def extract(cls, node):
param = node.pb.inner_product_param
pb_model = node.model_pb
attrs = {
'out-size': param.num_output,
'transpose_weights': not param.transpose,
}
attrs.update(weights_biases(param.bias_term, pb_model))
FullyConnected.update_node_stat(node, attrs)
return cls.enabled
def extract(cls, node):
attr = get_mxnet_layer_attrs(node.symbol_dict)
num_hidden = attr.int('num_hidden', None)
assert num_hidden is not None, "{} node with no `num_hidden` parameter found".format(cls.op)
attrs = {
'out-size': num_hidden,
'transpose_weights': True,
}
FullyConnected.update_node_stat(node, attrs)
return cls.enabled
def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<Params>')
weights, weights_shape = read_binary_matrix(pb)
mapping_rule = {
'out-size': weights_shape[0],
'transpose_weights': True,
}
embed_input(mapping_rule, 1, 'weights', weights)
FullyConnected.update_node_stat(node, mapping_rule)
return cls.enabled
def extract(cls, node):
pb = node.parameters
read_learning_info(pb)
weights, weights_shape = read_binary_matrix(pb)
biases = read_binary_vector(pb)
mapping_rule = {
'out-size': weights_shape[0],
'transpose_weights': True,
}
embed_input(mapping_rule, 1, 'weights', weights)
embed_input(mapping_rule, 2, 'biases', biases)
FullyConnected.update_node_stat(node, mapping_rule)
return cls.enabled
def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<LinearParams>')
weights, weights_shape = read_binary_matrix(pb)
tag = find_next_tag(pb)
read_placeholder(pb, 1)
if tag != '<BiasParams>':
raise Error('FixedAffineComponent must contain BiasParams')
biases = read_binary_vector(pb)
mapping_rule = {
'out-size': weights_shape[0],
'transpose_weights': True,
}
embed_input(mapping_rule, 1, 'weights', weights)
embed_input(mapping_rule, 2, 'biases', biases)
FullyConnected.update_node_stat(node, mapping_rule)
return cls.enabled
def replace_op(self, graph: Graph, node: Node):
input_node = node.in_node()
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([input_node])
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)
prev_lstm_output = Memory(
graph, {
'name': 'prev_memory_output',
'id': memory_pair_input,
'index': 1,
'size': 2,
'shape': np.array([node.gifo_r_weights_shape[1]],
dtype=np.int64)
}).create_node()
# *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([prev_lstm_output])
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([fc_layer_from_prev_state, fc_layer_after_input])
# *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 = Split(
graph, {
'name': 'join_input_split',
'axis': 1,
'num_split': 4,
'out_ports_count': 4,
}).create_node([join_input_prev_state_sum])
prev_lstm_state = Memory(
graph, {
'name':
'prev_memory_state',
'id':
memory_pair_output,
'index':
1,
'size':
2,
'shape':
np.array([node.input_gate_weights.shape[0]], dtype=np.int64)
}).create_node()
# *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([prev_lstm_state])
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([prev_lstm_state])
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([(split_joined_input, 1), state_input_scaleshift])
# 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([(split_joined_input, 2), state_forget_scaleshift])
# Split -> Tanh
remember_tahn = Tanh(graph, {
'name': 'remember_tahnv'
}).create_node([(split_joined_input, 0)])
# Split -> (2)Eltwise(sum) -> *Sigmoid
remember_sigmoid = Sigmoid(graph, {
'name': 'remember_sigmoid'
}).create_node([join_prev_lstm_input_joined_input_sum])
# Split -> (3)Eltwise(sum) -> **Sigmoid
forget_sigmoid = Sigmoid(graph, {
'name': 'forget_sigmoid'
}).create_node([join_prev_lstm_input_joined_forget_sum])
# *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([forget_sigmoid, prev_lstm_state])
# Split -> Tahn \
# (5)Eltwise(mul)
# Split -> (2)Eltwise(sum) -> *Sigmoid /
join_remember_candidates_mul = Mul(
graph, {
'name': 'join_remember_candidates_mul',
}).create_node([remember_tahn, remember_sigmoid])
# (5)Eltwise(mul) \
# (7)Eltwise(sum)
# (6)Eltwise(mul) /
join_forget_remember_sum = Add(graph, {
'name': 'join_forget_remember_sum',
}).create_node(
[join_forget_prev_state_mul, join_remember_candidates_mul])
# (7)Eltwise(sum) -> Clamp
join_forget_clamp = Clamp(
graph, {
'name': 'join_forget_clamp',
'max': node.clip_value,
'min': -node.clip_value
}).create_node([join_forget_remember_sum])
#
# Clamp -> (2)Memory(state)
next_lstm_state = Memory(
graph, {
'name':
'next_lstm_state',
'id':
memory_pair_output,
'index':
0,
'size':
2,
'shape':
np.array([node.input_gate_weights.shape[0]], dtype=np.int64)
}).create_node([join_forget_clamp])
Result(graph, {
'name': 'next_lstm_state_out'
}).create_node([next_lstm_state])
# Clamp -> (2)Tahn
state_filtered_tahn = Tanh(graph, {
'name': 'state_filtered_tahn'
}).create_node([join_forget_clamp])
# Clamp -> (2)ScaleShift
clamp_scaleshift_attrs = {
'name': 'clamp_scaleshift',
'bias_term': False
}
clamp_scaleshift = ScaleShiftOp(
graph, clamp_scaleshift_attrs).create_node([join_forget_clamp])
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([(split_joined_input, 3), clamp_scaleshift])
# (4)Eltwise(sum) -> (3)Sigmoid
output_sigmoid = Sigmoid(graph, {
'name': 'output_sigmoid'
}).create_node([join_next_lstm_input_joined_input_sum])
# (4)Eltwise(sum) -> (3)Sigmoid \
# (5)Eltwise(mul)
# Clamp -> (2)Tahn /
joined_output_mul = Mul(graph, {
'name': 'joined_output_mul'
}).create_node([state_filtered_tahn, output_sigmoid])
# (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(
[joined_output_mul])
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 = Memory(
graph, {
'name': 'next_lstm_output',
'id': memory_pair_input,
'index': 0,
'size': 2,
'shape': np.array([node.gifo_r_weights_shape[1]],
dtype=np.int64)
}).create_node([fc_output])
Result(graph, {
'name': 'next_lstm_output_out'
}).create_node([next_lstm_output])
return [fc_output.id]
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_zero_value_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))
# prev_lstm_state = Memory(graph, {'name': 'prev_memory_state',
# 'id': memory_pair_output,
# 'index': 1,
# 'size': 2,
# 'shape': np.array([node.input_gate_weights.shape[0]], dtype=np.int64)
# }).create_node()
init_value_prev_lstm_state = create_zero_value_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):
node = match['matmul']
name = node.soft_get('name', node.id)
A_shape = node.in_port(0).data.get_shape()
B_shape = node.in_port(1).data.get_shape()
out_shape = node.out_port(0).data.get_shape()
assert A_shape is not None and B_shape is not None and out_shape is not None
B_value = node.in_port(1).data.get_value()
if (B_value is not None or node.in_port(1).get_source().node.has_and_set('stop_value_propagation')) and B_shape[
B_shape != 1].size <= 2:
# transferring from MatMul representation: [B, I, K] * [B, K, O] = [B, I, O]
# to FullyConnected representation: [I, K] * [O, K] = [I, O]
B, I, K, O, aligned_A_shape, aligned_B_shape = MatMulToFullyConnected.get_matmul_BIKO(node)
# weights normalization
if not node.transpose_b:
# FullyConnected weights layout is OI
# MatMul second input layout is (B)IO
transpose_order = list(range(B_shape.size))
transpose_order[-1], transpose_order[-2] = transpose_order[-2], transpose_order[-1]
order = Const(graph, {'value': int64_array(transpose_order)}).create_node()
transpose = Transpose(graph, {'name': name + '/weights_transpose'}).create_node()
weights_source = node.in_port(1).get_source()
node.in_port(1).get_connection().set_source(transpose.out_port(0))
transpose.in_port(0).connect(weights_source)
transpose.in_port(1).connect(order.out_port(0))
order.infer(order)
transpose.infer(transpose)
if node.in_port(1).data.get_shape().size != 2:
const = Const(graph, {'value': int64_array([-1, K])}).create_node()
reshape = Reshape(graph, {'name': name + '/weights_reshape'}).create_node()
weights_source = node.in_port(1).get_source()
node.in_port(1).get_connection().set_source(reshape.out_port(0))
reshape.in_port(0).connect(weights_source)
reshape.in_port(1).connect(const.out_port(0))
const.infer(const)
reshape.infer(reshape)
assert np.all(np.array_equal(node.in_port(1).data.get_shape(), int64_array([O, K]))), \
"MatMul `{}` was not converted to FullyConnected: wrong weights shape: {}, " \
"B={}, I={}, K={}, O={}".format(name, node.in_port(1).data.get_shape(), B, I, K, O)
node.in_port(1).bin = 'weights'
del node['transpose_b']
# input normalization
if node.transpose_a:
transpose_order = list(range(A_shape.size))
transpose_order[-1], transpose_order[-2] = transpose_order[-2], transpose_order[-1]
order = Const(graph, {'value': int64_array(transpose_order)}).create_node()
transpose = Transpose(graph, {'name': name + '/input_transpose'}).create_node()
input_source = node.in_port(0).get_source()
node.in_port(0).get_connection().set_source(transpose.out_port(0))
transpose.in_port(0).connect(input_source)
transpose.in_port(1).connect(order.out_port(0))
order.infer(order)
transpose.infer(transpose)
if A_shape.size != 2:
const = Const(graph, {'value': int64_array([-1, K])}).create_node()
reshape = Reshape(graph, {'name': name + '/input_reshape'}).create_node()
input_source = node.in_port(0).get_source()
node.in_port(0).get_connection().set_source(reshape.out_port(0))
reshape.in_port(0).connect(input_source)
reshape.in_port(1).connect(const.out_port(0))
const.infer(const)
reshape.infer(reshape)
assert np.all(np.array_equal(node.in_port(0).data.get_shape(), int64_array([np.prod(B) * I, K]))), \
"MatMul `{}` wasn't converted to FullyConnected: wrong input shape: {}, " \
"B={}, I={}, K={}, O={}".format(name, node.in_port(0).data.get_shape(), B, I, K, O)
del node['transpose_a']
FullyConnected.update_node_stat(node, {'out-size': O})
# output normalization
if out_shape.size != 2:
const = Const(graph, {'value': int64_array([*B, I, O])}).create_node()
reshape = Reshape(graph, {'name': name + '/output_reshape'}).create_node()
dst = node.out_port(0).get_destination()
node.out_port(0).get_connection().set_destination(reshape.in_port(0))
const.out_port(0).connect(reshape.in_port(1))
reshape.out_port(0).connect(dst)
node.infer(node)
const.infer(const)
reshape.infer(reshape)
else:
assert A_shape.size == out_shape.size
assert B_shape.size <= out_shape.size
if B_shape.size != out_shape.size:
unsqueeze_dim = Const(graph, {'value': int64_array(list(range(out_shape.size - B_shape.size)))
}).create_node()
unsqueeze = Unsqueeze(graph, {}).create_node()
B_source = node.in_port(1).get_source()
node.in_port(1).get_connection().set_source(unsqueeze.out_port(0))
unsqueeze.in_port(0).connect(B_source)
unsqueeze.in_port(1).connect(unsqueeze_dim.out_port(0))
unsqueeze_dim.infer(unsqueeze_dim)
unsqueeze.infer(unsqueeze)
Gemm.update_node_stat(node, {
'transpose_a': node.has_and_set('transpose_a'),
'transpose_b': node.has_and_set('transpose_b'),
})
