def extract(cls, node: Node):
pb = node.pb
Split.update_node_stat(node, {
'num_splits': pb.attr['num_split'].i,
'input_port': 1,
})
return cls.enabled
def split_data(self, data: Node):
""" Helper. Split data node into two part along 0 axis """
assert len(data.shape) == 3
assert data.shape[0] == 2
output_data = [Op._create_data_node(data.graph,
name=data.name + '/SplittedBiLSTM/{}'.format(['forward', 'reverse'][i])) for
i in [0, 1]]
split_op = Split(data.graph, dict(name=data.name + '/DecomposedBiLSTM_0', num_splits=2))
axis_const = Const(data.graph, {'name': data.name + '/DecomposedBiLSTM_0' + '/Split_axis',
'value': np.int64(0)}).create_node_with_data()
return split_op.create_node_with_data([data, axis_const], data_nodes=output_data)
def replace_op(self, graph: Graph, node: Node):
# split input to (i_part, f_part, c_part, o_part, ct_1)
split_node_axis = Const(graph, {'value': np.int64(1)}).create_node()
split_node = Split(graph, {
'name': 'Split_lstm_input_',
'num_splits': 5
}).create_node()
node.in_port(0).get_connection().set_destination(split_node.in_port(0))
split_node.in_port(1).connect(split_node_axis.out_port(0))
# i_t = Sigmoid(i_part + w_ic*ct_1)
i_scale_attrs = {'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)
split_node.out_port(4).connect(i_scale.in_port(0))
sum_i_c = Add(graph, {'name': 'sum_i_c_'}).create_node()
split_node.out_port(0).connect(sum_i_c.in_port(0))
i_scale.out_port(0).connect(sum_i_c.in_port(1))
i_sigmoid = Sigmoid(graph, {'name': 'i_sigmoid'}).create_node()
sum_i_c.out_port(0).connect(i_sigmoid.in_port(0))
# f_t = Sigmoid(f_part + w_fc*ct_1)
f_scale_attrs = {'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)
split_node.out_port(4).connect(f_scale.in_port(0))
sum_f_c = Add(graph, {'name': 'sum_f_c_'}).create_node()
split_node.out_port(1).connect(sum_f_c.in_port(0))
f_scale.out_port(0).connect(sum_f_c.in_port(1))
f_sigmoid = Sigmoid(graph, {'name': 'f_sigmoid'}).create_node()
sum_f_c.out_port(0).connect(f_sigmoid.in_port(0))
# c_t = f_t*ct_1 + i_t * tanh(c_part)
c_tanh = Tanh(graph, {'name': 'c_tanh'}).create_node()
split_node.out_port(2).connect(c_tanh.in_port(0))
prod_i_c_tanh = Mul(graph, {'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': 'prod_f_ct_1_'}).create_node()
f_sigmoid.out_port(0).connect(prod_f_ct_1.in_port(0))
split_node.out_port(4).connect(prod_f_ct_1.in_port(1))
sum_f_i = Add(graph, {'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': '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': 'sum_o_c_'}).create_node()
split_node.out_port(3).connect(sum_o_c.in_port(0))
o_scale.out_port(0).connect(sum_o_c.in_port(1))
o_sigmoid = Sigmoid(graph, {'name': 'o_sigmoid'}).create_node()
sum_o_c.out_port(0).connect(o_sigmoid.in_port(0))
# m_t = o_t * Tanh(c_t)
c_t_tanh = Tanh(graph, {'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': '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': '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]
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_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_axis = Const(graph, {
'value': np.int64(1)
}).create_node()
split_joined_input = Split(graph, {
'name': 'join_input_split',
'num_splits': 4,
}).create_node([join_input_prev_state_sum, split_joined_input_axis])
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]