def replace_interpolate_pattern(graph: Graph, match: dict):
split = match['split']
scale = float32_array([get_split_scale(split)])
axis = int(split.in_port(1).get_connection().get_source().node.value)
split_node_name = split.name
axis_node = Const(graph, {'name': split_node_name + '/axis', 'value': int64_array([axis])}).create_node()
shape_node = Shape(graph, dict(name=split_node_name + '/Shape')).create_node()
scales_node = Const(graph, dict(name=split_node_name + '/scales', value=scale)).create_node()
mul_node = Mul(graph, dict(name=split_node_name + '/Mul')).create_node()
scales_node.out_port(0).connect(mul_node.in_port(1))
strided_slice_node = create_op_with_const_inputs(graph,
StridedSlice,
{1: int64_array([axis]), 2: int64_array([axis + 1])},
{
'name': split_node_name + '/StridedSlice',
'begin_mask': int64_array([1]),
'end_mask': int64_array([1]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
shape_node.out_port(0).connect(strided_slice_node.in_port(0))
cast_shape_to_float = Cast(graph, {'dst_type': np.float32}).create_node()
strided_slice_node.out_port(0).connect(cast_shape_to_float.in_port(0))
cast_shape_to_float.out_port(0).connect(mul_node.in_port(0))
interp_node = Interpolate(graph,
dict(name=split_node_name + '/Interpolate',
mode='nearest',
antialias=0, pads_begin=int64_array([0]), pads_end=int64_array([0]),
coordinate_transformation_mode='half_pixel', nearest_mode='round_prefer_floor',
cube_coeff=-0.75, version='opset4', shape_calculation_mode='scales',
in_ports_count=4, maybe_part_of_sequence=True)).create_node()
floor_node = Floor(graph, {'name': split_node_name + '/Floor'}).create_node()
cast_mul_result_to_int = Cast(graph, {'dst_type': np.int64}).create_node()
mul_node.out_port(0).connect(floor_node.in_port(0))
floor_node.out_port(0).connect(cast_mul_result_to_int.in_port(0))
cast_mul_result_to_int.out_port(0).connect(interp_node.in_port(1))
scales_node.out_port(0).connect(interp_node.in_port(2))
axis_node.out_port(0).connect(interp_node.in_port(3))
match['concat'].out_port(0).get_connection().set_source(interp_node.out_port(0))
split_connection = split.in_port(0).get_connection()
split_connection.set_destination(interp_node.in_port(0))
split_connection.get_source().connect(shape_node.in_port(0))
def find_and_replace_pattern(self, graph: Graph):
for node in graph.get_op_nodes(op='LayerNorm'):
node_name = node.soft_get('name', node.id)
if node.output_mean_var is True:
if not node.out_port(1).disconnected() or not node.out_port(2).disconnected():
raise Error("Node {} is supported with only one output".format(node_name))
log.error('LayerNorm node {} with attribute "output_mean_var" = True is not supported.'
'But since the node has one output, the conversion will continue.'.format(node_name),
extra={'is_warning': True})
input_shape = node.in_port(0).data.get_shape()
assert node.has_valid('axis'), 'Incorrect axis value for the node {}'.format(node_name)
axis = node.axis
mvn = create_op_node_with_second_input(graph, MVN, int64_array([axis]),
dict(eps=node.epsilon, name=node_name + '/LayerNorm/MVN_',
across_channels=1, normalize_variance=1, eps_mode='inside_sqrt'))
mul = Mul(graph, {'name': node_name + '/LayerNorm/mul_'}).create_node()
add = Add(graph, {'name': mul.name + '/LayerNorm/add_'}).create_node()
node.in_port(0).get_connection().set_destination(mvn.in_port(0))
node.in_port(1).get_connection().set_destination(mul.in_port(1))
node.in_port(2).get_connection().set_destination(add.in_port(1))
mvn.out_port(0).connect(mul.in_port(0))
mul.out_port(0).connect(add.in_port(0))
node.out_port(0).get_connection().set_source(add.out_port(0))
# MXNet LayerNorm gamma and beta attributes are 1D tensors with shape = [input_shape[axis]]
# We have to unsqueeze values for Mul and Add operations to avoid shapes incompatibility problems
# if axis != -1
canonical_axis = get_canonical_axis_index(input_shape, axis)
unsqueeze_value = []
for idx, val in enumerate(input_shape):
if idx != canonical_axis:
unsqueeze_value.append(idx)
mul_const_unsqueeze = create_op_node_with_second_input(graph, Unsqueeze,
int64_array(unsqueeze_value),
dict(name=mul.name + '/Unsqueeze',
override_output_shape=True))
add_const_unsqueeze = create_op_node_with_second_input(graph, Unsqueeze,
int64_array(unsqueeze_value),
dict(name=add.name + '/Unsqueeze',
override_output_shape=True))
mul.in_port(1).get_connection().insert_node(mul_const_unsqueeze)
add.in_port(1).get_connection().insert_node(add_const_unsqueeze)
rename_nodes([(node, node_name + '/ShouldBeDeleted'), (add, node_name)])
def div_to_mul_replacement(div: Node):
# we execute this transformation for V10 IR later on middle phase despite graph_condition
# so we prevent Div replacement on shape-calculating sub-graphs
if div.in_port(0).data.get_value() is not None and div.in_port(1).data.get_value() is not None:
return
# cannot replace Div with Mul when the divisor is integer because the reciprocal number will be 0
value = div.in_port(1).data.get_value()
if value is not None and type(value.item(0)) == int:
return
graph = div.graph
name = div.soft_get('name', div.id)
# keep Mul name the same as Div -- because of mathematical equality of output tensors
rename_node(node=div, name=name + '/to_be_removed')
# reconnect Div in(out)puts to Mul
mul = Mul(graph, {'name': name}).create_node()
rename_node(mul, name)
div.in_port(0).get_connection().set_destination(mul.in_port(0))
div.in_port(1).get_connection().set_destination(mul.in_port(1))
div.out_port(0).get_connection().set_source(mul.out_port(0))
# restore mathematical equivalence to Div operation: Div(A, B) = Mul(A, Pow(B, -1))
reciprocal = create_op_with_const_inputs(graph, Pow, {1: np.float64(-1)}, {'name': name + '/reciprocal_'})
mul.in_port(1).get_connection().insert_node(reciprocal)
def replace_sub_graph(self, graph: Graph, match: dict):
op = match['op']
out_port = op.in_port(0).get_source()
if op.soft_get('scale', 1) != 1:
const = Const(graph, {'value': np.array(op.scale)}).create_node()
mul = Mul(graph, {'name': op.name + '/mul_'}).create_node()
const.out_port(0).connect(mul.in_port(1))
out_port.connect(mul.in_port(0))
out_port = mul.out_port(0)
if op.soft_get('shift', 0) != 0:
const = Const(graph, {'value': np.array(op.shift)}).create_node()
add = Add(graph, {'name': op.name + '/add_'}).create_node()
const.out_port(0).connect(add.in_port(1))
out_port.connect(add.in_port(0))
out_port = add.out_port(0)
if op.soft_get('power', 1) != 1:
const = Const(graph, {'value': np.array(op.power)}).create_node()
pow = Pow(graph, {'name': op.name + '/pow_'}).create_node()
const.out_port(0).connect(pow.in_port(1))
out_port.connect(pow.in_port(0))
out_port = pow.out_port(0)
op.out_port(0).get_connection().set_source(out_port)
def find_and_replace_pattern(self, graph: Graph):
for dequantize_node in graph.get_op_nodes(op='DequantizeLinear'):
node_name = dequantize_node.soft_get('name', dequantize_node.id)
axis = dequantize_node.soft_get('axis', None)
scale_y_shape = dequantize_node.in_port(1).data.get_shape()
model_data_type = data_type_str_to_np(
graph.graph['cmd_params'].data_type)
cast = Cast(graph, {
'dst_type': model_data_type,
'name': node_name + '/Cast'
}).create_node()
dequantize_node.in_port(0).get_connection().set_destination(
cast.in_port(0))
mul = Mul(graph, {'can_be_fused': False}).create_node()
is_second_port_connected = dequantize_node.is_in_port_connected(2)
if is_second_port_connected:
# its is necessary not to replace subrtract for pattern in offline transformations
# See ConvertQuantizeDequantize transformation in ngraph
sub = Sub(graph, {
'name': node_name + '/Sub',
'zero_point_sub': True
}).create_node()
cast.out_port(0).connect(sub.in_port(0))
dequantize_node.in_port(2).get_connection().set_destination(
sub.in_port(1))
sub.out_port(0).connect(mul.in_port(0))
else:
cast.out_port(0).connect(mul.in_port(0))
dequantize_node.in_port(1).get_connection().set_destination(
mul.in_port(1))
dequantize_node.out_port(0).get_connection().set_source(
mul.out_port(0))
rename_nodes([(dequantize_node, node_name + '/TBD'),
(mul, node_name)])
assert scale_y_shape is not None
if axis is not None and len(
scale_y_shape) > 0 and scale_y_shape[0] > 1:
input_shape = cast.in_port(0).data.get_shape()
target_shape = np.ones(len(input_shape), np.int64)
target_shape[axis] = input_shape[axis]
mul_reshape = create_op_with_const_inputs(
graph, Reshape, {1: int64_array(target_shape)},
{'name': node_name + '/Reshape/Mul'})
mul.in_port(1).get_connection().set_destination(
mul_reshape.in_port(0))
mul_reshape.out_port(0).connect(mul.in_port(1))
if is_second_port_connected:
sub_reshape = create_op_with_const_inputs(
graph, Reshape, {1: int64_array(target_shape)},
{'name': node_name + '/Reshape/Sub'})
sub.in_port(1).get_connection().set_destination(
sub_reshape.in_port(0))
sub_reshape.out_port(0).connect(sub.in_port(1))
def replace_pattern(self, graph: Graph, match: dict):
quantize = match['quantize']
sum_node = Add(graph, dict()).create_node()
const = Const(graph, {'value': mo_array(0.5)}).create_node()
mul_node = Mul(graph, dict()).create_node()
mul_node.in_port(0).connect(sum_node.out_port(0))
mul_node.in_port(1).connect(const.out_port(0))
quantize.in_port(1).get_connection().get_source().connect(sum_node.in_port(0))
quantize.in_port(2).get_connection().get_source().connect(sum_node.in_port(1))
quantize.in_port(1).disconnect()
quantize.in_port(2).disconnect()
mul_node.out_port(0).connect(quantize.in_port(1))
mul_node.out_port(0).connect(quantize.in_port(2))
def _fused_batch_norm_decomposition(graph: Graph, tinput: Port, toutput: Port, gamma: Port, beta: Port,
mean: np.ndarray, variance: np.ndarray, can_be_fused=True):
"""
This is common function for TF, Caffe and MXNet
It creates Mul->Add->Mul->Add sub graph
"""
batch_norm_name = tinput.get_connection().get_destination().node.name
# Create first Mul & Add operations
mul1_node = Mul(graph, dict(name=batch_norm_name + "/mean", can_be_fused=can_be_fused)).create_node()
add1_node = Add(graph, dict(name=batch_norm_name + "/variance", can_be_fused=can_be_fused)).create_node()
const_mul1_node = Const(graph, dict(name="data_mul_", value=mo_array(mean))).create_node()
const_add1_node = Const(graph, dict(name="data_add_", value=mo_array(variance))).create_node()
# Broadcast const from scalar
# We can broadcast only when const.value is scalar
if gamma.data.get_shape()[0] != gamma.data.get_value().shape[0]:
value = gamma.data.get_value()
value.resize(gamma.data.get_shape()).fill(value[0])
gamma.data.set_value(value)
# Create second Mul & Add
mul2_node = Mul(graph, dict(name=batch_norm_name + "/gamma", can_be_fused=can_be_fused)).create_node()
add2_node = Add(graph, dict(name=batch_norm_name + "/beta", can_be_fused=can_be_fused)).create_node()
# Connect edges Mul1->Add1->Mul2->Add2
tinput.get_connection().set_destination(mul1_node.in_port(0))
mul1_node.in_port(1).get_connection().set_source(const_mul1_node.out_port(0))
add1_node.in_port(0).get_connection().set_source(mul1_node.out_port(0))
add1_node.in_port(1).get_connection().set_source(const_add1_node.out_port(0))
mul2_node.in_port(0).get_connection().set_source(add1_node.out_port(0))
gamma.get_connection().set_destination(mul2_node.in_port(1))
add2_node.in_port(0).get_connection().set_source(mul2_node.out_port(0))
beta.get_connection().set_destination(add2_node.in_port(1))
toutput.get_connection().set_source(add2_node.out_port(0))
def replace_op(self, graph: Graph, node: Node):
name = node.soft_get('name', node.id)
# create range of axes for MVN based on `start_axis` and rank of input
rank = Rank(graph, {'name': name + '/Rank'}).create_node()
rng = create_op_with_const_inputs(graph, Range, {
0: int64_array(2),
2: int64_array(1)
}, {
'name': name + '/Range',
'output_type': np.int64
})
mvn = MVN(
graph, {
'eps': node.epsilon,
'eps_mode': 'inside_sqrt',
'normalize_variance': 1,
'name': name + '/Ins_Norm/MVN_',
}).create_node()
node.in_port(0).get_connection().set_destination(mvn.in_port(0))
rng.out_port(0).connect(mvn.in_port(1))
mul = Mul(graph, {
'axis': 1,
'name': name + '/Ins_Norm/mul_'
}).create_node()
mvn.out_port(0).connect(mul.in_port(0))
node.in_port(1).get_connection().set_destination(mul.in_port(1))
add = Add(graph, {
'axis': 1,
'name': name + '/Ins_Norm/add_'
}).create_node()
mul.out_port(0).connect(add.in_port(0))
node.in_port(2).get_connection().set_destination(add.in_port(1))
mvn.in_port(0).get_connection().add_destination(rank.in_port(0))
rng.in_port(1).connect(rank.out_port(0))
rename_nodes([(node, name + '/TBD'), (add, name)])
return [add.id]
def replace_pattern(graph: Graph, match: dict):
node = match['normalize']
# rename normalize node since it will be no longer output node after the transformation
output_name = node.soft_get('name', node.id)
normalizel2_name = output_name + '/normalizel2'
rename_node(node, normalizel2_name)
assert node.in_port(0).data.get_shape().size in [2, 3, 4]
assert node.has_valid('across_spatial')
assert node.has_valid('channel_shared')
assert node.has_valid('eps')
if 'bin' in node.in_edge(1):
del node.in_edge(1)['bin']
weights = node.in_port(1).data.get_value()
assert weights is not None
# in the code below we intentionally use get_source() to get the out port. Because updating the out port will
# update the Const node 'value' and 'shape' attributes
if node.channel_shared or all(weights == weights[0]):
node.in_port(1).get_source().data.set_value(np.array([weights[0]]))
else:
new_shape = np.ones((len(node.in_port(0).data.get_shape())),
dtype=np.int64)
new_shape[1] = -1
node.in_port(1).get_source().data.set_value(
np.array(weights).reshape(new_shape))
mul = Mul(graph, {'name': output_name}).create_node()
rename_node(mul, output_name)
if not node.across_spatial:
axes = int64_array([1])
else:
axes = int64_array(
np.arange(start=1, stop=node.in_port(0).data.get_shape().size))
normalizel2 = create_op_with_const_inputs(graph, NormalizeL2Op,
{1: axes}, {
'eps_mode': 'add',
'eps': node.eps
})
node.out_port(0).get_connection().set_source(mul.out_port(0))
node.in_port(1).get_connection().get_source().connect(mul.in_port(1))
normalizel2.out_port(0).connect(mul.in_port(0))
node.in_port(0).get_connection().set_destination(
normalizel2.in_port(0))
def find_and_replace_pattern(self, graph: Graph):
for node in graph.get_op_nodes(op='ThresholdedRelu'):
name = node.soft_get('name', node.id)
greater = create_op_with_const_inputs(
graph, Greater, {1: float_array([node.alpha])})
greater.in_port(0).connect(node.in_port(0).get_source())
float_greater = Cast(
graph, {
'dst_type':
data_type_str_to_np(graph.graph['cmd_params'].data_type)
}).create_node()
greater.out_port(0).connect(float_greater.in_port(0))
mul = Mul(graph, {}).create_node()
node.out_port(0).get_connection().set_source(mul.out_port(0))
mul.in_port(0).connect(node.in_port(0).get_source())
mul.in_port(1).connect(float_greater.out_port(0))
rename_nodes([(node, name + '/TBR'), (mul, name)])
graph.remove_node(node.id)
def replace_sub_graph(self, graph: Graph, match: dict):
# This replacer replace ImageScalar operation to Mul->Add sequence
# Also it check that weights and biases are good
op = match['op']
# Check that weights and biases are not useless
has_bias, has_weights = True, True
if all([x == 1 for x in np.nditer(op.scale)]):
has_weights = False
if all([x == 0 for x in np.nditer(op.bias)]):
has_bias = False
assert len(op.in_ports()) == 1
last_port = op.in_port(0).get_source()
# Create Mul & Add nodes
if has_weights:
mul_weights = Const(graph,
dict(value=op.scale,
shape=op.scale.shape)).create_node()
mul_op = Mul(graph, dict(name=op.id + '/mul_')).create_node()
op.in_port(0).get_connection().set_destination(mul_op.in_port(0))
mul_weights.out_port(0).connect(mul_op.in_port(1))
last_port = mul_op.out_port(0)
if has_bias:
add_bias = Const(graph, dict(value=op.bias,
shape=op.bias.shape)).create_node()
add_op = Add(graph, dict(name=op.id + '/add_')).create_node()
last_port.get_connection().set_destination(add_op.in_port(0))
add_bias.out_port(0).connect(add_op.in_port(1))
last_port = add_op.out_port(0)
op.in_port(0).disconnect()
op.out_port(0).get_connection().set_source(last_port)
def replace_sub_graph(self, graph: Graph, match: Dict[str, Node]):
node = match['op']
name = node.name
min_port_tuple = (node.in_port(1).get_source().node,
node.in_port(1).get_source().idx)
max_port_tuple = (node.in_port(2).get_source().node,
node.in_port(2).get_source().idx)
node.in_port(1).disconnect()
node.in_port(2).disconnect()
# make sure min < max
min_less_max = Less(graph, {
'name': name + '/if_min_less_max'
}).create_node([min_port_tuple, max_port_tuple])
minimum = Select(graph, {
'name': name + '/minimum'
}).create_node([min_less_max, min_port_tuple, max_port_tuple])
maximum = Select(graph, {
'name': name + '/maximum'
}).create_node([min_less_max, max_port_tuple, min_port_tuple])
# to create zero of limits data type, we multiply it by integer zero
zero = create_op_node_with_second_input(graph,
Mul,
int64_array(0),
{'name': name + '/zero'},
input_node=minimum)
# if 0 < min < max: min_adj = 0 and max_adj = max - min
min_greater_zero = Greater(graph, {
'name': name + '/if_minimum_greater_zero'
}).create_node([minimum, zero])
max_minus_min = Sub(graph, {
'name': name + '/max_minus_min'
}).create_node([maximum, minimum])
minimum = Select(graph, {
'name': name + '/first_adj_min'
}).create_node([min_greater_zero, zero, minimum])
maximum = Select(graph, {
'name': name + '/first_adj_max'
}).create_node([min_greater_zero, max_minus_min, maximum])
# if min < max < 0: min_adj = min - max and max_adj = 0
max_less_zero = Less(graph, {
'name': name + '/if_max_less_zero'
}).create_node([maximum, zero])
min_minus_max = Sub(graph, {
'name': name + '/min_minus_max'
}).create_node([minimum, maximum])
minimum = Select(graph, {
'name': name + '/second_adj_min'
}).create_node([max_less_zero, min_minus_max, minimum])
maximum = Select(graph, {
'name': name + '/second_adj_max'
}).create_node([max_less_zero, zero, maximum])
# scale = (max - min) / (2 ^ num_bits - 1),
float_range = Sub(graph, {
'name': name + '/float_range'
}).create_node([maximum, minimum])
quant_min_value, quant_max_value = int(
node.narrow_range), 2**node.num_bits - 1
int_range = Const(
graph,
dict(name=name + '/int_range',
value=quant_max_value - quant_min_value)).create_node()
scale = Div(graph, {
'name': name + '/scale'
}).create_node([float_range, int_range])
# min_adj = scale * round(min / scale)
descaled_min = Div(graph, {
'name': name + '/descaled_min'
}).create_node([minimum, scale])
rounded_descaled_min = Round(graph, {
'name': name + '/rounded_descaled_min'
}).create_node([descaled_min])
min_adj = Mul(graph, {
'name': name + '/min_adj'
}).create_node([scale, rounded_descaled_min])
# max_adj = max + min_adj - min.
adjustment = Sub(graph, {
'name': name + '/limits_adjustment'
}).create_node([min_adj, minimum])
max_adj = Add(graph, {
'name': name + '/max_adj'
}).create_node([maximum, adjustment])
# FakeQuantize operation has 5 inputs instead of 3 inputs in TensorFlow
node.add_input_port(3, skip_if_exist=True)
node.add_input_port(4, skip_if_exist=True)
node.in_port(1).connect(min_adj.out_port(0))
node.in_port(2).connect(max_adj.out_port(0))
node.in_port(3).connect(min_adj.out_port(0))
node.in_port(4).connect(max_adj.out_port(0))
FakeQuantize.update_node_stat(node, {'levels': node['levels']})
def replace_sub_graph(self, graph: Graph, match: dict):
node = match['op']
if 1 not in node.in_ports() or node.in_port(1).disconnected():
if node.has_valid('factor') and not node.has_valid('width') and not node.has_valid('height'):
factor = Const(graph, {'value': np.array(node.factor)}).create_node()
shape = Shape(graph, {'name': node.name + '/shape'}).create_node()
begin = Const(graph, {'value': np.array([2])}).create_node()
end = Const(graph, {'value': np.array([4])}).create_node()
stride = Const(graph, {'value': np.array([1])}).create_node()
ss = StridedSlice(graph, {'name': node.name + '/ss_0_port', 'begin_mask': np.array([1]),
'end_mask': np.array([0]), 'new_axis_mask': np.array([0]),
'shrink_axis_mask': np.array([0]),
'ellipsis_mask': np.array([0])}).create_node()
mul = Mul(graph, {'name': node.name + '/factor_mul_'}).create_node()
source = node.in_port(0).get_connection().get_source()
source.connect(shape.in_port(0))
shape.out_port(0).connect(ss.in_port(0))
begin.out_port(0).connect(ss.in_port(1))
end.out_port(0).connect(ss.in_port(2))
stride.out_port(0).connect(ss.in_port(3))
ss.out_port(0).connect(mul.in_port(0))
factor.out_port(0).connect(mul.in_port(1))
node.add_input_port(1, skip_if_exist=True)
assert node.in_port(1).disconnected()
mul.out_port(0).connect(node.in_port(1))
else:
shape = Shape(graph, {'name': node.name + '/shape'}).create_node()
begin = Const(graph, {'value': np.array([2])}).create_node()
end = Const(graph, {'value': np.array([4])}).create_node()
stride = Const(graph, {'value': np.array([1])}).create_node()
ss = StridedSlice(graph, {'name': node.name + '/ss_0_port', 'begin_mask': np.array([1]),
'end_mask': np.array([0]), 'new_axis_mask': np.array([0]),
'shrink_axis_mask': np.array([0]),
'ellipsis_mask': np.array([0])}).create_node()
source = node.in_port(0).get_connection().get_source()
source.connect(shape.in_port(0))
shape.out_port(0).connect(ss.in_port(0))
begin.out_port(0).connect(ss.in_port(1))
end.out_port(0).connect(ss.in_port(2))
stride.out_port(0).connect(ss.in_port(3))
pads_value = node.pads_begin + node.pads_end
pads_const = Const(graph, {'value': np.array(pads_value)}).create_node()
add = Add(graph, {'name': node.name + '/pad_add'}).create_node()
ss.out_port(0).connect(add.in_port(0))
add.in_port(1).connect(pads_const.out_port(0))
if node.soft_get('shrink_factor') != 1 and node.soft_get('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
const = Const(graph, {'name': node.name + '/pre_shrink_sub_const',
'value': np.array(-1)}).create_node()
sub = Add(graph, {'name': node.name + '/pre_shrink_sub'}).create_node()
add.out_port(0).connect(sub.in_port(0))
sub.in_port(1).connect(const.out_port(0))
const = Const(graph, {'value': np.array(1 / shrink_factor),
'name': node.name + 'shrink_factor_div_const'}).create_node()
div = Mul(graph, {'name': node.name + 'shrink_factor_div'}).create_node()
sub.out_port(0).connect(div.in_port(0))
div.in_port(1).connect(const.out_port(0))
const = Const(graph, {'name': node.name + '/shrink_factor_add_one_const', 'value': np.array(1)
}).create_node()
add = Add(graph, {'name': node.name + '/shrink_factor_add_one'}).create_node()
div.out_port(0).connect(add.in_port(0))
const.out_port(0).connect(add.in_port(1))
node.add_input_port(1, skip_if_exist=True)
assert node.in_port(1).disconnected()
add.out_port(0).connect(node.in_port(1))
elif node.soft_get('shrink_factor') == 1 and node.soft_get('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'] = 'Interpolate layer replacer may be wrong, please, try to update it in the' \
' file (openvino/tools/mo/front/InterpolateNormalizer.py 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
# const = Const(graph, {'value': np.array(-1),
# 'name': node.name + 'zoom_factor_deeplab-caffe_sub_const'}).create_node()
# sub = Add(graph, {'name': node.name + 'zoom_factor_deeplab-caffe_sub'}).create_node()
# add.out_port(0).connect(sub.in_port(0))
# const.out_port(0).connect(sub.in_port(1))
#
# const = Const(graph, {'value': np.array(zoom_factor - 1),
# 'name': node.name + 'zoom_factor_deeplab-caffe_mul_const'}).create_node()
# mul = Mul(graph, {'name': node.name + 'zoom_factor_deeplab-caffe_mul'}).create_node()
# sub.out_port(0).connect(mul.in_port(0))
# const.out_port(0).connect(mul.in_port(1))
#
# sum = Add(graph, {'name': node.name + 'zoom_factor_deeplab-caffe_sum'}).create_node()
# add.out_port(0).connect(sum.in_port(0))
# mul.out_port(0).connect(sum.in_port(1))
#
# node.add_input_port(1, skip_if_exist=True)
# assert node.in_port(1).disconnected()
# sum.out_port(0).connect(node.in_port(1))
# Comment out the following lines if you use the reshape method from previous section
const = Const(graph, {'value': np.array(zoom_factor),
'name': node.name + '/zoom_factor_mul_const'}).create_node()
mul = Mul(graph, {'name': node.name + '/zoom_factor_mul'}).create_node()
add.out_port(0).connect(mul.in_port(0))
const.out_port(0).connect(mul.in_port(1))
node.add_input_port(1, skip_if_exist=True)
assert node.in_port(1).disconnected()
mul.out_port(0).connect(node.in_port(1))
elif node.soft_get('width') != 0 and node.soft_get('height') != 0:
const = Const(graph, {'value': np.array([node.height, node.width])}).create_node()
node.add_input_port(1, skip_if_exist=True)
assert node.in_port(1).disconnected()
const.out_port(0).connect(node.in_port(1))
elif node.soft_get('shrink_factor') != 1 and node.soft_get('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
const = Const(graph, {'value': np.array(-1)}).create_node()
sub = Add(graph, {'name': node.name + '/shrink_zoom_factor_sub'}).create_node()
add.out_port(0).connect(sub.in_port(0))
const.out_port(0).connect(sub.in_port(1))
const = Const(graph, {'value': np.array(1 / (shrink_factor + 1))}).create_node()
div = Mul(graph, {'name': node.name + '/shrink_factor_div'}).create_node()
sub.out_port(0).connect(div.in_port(0))
const.out_port(0).connect(div.in_port(1))
const = Const(graph, {'value': np.array(-1),
'name': node.name + 'shrink_zoom_factor_sum_const'}).create_node()
sum = Add(graph, {'name': node.name + '/shrink_zoom_factor_sum'}).create_node()
div.out_port(0).connect(sum.in_port(0))
const.out_port(0).connect(sum.in_port(1))
const = Const(graph, {'value': np.array(zoom_factor - 1)}).create_node()
mul = Mul(graph, {'name': node.name + '/zoom_factor_mul'}).create_node()
sum.out_port(0).connect(mul.in_port(0))
const.out_port(0).connect(mul.in_port(1))
sum = Add(graph, {'name': node.name + '/final_shrink_zoom_factor_sum'}).create_node()
div.out_port(0).connect(sum.in_port(0))
mul.out_port(0).connect(sum.in_port(1))
node.add_input_port(1, skip_if_exist=True)
assert node.in_port(1).disconnected()
sum.out_port(0).connect(node.in_port(1))
else:
if node.soft_get('fw') == 'caffe':
shape = Shape(graph, {'name': node.name + '/shape'}).create_node()
begin = Const(graph, {'value': np.array([2])}).create_node()
end = Const(graph, {'value': np.array([4])}).create_node()
stride = Const(graph, {'value': np.array([1])}).create_node()
ss = StridedSlice(graph, {'name': node.name + '/ss_0_port', 'begin_mask': np.array([1]),
'end_mask': np.array([0]), 'new_axis_mask': np.array([0]),
'shrink_axis_mask': np.array([0]),
'ellipsis_mask': np.array([0])}).create_node()
source = node.in_port(1).get_connection().get_source()
node.in_port(1).disconnect()
source.connect(shape.in_port(0))
shape.out_port(0).connect(ss.in_port(0))
begin.out_port(0).connect(ss.in_port(1))
end.out_port(0).connect(ss.in_port(2))
stride.out_port(0).connect(ss.in_port(3))
ss.out_port(0).connect(node.in_port(1))
def replace_pattern(self, graph: Graph, match: Dict[str, Node]):
group_norm_node = match['op']
group_norm_num_input_dims = len(group_norm_node.in_port(0).data.get_shape())
# node computing initial GroupNorm input shape
initial_shape_op_node = Shape(graph, {'name': group_norm_node.name + '/Shape'}).create_node()
initial_shape_op_node.in_port(0).connect(group_norm_node.in_port(0).get_source())
initial_shape_op_node_float = Cast(
graph, {'name': initial_shape_op_node.name + '/to_float',
'dst_type': data_type_str_to_np(graph.graph['cmd_params'].data_type)}).create_node()
initial_shape_op_node.out_port(0).connect(initial_shape_op_node_float.in_port(0))
initial_batch_dim_node = node_to_get_batch_value(initial_shape_op_node_float)
initial_features_dim_node = node_to_get_features_dimension_value(initial_shape_op_node_float)
initial_spatial_dims_node_int = node_to_get_spatial_dimensions_value(initial_shape_op_node)
initial_spatial_dims_node = Cast(
graph, {'name': initial_spatial_dims_node_int.name + '/to_float',
'dst_type': data_type_str_to_np(graph.graph['cmd_params'].data_type)}).create_node()
initial_spatial_dims_node_int.out_port(0).connect(initial_spatial_dims_node.in_port(0))
group_size_node = Const(graph, {'value': int64_array([group_norm_node.num_groups]),
'name': group_norm_node.name + '/GroupSize'}).create_node()
# calculate "features // group_size" value
reciprocal_group_size_node = Const(graph, {'value': np.array([1.0 / group_norm_node.num_groups]),
'name': group_norm_node.name + '/ReciprocalGroupSize'}).create_node()
c_div_g_node = Mul(graph, {}).create_node()
c_div_g_node.in_port(0).connect(initial_features_dim_node.out_port(0))
c_div_g_node.in_port(1).connect(reciprocal_group_size_node.out_port(0))
batch_mul_group_size_node = Mul(graph, {}).create_node()
batch_mul_group_size_node.in_port(0).connect(initial_batch_dim_node.out_port(0))
batch_mul_group_size_node.in_port(1).connect(group_size_node.out_port(0))
# create new node which concatenates several dims to one
new_shape_node_float = new_shape_node_from_shape_nodes([batch_mul_group_size_node, c_div_g_node,
initial_spatial_dims_node])
new_shape_node = Cast(graph,
{'name': new_shape_node_float.name + '/to_int64', 'dst_type': np.int64}).create_node()
new_shape_node_float.out_port(0).connect(new_shape_node.in_port(0))
reshape_for_mvn_node = Reshape(graph, {}).create_node()
group_norm_node.in_port(0).get_connection().set_destination(reshape_for_mvn_node.in_port(0))
reshape_for_mvn_node.in_port(1).connect(new_shape_node.out_port(0))
# Reshape the gamma and beta constants to correct layout from [C] to [1,C], [1,C,1], [1,C,1,1] etc
gamma_beta_shape = np.ones([group_norm_num_input_dims], dtype=np.int64)
gamma_beta_shape[1] = -1
gamma_value = group_norm_node.in_port(1).get_source().data.get_value()
beta_value = group_norm_node.in_port(2).get_source().data.get_value()
assert gamma_value is not None, 'The gamma should be constant'
assert beta_value is not None, 'The beta should be constant'
gamma_value = np.reshape(gamma_value, gamma_beta_shape)
group_norm_node.in_port(1).get_source().data.set_value(gamma_value)
beta_value = np.reshape(beta_value, gamma_beta_shape)
group_norm_node.in_port(2).get_source().data.set_value(beta_value)
# MVN
mvn_node = MVN(graph, {'name': group_norm_node.name + '/MVN',
'normalize_variance': 1,
'eps': group_norm_node.eps,
'eps_mode': 'inside_sqrt'}).create_node()
mvn_node.in_port(0).connect(reshape_for_mvn_node.out_port(0))
# MVN axes
_, rank = get_shape_and_rank_nodes_by_port(mvn_node.in_port(0).get_connection().get_source(),
return_as_a_scalar=True)
rng = create_op_with_const_inputs(graph, Range, {0: int64_array(1), 2: int64_array(1)},
{'name': group_norm_node.name + '/Range', 'output_type': np.int64})
mvn_node.in_port(1).connect(rng.out_port(0))
rng.in_port(1).connect(rank.out_port(0))
# reshape to the initial shape before multiplying with gamma and adding beta
reshape_to_initial_shape_node = Reshape(graph, {}).create_node()
reshape_to_initial_shape_node.in_port(0).connect(mvn_node.out_port(0))
reshape_to_initial_shape_node.in_port(1).connect(initial_shape_op_node.out_port(0))
mul_node = Mul(graph, {'name': mvn_node.name + '/Mul'}).create_node()
mul_node.in_port(0).connect(reshape_to_initial_shape_node.out_port(0))
group_norm_node.in_port(1).get_connection().set_destination(mul_node.in_port(1))
add_node = Add(graph, {'name': mul_node.name + '/Add'}).create_node()
add_node.in_port(0).connect(mul_node.out_port(0))
group_norm_node.in_port(2).get_connection().set_destination(add_node.in_port(1))
group_norm_node.out_port(0).get_connection().set_source(add_node.out_port(0))
def replace_op(self, graph: Graph, node: Node):
input_out_port = node.in_port(0).get_source()
memory_pair_input = unique_id('id')
memory_pair_output = unique_id('id')
# Input -> FullyConnected
fc_layer_after_input_attrs = {
'name': 'input_fullyconnected',
'out-size': node.gifo_x_weights_shape[0],
'transpose_weights': True,
'bias_term': True,
}
fc_layer_after_input = FullyConnected(
graph, fc_layer_after_input_attrs).create_node()
fc_layer_after_input.in_port(0).connect(input_out_port)
input_as_const(fc_layer_after_input, fc_layer_after_input_attrs, 1,
'weights', node.gifo_x_weights)
input_as_const(fc_layer_after_input, fc_layer_after_input_attrs, 2,
'biases', node.gifo_biases)
init_value_prev_lstm_output = create_const_with_batch_from_input(
input_out_port, node.gifo_r_weights_shape[1])
prev_lstm_output = ReadValue(graph, {
'name': 'prev_memory_output',
'variable_id': memory_pair_input
}).create_node()
prev_lstm_output.in_port(0).connect(
init_value_prev_lstm_output.out_port(0))
# *Memory(output) -> FullyConnected
fc_layer_from_prev_state_attrs = {
'name': 'prev_memory_output_fullyconnected',
'out-size': node.gifo_r_weights_shape[0],
'transpose_weights': True,
'bias_term': False,
}
fc_layer_from_prev_state = FullyConnected(
graph, fc_layer_from_prev_state_attrs).create_node()
fc_layer_from_prev_state.in_port(0).connect(
prev_lstm_output.out_port(0))
input_as_const(fc_layer_from_prev_state,
fc_layer_from_prev_state_attrs, 1, 'weights',
node.gifo_r_weights)
# Memory -> FullyConnected \
# *Eltwise(sum)
# Input -> FullyConnected /
join_input_prev_state_sum = Add(graph, {
'name': 'join_input_eltwise'
}).create_node()
join_input_prev_state_sum.in_port(0).connect(
fc_layer_from_prev_state.out_port(0))
join_input_prev_state_sum.in_port(1).connect(
fc_layer_after_input.out_port(0))
# *Eltwise(sum) -> Split
# it is split into 4 nodes: Act, Eltw*3
# the following order is mandatory
# ___Tanh
# /
# Split ---(2)Eltwise(sum)
# |\
# | \__(3)Eltwise(sum)
# |____(4)Eltwise(sum)
split_joined_input_axis = Const(graph, {
'value': np.int64(1)
}).create_node()
split_joined_input = Split(graph, {
'name': 'join_input_split',
'num_splits': 4,
'out_ports_count': 4
}).create_node()
split_joined_input.in_port(0).connect(
join_input_prev_state_sum.out_port(0))
split_joined_input.in_port(1).connect(
split_joined_input_axis.out_port(0))
init_value_prev_lstm_state = create_const_with_batch_from_input(
split_joined_input.out_port(0), node.input_gate_weights.shape[0])
prev_lstm_state = ReadValue(graph, {
'name': 'prev_memory_state',
'variable_id': memory_pair_output
}).create_node()
prev_lstm_state.in_port(0).connect(
init_value_prev_lstm_state.out_port(0))
# *Memory(state) -> *ScaleShift(input)
state_input_scaleshift_attrs = {
'name': 'input_scaleshift',
'bias_term': False
}
state_input_scaleshift = ScaleShiftOp(
graph, state_input_scaleshift_attrs).create_node()
state_input_scaleshift.in_port(0).connect(prev_lstm_state.out_port(0))
input_as_const(state_input_scaleshift, state_input_scaleshift_attrs, 1,
'weights', node.input_gate_weights)
# *Memory(state) -> *ScaleShift(forget)
state_forget_scaleshift_attrs = {
'name': 'forget_scaleshift',
'bias_term': False
}
state_forget_scaleshift = ScaleShiftOp(
graph, state_forget_scaleshift_attrs).create_node()
state_forget_scaleshift.in_port(0).connect(prev_lstm_state.out_port(0))
input_as_const(state_forget_scaleshift, state_forget_scaleshift_attrs,
1, 'weights', node.forget_gate_weights)
# Split \
# (2)Eltwise(sum)
# Memory(state) -> *ScaleShift(input) /
join_prev_lstm_input_joined_input_sum = Add(
graph, {
'name': 'join_prev_lstm_input_joined_input_eltwise'
}).create_node()
join_prev_lstm_input_joined_input_sum.in_port(0).connect(
split_joined_input.out_port(1))
join_prev_lstm_input_joined_input_sum.in_port(1).connect(
state_input_scaleshift.out_port(0))
# Split \
# (3)Eltwise(sum)
# Memory(state) -> *ScaleShift(forget) /
join_prev_lstm_input_joined_forget_sum = Add(
graph, {
'name': 'join_prev_lstm_input_joined_forget_sum',
}).create_node()
join_prev_lstm_input_joined_forget_sum.in_port(0).connect(
split_joined_input.out_port(2))
join_prev_lstm_input_joined_forget_sum.in_port(1).connect(
state_forget_scaleshift.out_port(0))
# Split -> Tanh
remember_tahn = Tanh(graph, {'name': 'remember_tahnv'}).create_node()
remember_tahn.in_port(0).connect(split_joined_input.out_port(0))
# Split -> (2)Eltwise(sum) -> *Sigmoid
remember_sigmoid = Sigmoid(graph, {
'name': 'remember_sigmoid'
}).create_node()
remember_sigmoid.in_port(0).connect(
join_prev_lstm_input_joined_input_sum.out_port(0))
# Split -> (3)Eltwise(sum) -> **Sigmoid
forget_sigmoid = Sigmoid(graph, {
'name': 'forget_sigmoid'
}).create_node()
forget_sigmoid.in_port(0).connect(
join_prev_lstm_input_joined_forget_sum.out_port(0))
# *Memory(state) \
# (6)Eltwise(mul)
# Split -> (3)Eltwise(sum) -> **Sigmoid /
join_forget_prev_state_mul = Mul(graph, {
'name': 'join_forget_prev_state_mul'
}).create_node()
join_forget_prev_state_mul.in_port(0).connect(
forget_sigmoid.out_port(0))
join_forget_prev_state_mul.in_port(1).connect(
prev_lstm_state.out_port(0))
# Split -> Tahn \
# (5)Eltwise(mul)
# Split -> (2)Eltwise(sum) -> *Sigmoid /
join_remember_candidates_mul = Mul(
graph, {
'name': 'join_remember_candidates_mul'
}).create_node()
join_remember_candidates_mul.in_port(0).connect(
remember_tahn.out_port(0))
join_remember_candidates_mul.in_port(1).connect(
remember_sigmoid.out_port(0))
# (5)Eltwise(mul) \
# (7)Eltwise(sum)
# (6)Eltwise(mul) /
join_forget_remember_sum = Add(graph, {
'name': 'join_forget_remember_sum'
}).create_node()
join_forget_remember_sum.in_port(0).connect(
join_forget_prev_state_mul.out_port(0))
join_forget_remember_sum.in_port(1).connect(
join_remember_candidates_mul.out_port(0))
# (7)Eltwise(sum) -> Clamp
join_forget_clamp = create_op_with_const_inputs(
graph, Clamp, {
1: np.array(-node.clip_value, dtype=np.float32),
2: np.array(node.clip_value, dtype=np.float32)
}, {'name': 'join_forget_clamp'}, join_forget_remember_sum)
#
# Clamp -> (2)Memory(state)
next_lstm_state = Assign(graph, {
'name': 'next_lstm_state',
'variable_id': memory_pair_output
}).create_node()
next_lstm_state.in_port(0).connect(join_forget_clamp.out_port(0))
res_node = Result(graph, {'name': 'next_lstm_state_out'}).create_node()
res_node.in_port(0).connect(next_lstm_state.out_port(0))
# Clamp -> (2)Tahn
state_filtered_tahn = Tanh(graph, {
'name': 'state_filtered_tahn'
}).create_node()
state_filtered_tahn.in_port(0).connect(join_forget_clamp.out_port(0))
# Clamp -> (2)ScaleShift
clamp_scaleshift_attrs = {
'name': 'clamp_scaleshift',
'bias_term': False
}
clamp_scaleshift = ScaleShiftOp(graph,
clamp_scaleshift_attrs).create_node()
clamp_scaleshift.in_port(0).connect(join_forget_clamp.out_port(0))
input_as_const(clamp_scaleshift, clamp_scaleshift_attrs, 1, 'weights',
node.output_gate_weights)
# Split \
# (4)Eltwise(sum)
# Clamp -> (2)ScaleShift /
join_next_lstm_input_joined_input_sum = Add(
graph, {
'name': 'join_next_lstm_input_joined_input_sum',
}).create_node()
join_next_lstm_input_joined_input_sum.in_port(0).connect(
split_joined_input.out_port(3))
join_next_lstm_input_joined_input_sum.in_port(1).connect(
clamp_scaleshift.out_port(0))
# (4)Eltwise(sum) -> (3)Sigmoid
output_sigmoid = Sigmoid(graph, {
'name': 'output_sigmoid'
}).create_node()
output_sigmoid.in_port(0).connect(
join_next_lstm_input_joined_input_sum.out_port(0))
# (4)Eltwise(sum) -> (3)Sigmoid \
# (5)Eltwise(mul)
# Clamp -> (2)Tahn /
joined_output_mul = Mul(graph, {
'name': 'joined_output_mul'
}).create_node()
joined_output_mul.in_port(0).connect(state_filtered_tahn.out_port(0))
joined_output_mul.in_port(1).connect(output_sigmoid.out_port(0))
# (5)Eltwise(mul) -> (3)FullyConnected
fc_output_attrs = {
'name': 'FullyConnected',
'out-size': node.projection_weights_shape[0],
'transpose_weights': True,
'bias_term': False
}
fc_output = FullyConnected(graph, fc_output_attrs).create_node()
fc_output.in_port(0).connect(joined_output_mul.out_port(0))
input_as_const(fc_output, fc_output_attrs, 1, 'weights',
node.projection_weights)
# / (2)Memory(output)
# (3)FullyConnected
# \ Output (any next node) (edge created automatically after replacement)
next_lstm_output = Assign(graph, {
'name': 'next_lstm_output',
'variable_id': memory_pair_input
}).create_node()
next_lstm_output.in_port(0).connect(fc_output.out_port(0))
res_node_lstm_output = Result(graph, {
'name': 'next_lstm_output_out'
}).create_node()
res_node_lstm_output.in_port(0).connect(next_lstm_output.out_port(0))
return [fc_output.id]
def replace_pattern(graph: Graph, match: dict):
log.debug(
'================== GNMTBeforeConditionFind ==================')
input_sequence_lengths = match['Max'].in_port(0).get_source().node
encoder_sequence_lengths = looking_for_op_in_list([
port.node
for port in input_sequence_lengths.out_port(0).get_destinations()
], 'Identity')
# Looking for Sequence_length node in encoder looks like:
# Sequence_length -> CheckSeqLen -> Max -> Maximum -> Minimum
check_seq_len = looking_for_op_in_list([
port.node for port in encoder_sequence_lengths.out_port(
0).get_destinations()
], 'Identity')
max = looking_for_op_in_list([
port.node for port in check_seq_len.out_port(0).get_destinations()
], 'ReduceMax')
maximum = max.out_port(0).get_destinations()[0].node
assert maximum.op == 'Maximum'
minimum = maximum.out_port(0).get_destinations()[0].node
assert minimum.op == 'Minimum'
tensor_seq_len = looking_for_op_in_list([
minimum.in_port(port).get_source().node
for port in minimum.in_ports()
], 'StridedSlice')
# Create node for multiplying seq_len by 2
const = Const(graph, {
'name': 'FakeSeqLenMultiplyer',
'value': mo_array(2)
}).create_node()
mul_op = Mul(graph, {'name': 'FakeSeqLen'}).create_node()
const.out_port(0).get_connection().set_destination(mul_op.in_port(1))
tensor_seq_len.out_port(0).get_connection().add_destination(
mul_op.in_port(0))
# Connect seq_len * 2 to TensorArray from GNMT loop
ta_writes = [
port.node
for port in match['Identity_1'].out_port(0).get_destinations()
if port.node.op == 'TensorArrayWriteV3'
]
for ta_write in ta_writes:
ta = ta_write.in_port(0).get_source().node.in_port(
0).get_source().node
ta.in_port(0).disconnect()
ta.in_port(0).get_connection().set_source(mul_op.out_port(0))
if not graph.graph['cmd_params'].static_shape:
log.error(
"Model can not be translated in a reshape-able way.\n"
"Model Optimizer key static_shape was turned on to prevent related errors.\n"
"There will be no success changing input shapes of the model with the help of "
"InferenceEngine reshape method",
extra={'is_warning': True})
graph.graph['cmd_params'].static_shape = True
def replace_resize(graph: Graph, resize: Node):
log.debug("Converting of ONNX Resize-11 to Interpolate-4 "
"is triggered for node {}.".format(
resize.soft_get('name', resize.id)))
input_shape = resize.in_port(0).data.get_shape()
input_rank = len(input_shape)
resize_name = resize.soft_get('name', resize.id)
if input_rank not in {4, 5}:
log.warning(
'The input shape is not 4D or 5D for op with name {}'.format(
resize_name))
return
assert (resize.is_in_port_connected(0) and (resize.is_in_port_connected(2) or resize.is_in_port_connected(3))), \
"Scales or sizes inputs must be connected to Node {} with op {}.".format(resize.soft_get("name", resize.id),
resize.op)
assert resize.soft_get('coordinate_transformation_mode') != 'tf_crop_and_resize', \
'Mode tf_crop_and_resize is not supported for op {} with name {}'.format(resize.op,
resize.soft_get("name", resize.id))
layout = graph.graph['layout']
if input_rank == 4:
begin_dim = get_height_dim(layout, input_rank)
end_dim = get_width_dim(layout, input_rank) + 1
else:
begin_dim = get_depth_dim(layout, input_rank)
end_dim = get_width_dim(layout, input_rank) + 1
sizes_ss = create_op_with_const_inputs(
graph, StridedSlice, {
1: int64_array([begin_dim]),
2: int64_array([end_dim]),
3: int64_array([1])
}, {
'name': resize_name + '/StridedSlice_sizes',
'begin_mask': int64_array([1]),
'end_mask': int64_array([1]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
scales_ss = create_op_with_const_inputs(
graph, StridedSlice, {
1: int64_array([begin_dim]),
2: int64_array([end_dim]),
3: int64_array([1])
}, {
'name': resize_name + '/StridedSlice_scales',
'begin_mask': int64_array([1]),
'end_mask': int64_array([1]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
axes_node = Const(
graph, {
'name': resize_name + '/axis',
'value': int64_array(np.arange(begin_dim, end_dim))
}).create_node()
shape_calculation_mode = 'sizes' if resize.is_in_port_connected(
3) else 'scales'
interpolate_node = Interpolate(
graph, {
'version': 'opset4',
'mode': convert_mode(resize.mode),
'coordinate_transformation_mode':
resize.coordinate_transformation_mode,
'cube_coeff': resize.cube_coeff,
'nearest_mode': resize.nearest_mode,
'pads_begin': int64_array([0]),
'pads_end': int64_array([0]),
'antialias': 0,
'shape_calculation_mode': shape_calculation_mode,
'in_ports_count': 4
}).create_node()
axes_node.out_port(0).connect(interpolate_node.in_port(3))
shape_of = Shape(graph, {'name': resize_name + '/ShapeOf'}).create_node()
add_node = create_op_with_const_inputs(graph, Add,
{1: float_array([1.0e-5])},
{'name': resize_name + '/Add'})
dst_dtype = np.float32 # even if data_type=FP16 use float32 for shape values
if not resize.is_in_port_connected(3):
cast_shape_to_float = Cast(graph, {
'dst_type': dst_dtype
}).create_node()
mul_node = Mul(graph, {'name': resize_name + '/Mul'}).create_node()
shape_of.out_port(0).connect(cast_shape_to_float.in_port(0))
cast_shape_to_float.out_port(0).connect(mul_node.in_port(0))
cast_add_result_to_int = Cast(graph, {
'dst_type': np.int64
}).create_node()
floor_node = Floor(graph, {
'name': resize_name + '/Floor'
}).create_node()
mul_node.out_port(0).connect(add_node.in_port(0))
add_node.out_port(0).connect(floor_node.in_port(0))
floor_node.out_port(0).connect(cast_add_result_to_int.in_port(0))
cast_add_result_to_int.out_port(0).connect(sizes_ss.in_port(0))
sizes_ss.out_port(0).connect(interpolate_node.in_port(1))
scales_ss.out_port(0).connect(interpolate_node.in_port(2))
connection_of_resize_input = resize.in_port(0).get_connection()
connection_of_resize_input.set_destination(interpolate_node.in_port(0))
connection_of_scales = resize.in_port(2).get_connection()
connection_of_scales.set_destination(scales_ss.in_port(0))
connection_of_resize_input.get_source().connect(shape_of.in_port(0))
connection_of_scales.get_source().connect(mul_node.in_port(1))
else:
cast_shape_to_float = Cast(graph, {
'dst_type': dst_dtype
}).create_node()
cast_sizes_to_float = Cast(graph, {
'dst_type': dst_dtype
}).create_node()
div_node = Div(graph, {'name': resize_name + '/Div'}).create_node()
cast_sizes_to_float.out_port(0).connect(div_node.in_port(0))
cast_shape_to_float.out_port(0).connect(div_node.in_port(1))
shape_of.out_port(0).connect(cast_shape_to_float.in_port(0))
div_node.out_port(0).connect(add_node.in_port(0))
add_node.out_port(0).connect(scales_ss.in_port(0))
scales_ss.out_port(0).connect(interpolate_node.in_port(2))
sizes_ss.out_port(0).connect(interpolate_node.in_port(1))
connection_of_resize_input = resize.in_port(0).get_connection()
connection_of_resize_input.set_destination(interpolate_node.in_port(0))
connection_of_sizes = resize.in_port(3).get_connection()
connection_of_sizes.set_destination(sizes_ss.in_port(0))
connection_of_resize_input.get_source().connect(shape_of.in_port(0))
connection_of_sizes.get_source().connect(
cast_sizes_to_float.in_port(0))
rename_nodes([(resize, resize_name + '/delete'),
(interpolate_node, resize_name)])
resize.out_port(0).get_connection().set_source(
interpolate_node.out_port(0))
def dequantize_data(fake_quantize: Node, dst_type: type,
quantized_type: type) -> Node:
graph = fake_quantize.graph
quantized_data = fake_quantize.in_port(0).get_source().node
name = fake_quantize.soft_get('name', fake_quantize.id)
assert quantized_data.soft_get('type') == 'Convert' and quantized_data.dst_type == quantized_type, \
'Weights aren`t compressed as expected for node {}'.format(fake_quantize.soft_get('name', fake_quantize.id))
dequantizing_cast = Cast(
graph,
dict(name=quantized_data.name +
"/to_{}".format(np_data_type_to_destination_type(dst_type)),
dst_type=dst_type,
stop_value_propagation=True)).create_node()
fake_quantize.in_port(0).get_connection().set_destination(
dequantizing_cast.in_port(0))
# limits of dequantize
in_low = fake_quantize.in_port(1).get_source()
in_high = fake_quantize.in_port(2).get_source()
out_low = fake_quantize.in_port(3).get_source()
out_high = fake_quantize.in_port(4).get_source()
# scale calculation
output_range = Sub(graph, {
'name': name + '/output_range'
}).create_node()
output_range.in_port(0).connect(out_high)
output_range.in_port(1).connect(out_low)
input_range = Sub(graph, {'name': name + '/input_range'}).create_node()
input_range.in_port(0).connect(in_high)
input_range.in_port(1).connect(in_low)
scale = Div(graph, {'name': name + '/scale'}).create_node()
scale.in_port(0).connect(output_range.out_port(0))
scale.in_port(1).connect(input_range.out_port(0))
# shift calculation
descaled_output_low = Div(graph, {
'name': name + '/descaled_output_low'
}).create_node()
descaled_output_low.in_port(0).connect(out_low)
descaled_output_low.in_port(1).connect(scale.out_port(0))
shift = Sub(graph, {'name': name + '/shift'}).create_node()
shift.in_port(0).connect(in_low)
shift.in_port(1).connect(descaled_output_low.out_port(0))
zero = Const(graph, {
'name': name + '/zero',
'value': mo_array(0, dtype=dst_type)
}).create_node()
scale_eq_zero = Equal(graph, {
'name': name + '/scale_eq_zero'
}).create_node()
scale_eq_zero.in_port(0).connect(scale.out_port(0))
scale_eq_zero.in_port(1).connect(zero.out_port(0))
zero_point = Select(graph, {
'name': name + '/zero_point'
}).create_node()
zero_point.in_port(0).connect(scale_eq_zero.out_port(0))
zero_point.in_port(1).connect(zero.out_port(0))
zero_point.in_port(2).connect(shift.out_port(0))
# DeQuantize(x) == Mul(Sub(x, zero_point), scale)
sub_zp = Sub(graph, {'name': name + '/minus_zp'}).create_node()
sub_zp.in_port(0).connect(dequantizing_cast.out_port(0))
sub_zp.in_port(1).connect(zero_point.out_port(0))
mul_scale = Mul(graph, {
'name': name + '/mulpiply_by_scale'
}).create_node()
mul_scale.in_port(0).connect(sub_zp.out_port(0))
mul_scale.in_port(1).connect(scale.out_port(0))
fake_quantize.out_port(0).get_connection().set_source(
mul_scale.out_port(0))
graph.remove_nodes_from([fake_quantize.id, fake_quantize.out_node(0)])
def replace_op(self, graph: Graph, node: Node):
node_name = node.soft_get('name', node.id)
# check if we have dropout
input_port = node.in_port(0)
if node.has_and_set('use_dropout'):
split_dropout = AttributedVariadicSplit(graph,
{'name': node_name + '/split_dropout',
'size_splits': int64_array([-1, 1, 1, 1]),
'axis': int64_array(1)}).create_node()
input_port.get_connection().set_destination(split_dropout.in_port(0))
input_port = split_dropout.out_port(0)
i_drop_scale = split_dropout.out_port(1)
f_drop_scale = split_dropout.out_port(2)
o_drop_scale = split_dropout.out_port(3)
# split input to (i_part, f_part, c_part, o_part, ct_1)
split_node = create_op_with_const_inputs(graph, Split, {1: np.int64(1)},
{'name': node_name + '/split_lstm_input',
'num_splits': 5})
input_port.get_connection().set_destination(split_node.in_port(0))
i_part = split_node.out_port(0)
f_part = split_node.out_port(1)
c_part = split_node.out_port(2)
o_part = split_node.out_port(3)
ct_1 = split_node.out_port(4)
# i_t = Sigmoid(i_part + w_ic*ct_1)
i_scale_attrs = {'name': node_name + '/i_scaleshift',
'bias_term': False}
i_scale = ScaleShiftOp(graph, i_scale_attrs).create_node()
input_as_const(i_scale, i_scale_attrs, 1, 'weights', node.i_weights)
ct_1.connect(i_scale.in_port(0))
sum_i_c = Add(graph, {'name': node_name + '/sum_i_c_'}).create_node()
i_part.connect(sum_i_c.in_port(0))
i_scale.out_port(0).connect(sum_i_c.in_port(1))
i_sigmoid = Sigmoid(graph, {'name': node_name + '/i_sigmoid'}).create_node()
sum_i_c.out_port(0).connect(i_sigmoid.in_port(0))
if node['use_dropout']:
mul_dropout_i = Mul(graph, {'name': split_node.soft_get('name', split_node.id) + '/mul_i'}).create_node()
mul_dropout_i.in_port(0).connect(i_sigmoid.out_port(0))
mul_dropout_i.in_port(1).connect(i_drop_scale)
i_sigmoid = mul_dropout_i
# f_t = Sigmoid(f_part + w_fc*ct_1)
f_scale_attrs = {'name': node_name + '/f_scaleshift',
'bias_term': False}
f_scale = ScaleShiftOp(graph, f_scale_attrs).create_node()
input_as_const(f_scale, f_scale_attrs, 1, 'weights', node.f_weights)
ct_1.connect(f_scale.in_port(0))
sum_f_c = Add(graph, {'name': node_name + '/sum_f_c_'}).create_node()
f_part.connect(sum_f_c.in_port(0))
f_scale.out_port(0).connect(sum_f_c.in_port(1))
f_sigmoid = Sigmoid(graph, {'name': node_name + '/f_sigmoid'}).create_node()
sum_f_c.out_port(0).connect(f_sigmoid.in_port(0))
if node['use_dropout']:
mul_dropout_f = Mul(graph, {'name': split_node.soft_get('name', split_node.id) + '/mul_f'}).create_node()
mul_dropout_f.in_port(0).connect(f_sigmoid.out_port(0))
mul_dropout_f.in_port(1).connect(f_drop_scale)
f_sigmoid = mul_dropout_f
# c_t = f_t*ct_1 + i_t * tanh(c_part)
c_tanh = Tanh(graph, {'name': node_name + '/c_tanh'}).create_node()
c_part.connect(c_tanh.in_port(0))
prod_i_c_tanh = Mul(graph, {'name': node_name + '/prod_i_c_tanh_'}).create_node()
i_sigmoid.out_port(0).connect(prod_i_c_tanh.in_port(0))
c_tanh.out_port(0).connect(prod_i_c_tanh.in_port(1))
prod_f_ct_1 = Mul(graph, {'name': node_name + '/prod_f_ct_1_'}).create_node()
f_sigmoid.out_port(0).connect(prod_f_ct_1.in_port(0))
ct_1.connect(prod_f_ct_1.in_port(1))
sum_f_i = Add(graph, {'name': node_name + '/sum_f_i_'}).create_node()
prod_f_ct_1.out_port(0).connect(sum_f_i.in_port(0))
prod_i_c_tanh.out_port(0).connect(sum_f_i.in_port(1))
# o_t = Sigmoid(o_part + w_oc*c_t)
o_scale_attrs = {'name': node_name + '/o_scaleshift',
'bias_term': False}
o_scale = ScaleShiftOp(graph, o_scale_attrs).create_node()
input_as_const(o_scale, o_scale_attrs, 1, 'weights', node.o_weights)
sum_f_i.out_port(0).connect(o_scale.in_port(0))
sum_o_c = Add(graph, {'name': node_name + '/sum_o_c_'}).create_node()
o_part.connect(sum_o_c.in_port(0))
o_scale.out_port(0).connect(sum_o_c.in_port(1))
o_sigmoid = Sigmoid(graph, {'name': node_name + '/o_sigmoid'}).create_node()
sum_o_c.out_port(0).connect(o_sigmoid.in_port(0))
if node['use_dropout']:
mul_dropout_o = Mul(graph, {'name': split_node.soft_get('name', split_node.id) + '/mul_o'}).create_node()
mul_dropout_o.in_port(0).connect(o_sigmoid.out_port(0))
mul_dropout_o.in_port(1).connect(o_drop_scale)
o_sigmoid = mul_dropout_o
# m_t = o_t * Tanh(c_t)
c_t_tanh = Tanh(graph, {'name': node_name + '/c_t_tanh'}).create_node()
sum_f_i.out_port(0).connect(c_t_tanh.in_port(0))
prod_o_c_t_tanh = Mul(graph, {'name': node_name + '/prod_o_c_t_tanh_'}).create_node()
o_sigmoid.out_port(0).connect(prod_o_c_t_tanh.in_port(0))
c_t_tanh.out_port(0).connect(prod_o_c_t_tanh.in_port(1))
# add concat to create 1 output
concat = Concat(graph, {'name': node_name + '/concat_c_m'}).create_node()
concat.add_sequence_of_ports('in', range(2))
sum_f_i.out_port(0).connect(concat.in_port(0))
prod_o_c_t_tanh.out_port(0).connect(concat.in_port(1))
return [concat.id]