape
def infer(node: Node):
name = node.soft_get('name', node.id)
connected_in_ports = {idx: port for idx, port in node.in_ports().items() if not port.disconnected()}
assert len(connected_in_ports) == 3 and 0 in connected_in_ports and 1 in connected_in_ports and \
2 in connected_in_ports, "Gather should have 3 connected input port, but it doesn't for " \
"node: `{}`. Ports: {}".format(name, connected_in_ports)
data_shape = node.in_port(0).data.get_shape()
assert data_shape is not None
indices_shape = node.in_port(1).data.get_shape()
assert indices_shape is not None
axis = node.in_port(2).data.get_value()
assert axis is not None
axis = get_canonical_axis_index(data_shape, axis)
# we import PermuteInputs locally because it uses Gather inside and we have recursive imports
from mo.graph.perm_inputs import PermuteInputs
PermuteInputs().set_input_permutation(node.in_node(1), node, 'input:0', 'axis')
data_value = node.in_port(0).data.get_value()
indices_value = node.in_port(1).data.get_value()
if data_value is not None and indices_value is not None:
node.out_port(0).data.set_value(np.array(np.take(data_value, int64_array(indices_value), axis),
dtype=data_value.dtype))
return
shape = np.concatenate((data_shape[:axis], indices_shape))
if axis < len(data_shape) - 1:
shape = np.concatenate((shape, data_shape[axis + 1:]))
node.out_port(0).data.set_shape(int64_array(shape))
def infer(node: Node):
name = node.soft_get('name', node.id)
connected_in_ports = {idx: port for idx, port in node.in_ports().items() if not port.disconnected()}
assert len(connected_in_ports) == 2 and 0 in connected_in_ports and 1 in connected_in_ports, \
"AttributedGather should have 2 connected input port, but it doesn't for node: `{}`. Ports: {}" \
"".format(name, connected_in_ports)
axis = node.soft_get('axis', None)
assert axis is not None
data_shape = node.in_port(0).data.get_shape()
assert data_shape is not None
indices_shape = node.in_port(1).data.get_shape()
assert indices_shape is not None
# Convert negative axis
axis = get_canonical_axis_index(data_shape, axis)
node.axis = axis
PermuteAttrs.create_permute_attrs(node, attrs=[('axis', 'input:0')])
data_value = node.in_port(0).data.get_value()
indices_value = node.in_port(1).data.get_value()
if data_value is not None and indices_value is not None:
node.out_port(0).data.set_value(np.array(np.take(data_value, indices_value, axis), dtype=data_value.dtype))
return
shape = np.concatenate((data_shape[:axis], indices_shape))
if axis < len(data_shape) - 1:
shape = np.concatenate((shape, data_shape[axis + 1:]))
node.out_port(0).data.set_shape(int64_array(shape))
def infer(node):
input_shape = node.in_port(0).data.get_shape()
shape_like = node.in_port(1).data.get_shape()
new_shape = np.copy(input_shape)
if node.axes is not None:
node.axes = sorted(
[get_canonical_axis_index(input_shape, i) for i in node.axes])
for i in node.axes:
new_shape[i] = shape_like[i]
else:
assert input_shape.size == shape_like.size,\
'Input shape ranks are inconsistent: {} and {}'.format(input_shape.size, shape_like.size)
node.axes = int64_array(range(shape_like.size))
new_shape = np.copy(shape_like)
node.out_port(0).data.set_shape(new_shape)
if node.in_port(0).get_connection().data.get_value() is not None:
out_value = np.copy(node.in_port(0).data.get_value())
slice_indexes = []
for s in out_value.shape:
slice_indexes.append(slice(0, s))
for axis in node.axes:
slice_indexes[axis] = slice(0, new_shape[axis])
out_value = out_value[tuple(slice_indexes)]
node.out_port(0).data.set_value(out_value)
def infer(node):
input_shape = node.in_node(0).shape
if input_shape is None:
log.debug(
'The input shape for the layer "{}" is not defined'.format(
node.soft_get('name')))
return
axis = get_canonical_axis_index(input_shape, node.axis)
end_axis = node.end_axis if node.has('end_axis') else -1
end_axis = get_canonical_axis_index(input_shape, end_axis)
prod_axes = np.prod(input_shape[axis:end_axis + 1])
node.out_node(0).shape = int64_array(
[*input_shape[0:axis], prod_axes, *input_shape[end_axis + 1:]])
if node.in_node().has_valid('value'):
node.out_node().value = node.in_node().value.copy().reshape(
node.out_node(0).shape)
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 _two_inputs_infer(node: Node):
N = len(node.in_nodes())
shapes = [node.in_node(i).shape for i in range(N)]
if any(s is None for s in shapes):
log.error('Not all input shapes were defined for {} node'.format(node.name))
return
if not node.has_valid('axis'):
log.error('axis attribute is missing for {} node. should be set in crop extractor'.format(node.name))
return
if not node.has_valid('offset'):
log.error('offset attribute is missing for {} node. should be set in crop extractor'.format(node.name))
return
input_shape = np.array(shapes[0])
start_axis = get_canonical_axis_index(input_shape, node.axis)
node.axis = start_axis
reference_shape = np.array(shapes[1])
input_dim = input_shape.size
# set new shape to current shape
new_shape = input_shape.copy()
ir_axis = []
ir_offset = []
dim = []
for i in range(0, input_dim):
if i < start_axis:
new_shape[i] = input_shape[i]
continue
crop_offset = 0
if len(node.offset) == 1:
crop_offset = node.offset[0]
elif len(node.offset) > 1:
crop_offset = node.offset[i - start_axis]
if input_shape[i] - crop_offset < reference_shape[i]:
log.error('The crop for dimension is out of bounds in ' + node.node)
return
dim.append(reference_shape[i])
ir_axis.append(i)
ir_offset.append(crop_offset)
new_shape[i] = reference_shape[i]
node.axis = ir_axis
node.offset = ir_offset
node['dim'] = dim
node.out_node().shape = new_shape
PermuteAttrs.create_permute_attrs(node, attrs=[('axis', 'input:0')])
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 concat_infer(node):
if not node.has('axis'):
N = node.N
axis_input = node.in_node(N)
if axis_input.has_valid('value') and axis_input.value.size == 1:
node['axis'] = axis_input.value.item()
node.graph.remove_edge(
axis_input.node,
node.node) # TODO add skip attribute instead of deleting
else:
return
else:
N = len(node.in_nodes())
shapes = [node.in_node(i).shape for i in range(N)]
if any(s is None for s in shapes):
return
shape = np.array(shapes[0])
axis = get_canonical_axis_index(shape, node.axis)
node.axis = axis
mask = np.zeros_like(shape, dtype=np.bool)
mask[axis] = True # pylint: disable=unsupported-assignment-operation
not_mask = np.logical_not(mask) # pylint: disable=assignment-from-no-return
for s in shapes[1:]:
s = int64_array(s)
if np.all(shape[not_mask] ==
s[not_mask]): # TODO handle -1 in a special way
shape[mask] += s[mask]
else:
log.error('Concat input shapes do not match')
return
node.out_node(0).shape = shape
if len(shape) != 4:
# exclude it from NHWC to NCHW conversion
if 'axis' in node.dim_attrs:
node.dim_attrs.remove('axis')
PermuteAttrs.create_permute_attrs(node, attrs=[('axis', 'input:0')])
values = [node.in_node(i).value for i in range(N)]
if any(v is None for v in values):
return
node.out_node(0).value = np.concatenate(values, axis=node.axis).astype(
values[0].dtype, copy=False)
node.out_node(0).shape = np.array(node.out_node(0).value.shape,
dtype=np.int64)
def find_and_replace_pattern(self, graph: Graph):
for node in graph.get_op_nodes(op='Slice'):
node_name = node.soft_get('name', node.id)
input_shape = node.in_port(0).data.get_shape()
if node.is_in_port_connected(3):
axes = node.in_port(3).data.get_value().copy()
assert axes is not None, 'The input with axes is not constant for node {}'.format(node_name)
for i, val in enumerate(axes):
axes[i] = get_canonical_axis_index(input_shape, val)
else:
axes = int64_array(range(len(input_shape)))
ss_begin = create_ss_interval_border(graph, node.in_port(1).get_source(), input_shape, axes, node_name)
ss_end = create_ss_interval_border(graph, node.in_port(2).get_source(), input_shape, axes, node_name)
node.in_port(1).disconnect()
node.in_port(2).disconnect()
rename_nodes([(ss_begin, node_name + '/Begin'), (ss_end, node_name + '/End')])
if node.is_in_port_connected(4):
steps = node.in_port(4).data.get_value()
assert steps is not None, 'The input with steps is not constant for node {}'.format(node_name)
else:
steps = np.ones([axes.size])
ss_begin_mask = np.zeros(len(input_shape), dtype=np.int64)
ss_end_mask = np.zeros(len(input_shape), dtype=np.int64)
ss_step = np.ones(len(input_shape), dtype=np.int64)
for i, axis in enumerate(axes):
ss_begin_mask[axis] = 1
ss_end_mask[axis] = 1
ss_step[axis] = steps[i]
ss_strides = Const(graph, dict(name=node_name + '/Strides', value=ss_step)).create_node()
ss = StridedSlice(graph, dict(name='ss', new_axis_mask=np.zeros(len(input_shape), dtype=np.int64),
shrink_axis_mask=np.zeros(len(input_shape), dtype=np.int64),
ellipsis_mask=np.zeros(len(input_shape), dtype=np.int64),
begin_mask=ss_begin_mask,
end_mask=ss_end_mask)).create_node()
node.in_port(0).get_connection().set_destination(ss.in_port(0))
ss.in_port(1).connect(ss_begin.out_port(0))
ss.in_port(2).connect(ss_end.out_port(0))
ss.in_port(3).connect(ss_strides.out_port(0))
node.out_port(0).get_connection().set_source(ss.out_port(0))
rename_nodes([(node, node_name + '/ShouldBeDeleted'), (ss, node_name)])
def infer(node):
input_shape = node.in_port(0).data.get_shape()
shape_like = node.in_port(1).data.get_shape()
new_shape = np.copy(input_shape)
if node.axes is not None:
node.axes = sorted(
[get_canonical_axis_index(input_shape, i) for i in node.axes])
for i in node.axes:
new_shape[i] = shape_like[i]
else:
assert input_shape.size == shape_like.size,\
'Input shape ranks are inconsistent: {} and {}'.format(input_shape.size, shape_like.size)
node.axes = int64_array(range(shape_like.size))
new_shape = np.copy(shape_like)
node.out_port(0).data.set_shape(new_shape)
def infer(node: Node):
real_squeeze_dims = int64_array([])
input_shape = node.in_node().shape
if input_shape is None:
return
output_shape = input_shape.copy()
assert len(node.in_nodes()
) == 2, 'The Squeeze node {} must have 2 inputs'.format(
node.soft_get('name'))
# TODO remove the following 'if' statement when IE start support 0D tensors
squeeze_dims = node.in_port(1).data.get_value()
if squeeze_dims.ndim == 0:
squeeze_dims = squeeze_dims.reshape([1])
for dim in squeeze_dims:
if output_shape[dim] == 1:
real_squeeze_dims = np.append(
real_squeeze_dims,
get_canonical_axis_index(output_shape, dim))
else:
raise Error(
'Trying to squeeze dimension not equal to 1 for node "{}"'.
format(node.soft_get('name')))
output_shape = np.delete(output_shape, real_squeeze_dims)
node.out_node().shape = output_shape
# make dimensions positive to correctly translate from NHWC to NCHW layout
if node.in_port(1).get_source().node.op == 'Const':
node.in_port(1).data.set_value(real_squeeze_dims)
if node.in_port(0).data.get_value() is not None:
node.out_port(0).data.set_value(
node.in_port(0).data.get_value().reshape(output_shape))
# the squeeze_dim attribute will be converted to the second input in the end of the Middle phase
PermuteInputs().set_input_permutation(node.in_node(1), node, 'input:0',
'axis')
def arg_ops_infer(node: Node):
shape = node.in_port(0).data.get_shape()
node_name = node.soft_get('name', node.id)
assert shape is not None, "Input shape for the node {} is None".format(
node_name)
# there are two inputs in TensorFlow. The second input is the axis for ArgMax
connected_in_ports = [
port for port in node.in_ports().values() if not port.disconnected()
]
if len(connected_in_ports) == 2:
axis = node.in_port(1).data.get_value()
if axis is None:
log.debug('The second argument to {} is None'.format(
node.soft_get('name', node.id)))
return
node.axis = axis
# remove the unnecessary input
node.in_port(1).disconnect()
num_top_axes = shape.size
if num_top_axes < 3:
num_top_axes = 3
out_shape = np.ones(num_top_axes, dtype=np.int64)
if node.has_valid('axis'):
axis = get_canonical_axis_index(shape, node.axis)
node.axis = axis
out_shape = shape.copy()
out_shape[axis] = node.top_k
PermuteAttrs.create_permute_attrs(node, attrs=[('axis', 'input:0')])
else:
out_shape[0] = shape[0]
out_shape[2] = node.top_k
if node.has_and_set('out_max_val'):
out_shape[1] = 2
node.out_port(0).data.set_shape(out_shape)
def concat_infer(node):
node_name = node.soft_get('name', node.id)
if not node.has('axis'):
N = node.N
axis_input = node.in_node(N)
if axis_input.has_valid('value') and axis_input.value.size == 1:
node['axis'] = axis_input.value.item()
node.graph.remove_edge(
axis_input.node,
node.node) # TODO add skip attribute instead of deleting
else:
raise Error(
'Input with value is not specified for node "{}"'.format(
node_name))
else:
N = len(node.in_nodes())
shapes = [node.in_node(i).shape for i in range(N)]
if any(s is None for s in shapes):
raise Error(
'One of the input shapes is not defined for node "{}"'.format(
node_name))
shape = shape_array(shapes[0])
axis = get_canonical_axis_index(shape, node.axis)
node.axis = axis
mask = np.zeros_like(shape, dtype=np.bool)
mask[axis] = True # pylint: disable=unsupported-assignment-operation
not_mask = np.logical_not(mask) # pylint: disable=assignment-from-no-return
for s in shapes[1:]:
s = shape_array(s)
if np.ma.allequal(shape[not_mask], s[not_mask]):
shape[mask] += s[mask]
else:
raise Error(
'Concat input shapes do not match for node "{}" with axis {}'.
format(node_name, axis))
# dynamic dimensions in the output (except the concat axis) can be deduced from input shape
for pos in range(len(shape)):
if shape[pos] is dynamic_dimension and pos != axis:
for in_shape in shapes:
if in_shape[pos] is not dynamic_dimension:
shape[pos] = in_shape[pos]
node.out_port(0).data.set_shape(shape)
PermuteAttrs.create_permute_attrs(node, attrs=[('axis', 'input:0')])
values = [node.in_node(i).value for i in range(N)]
if any([v is None for v in values]):
return
# if one of the input values are dynamic, the output tensor type is inferred from one of the fully defined inputs
output_dtype = np.int64
for input in values:
if is_fully_defined(input):
output_dtype = input.dtype
if any(not is_fully_defined(v) for v in values):
node.out_port(0).data.set_value(
np.ma.concatenate(values, axis=node.axis).astype(output_dtype))
else: # there is a serious performance benefit to use concatenation as it is implemented below
node.out_node(0).value = np.concatenate(values, axis=node.axis).astype(
values[0].dtype, copy=False)
node.out_node(0).shape = shape_array(node.out_node(0).value.shape)
def infer(node: Node):
shape = np.copy(node.in_node().shape)
node.axis = get_canonical_axis_index(shape, node.axis)
rep = node.repeats
shape[node.axis] = shape[node.axis] * rep
node.out_node().shape = shape
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)}
)
def replace_pattern(self, graph: Graph, match: dict):
node = match['reduce']
if node.out_port(0).data.get_value() is not None:
# We leave Reduce* operations located in constant sub-graph as is
# to keep model reshapable with --keep_shape_ops cli key
return
reduce_type = node.type
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_port(0).data.get_shape()
output_shape = node.out_port(0).data.get_shape()
# normalize node axes to exclude negative indices
axes_data_value = node.in_port(1).data.get_value()
axes = int64_array([
axes_data_value.item()
]) if axes_data_value.size == 1 else axes_data_value
axes = [get_canonical_axis_index(input_shape, a) for a in axes]
axes = sorted(axes)
# Check that values in axes list are consecutive
for idx in range(1, len(axes)):
if axes[idx] != (axes[idx - 1] + 1):
log.error(
"Reduce with not consecutive axes {} is not supported ".
format(axes))
return
# 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 axes])
begin_dims = np.array([input_shape[idx] for idx in range(axes[0])])
end_dim = np.prod([
input_shape[idx] for idx in range(axes[-1] + 1, len(input_shape))
])
# 2. Create reshape with appropriate shape
if len(begin_dims) > 2:
if 0 not in axes:
begin_dims = int64_array(
[begin_dims[0], np.prod(begin_dims[1:])])
else:
begin_dims = int64_array(
[np.prod(begin_dims[0:-1]), begin_dims[-1]])
else:
# Expand begin_dims to 2
begin_dims = int64_array(
np.append(begin_dims, [1] * (2 - len(begin_dims))))
reshape_shape = int64_array([*begin_dims, reduction_dim, end_dim])
pool_window = int64_array([1, 1, reduction_dim, 1])
if end_dim == 1:
new_window = ReduceReplacer.initial_reshape_dim_normalizer(
reduction_dim)
reshape_shape = int64_array([*begin_dims, *new_window])
pool_window = int64_array([1, 1, *new_window])
# 3. Reduce => Reshape->Pooling->Reshape
reshape_op = Reshape(graph, {'name': node.id + '/Reshape'})
reshape_dim_const_data = Const(graph, {
'name': node.id + '/Reshape/Dim',
'value': reshape_shape
}).create_node_with_data()
final_reshape_op = Reshape(graph, {'name': node.id + '/FinalReshape'})
final_reshape_dim_const_data = Const(graph, {
'name': node.id + '/FinalReshape/Dim',
'value': output_shape
}).create_node_with_data()
pooling_op = Pooling(
graph,
dict(name=node.id + '/Pool',
window=pool_window,
output_spatial_shape=None,
batch_dims=int64_array([0]),
channel_dims=int64_array([1]),
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)
if np.array_equal(input_shape, reshape_shape):
input_to_pooling = input_data
else:
input_to_pooling = reshape_op.create_node_with_data(
inputs=[input_data, reshape_dim_const_data])
pooling = pooling_op.create_node_with_data(inputs=[input_to_pooling])
final_reshape_op.create_node_with_data(
inputs=[pooling, final_reshape_dim_const_data],
data_nodes=output_data)
# convert batch dimension to 0 to produce reshape-able IR over the batch dimension
if 0 not in axes:
reshape_dim_const_data.in_node(0).value[0] = 0
final_reshape_dim_const_data.in_node(0).value[0] = 0
# 4. If it is reduction with summation, we need to multiply by size of the reduction slice with Mul op
if reduce_type == 'ReduceSum':
output_data.in_node().insert_node_with_data_after(
output_data, AttributedPower, {
'name': node.name + '/Mul',
'scale': float(reduction_dim)
})
def test_posirive_index(self):
shape = [1, 2, 3, 4]
inds = [0, 1, 2, 3]
expected_inds = [0, 1, 2, 3]
for i in range(len(inds)):
assert get_canonical_axis_index(shape, inds[i]) == expected_inds[i]
def _two_inputs_infer(node: Node):
N = len(node.in_nodes())
shapes = [node.in_node(i).shape for i in range(N)]
if any(s is None for s in shapes):
log.error('Not all input shapes were defined for {} node'.format(
node.name))
return
if not node.has_valid('axis'):
log.error(
'axis attribute is missing for {} node. should be set in crop extractor'
.format(node.name))
return
if not node.has_valid('offset'):
log.error(
'offset attribute is missing for {} node. should be set in crop extractor'
.format(node.name))
return
input_shape = np.array(shapes[0])
start_axis = get_canonical_axis_index(input_shape, node.axis)
node.axis = start_axis
reference_shape = np.array(shapes[1])
if node.has_valid('axes'):
'''
The axes parameter contain shape indexes for second input and
show which shape indexes we need to use for dim attribute.
'''
input_dim = node.axes
node.in_port(1).disconnect()
else:
input_dim = list(range(0, input_shape.size))
# set new shape to current shape
new_shape = input_shape.copy()
ir_axis = []
ir_offset = []
dim = []
for i in input_dim:
if i < start_axis:
new_shape[i] = input_shape[i]
continue
crop_offset = 0
if len(node.offset) == 1:
crop_offset = node.offset[0]
elif len(node.offset) > 1:
crop_offset = node.offset[i - start_axis]
if input_shape[i] - crop_offset < reference_shape[i]:
log.error('The crop for dimension is out of bounds in ' +
node.node)
return
dim.append(reference_shape[i])
ir_axis.append(i)
ir_offset.append(crop_offset)
new_shape[i] = reference_shape[i]
node.axis = ir_axis
node.offset = ir_offset
node['dim'] = dim
node.out_node().shape = new_shape
if node.in_node(0).has_valid('value') and not node.graph.graph[
'cmd_params'].enable_ssd_gluoncv:
out_value = np.copy(node.in_node(0).value)
slice_indexes = []
for s in out_value.shape:
slice_indexes.append(slice(0, s))
for axis in input_dim:
slice_indexes[axis] = slice(0, new_shape[axis])
out_value = out_value[tuple(slice_indexes)]
node.out_node().value = out_value
PermuteAttrs.create_permute_attrs(node, attrs=[('axis', 'input:0')])