def replace_op(self, graph: Graph, node: Node):
attrs = {'name': node.id + "/ScaleShift_"}
param = graph.node[node.id]['pb'].bn_param
pb_model = graph.node[node.id]['model_pb']
blobs = pb_model.blobs
if len(blobs) != 4:
raise Error("Incorrect number of blobs in BN layer {}".format(
node.id))
mean = np.array(blobs[0].data)
var = np.array(blobs[1].data)
betta = np.array(blobs[2].data)
gamma = np.array(blobs[3].data)
gamma = gamma + np.repeat(param.eps, gamma.shape)
scale = 1.0 / np.sqrt(gamma) * mean
shift = var - betta * scale
ss = ScaleShiftOp(graph, attrs)
scale_shift = ss.create_node([node.in_node(0)])
input_as_const(scale_shift, attrs, 1, 'weights', scale)
input_as_const(scale_shift, attrs, 2, 'biases', shift)
return [scale_shift.id]
def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<Dim>')
dim = read_binary_integer32_token(pb)
collect_until_token(pb, b'<BlockDim>')
block_dim = read_binary_integer32_token(pb)
collect_until_token(pb, b'<Epsilon>')
eps = read_binary_float_token(pb)
collect_until_token(pb, b'<TargetRms>')
target_rms = read_binary_float_token(pb)
collect_until_token(pb, b'<StatsMean>')
mean = read_binary_vector(pb)
collect_until_token(pb, b'<StatsVar>')
var = read_binary_vector(pb)
scale = target_rms / np.sqrt(var + eps)
shift = -target_rms * mean / np.sqrt(var + eps)
scale = np.tile(scale, dim // block_dim)
shift = np.tile(shift, dim // block_dim)
attrs = {'out-size': dim}
embed_input(attrs, 1, 'weights', scale)
embed_input(attrs, 2, 'biases', shift)
ScaleShiftOp.update_node_stat(node, attrs)
return cls.enabled
def extract(cls, node):
pb = node.pb
model = node.model_pb
param = pb.scale_param
attrs = {
'axis': param.axis,
}
if model is None and len(pb.bottom) == 1:
# default weights and biases for scale layer if the caffemodel file doesn't contain them
model = NamedAttrsClass({
'blobs':
mo_array([
NamedAttrsClass({'data': mo_array([1])}),
NamedAttrsClass({'data': mo_array([0])})
])
})
# scale with 1 input and 1 or 2 blobs
if model and len(model.blobs) != 0 and len(pb.bottom) == 1:
attrs.update(weights_biases(param.bias_term, model))
# 2 inputs + bias
elif len(pb.bottom) == 2 and param.bias_term:
if model is None or len(model.blobs) == 0:
# default bias for scale layer with 2 inputs if the caffemodel file doesn't contain them
model = NamedAttrsClass({
'blobs':
mo_array([NamedAttrsClass({'data': mo_array([0])})])
})
embed_input(attrs, 1, 'biases', model.blobs[0].data)
ScaleShiftOp.update_node_stat(node, attrs)
return cls.enabled
def extract(cls, node):
pb = node.parameters
read_learning_info(pb)
weights = read_binary_vector(pb)
mapping_rule = {}
embed_input(mapping_rule, 1, 'weights', weights)
ScaleShiftOp.update_node_stat(node, mapping_rule)
return cls.enabled
def extract(cls, node):
pb = node.parameters
read_learning_info(pb)
biases = read_binary_vector(pb)
bias_term = True
mapping_rule = {'bias_term': bias_term}
embed_input(mapping_rule, 1, 'weights', np.ones(biases.shape))
embed_input(mapping_rule, 2, 'biases', biases)
ScaleShiftOp.update_node_stat(node, mapping_rule)
return cls.enabled
def replace_op(self, graph: Graph, node: Node):
in_node_0 = node.in_node(0)
in_node_1 = node.in_node(1)
in_node_2 = node.in_node(2)
ss = ScaleShiftOp(graph, {'name': node.id + "/ScaleShift_", 'axis': 0})
scale_shift = ss.create_node(inputs=[in_node_1, in_node_0])
el = Add(graph, {'name': node.id + "/Add_"})
el_node = el.create_node(inputs=[scale_shift, in_node_2])
return [el_node.id]
def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<Dim>')
dim = read_binary_integer32_token(pb)
collect_until_token(pb, b'<Scale>')
scale = read_binary_float_token(pb)
# TODO add real batch here
attrs = {}
embed_input(attrs, 1, 'weights', np.full([dim], scale))
ScaleShiftOp.update_node_stat(node, attrs)
return cls.enabled
def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<Params>')
weights = read_binary_vector(pb)
find_next_tag(pb)
read_placeholder(pb, 1)
mapping_rule = {
'layout': 'NCHW'
}
embed_input(mapping_rule, 1, 'weights', weights)
ScaleShiftOp.update_node_stat(node, mapping_rule)
return cls.enabled
def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<Bias>')
biases = read_binary_vector(pb)
find_next_tag(pb)
read_placeholder(pb, 1)
mapping_rule = {
'layout': 'NCHW',
'bias_term': True,
'out-size': biases.shape[0],
}
embed_input(mapping_rule, 2, 'biases', biases)
ScaleShiftOp.update_node_stat(node, mapping_rule)
return cls.enabled
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_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]