def decode(conv_output, anchors, stride, num_class, conf_thresh):
conv_shape = P.shape(conv_output)
batch_size = conv_shape[0]
n_grid = conv_shape[1]
anchor_per_scale = len(anchors)
conv_output = P.reshape(
conv_output,
(batch_size, n_grid, n_grid, anchor_per_scale, 5 + num_class))
conv_raw_dxdy = conv_output[:, :, :, :, 0:2]
conv_raw_dwdh = conv_output[:, :, :, :, 2:4]
conv_raw_conf = conv_output[:, :, :, :, 4:5]
conv_raw_prob = conv_output[:, :, :, :, 5:]
rows = P.range(0, n_grid, 1, 'float32')
cols = P.range(0, n_grid, 1, 'float32')
rows = P.expand(P.reshape(rows, (1, -1, 1)), [n_grid, 1, 1])
cols = P.expand(P.reshape(cols, (-1, 1, 1)), [1, n_grid, 1])
offset = P.concat([rows, cols], axis=-1)
offset = P.reshape(offset, (1, n_grid, n_grid, 1, 2))
offset = P.expand(offset, [batch_size, 1, 1, anchor_per_scale, 1])
pred_xy = (P.sigmoid(conv_raw_dxdy) + offset) * stride
pred_wh = (P.exp(conv_raw_dwdh) * P.assign(anchors))
pred_xywh = P.concat([pred_xy, pred_wh], axis=-1)
pred_conf = P.sigmoid(conv_raw_conf)
pred_prob = P.sigmoid(conv_raw_prob)
pred_xywh = P.reshape(pred_xywh, (batch_size, -1, 4)) # [-1, -1, 4]
pred_conf = P.reshape(pred_conf, (batch_size, -1, 1)) # [-1, -1, 1]
pred_prob = P.reshape(pred_prob,
(batch_size, -1, num_class)) # [-1, -1, 80]
return pred_xywh, pred_conf, pred_prob
def lstm_step(x_t, hidden_t_prev, cell_t_prev, size, para_name, args):
"""Util function for pointer network"""
def linear(inputs, para_name, args):
return layers.fc(input=inputs,
size=size,
param_attr=fluid.ParamAttr(name=para_name + '_w'),
bias_attr=fluid.ParamAttr(name=para_name + '_b'))
input_cat = layers.concat([hidden_t_prev, x_t], axis=1)
forget_gate = layers.sigmoid(x=linear(input_cat, para_name + '_lstm_f',
args))
input_gate = layers.sigmoid(x=linear(input_cat, para_name + '_lstm_i',
args))
output_gate = layers.sigmoid(x=linear(input_cat, para_name + '_lstm_o',
args))
cell_tilde = layers.tanh(x=linear(input_cat, para_name + '_lstm_c', args))
cell_t = layers.sums(input=[
layers.elementwise_mul(
x=forget_gate, y=cell_t_prev), layers.elementwise_mul(
x=input_gate, y=cell_tilde)
])
hidden_t = layers.elementwise_mul(x=output_gate, y=layers.tanh(x=cell_t))
return hidden_t, cell_t
def forward(self, input_tensor, cur_state):
h_cur = cur_state
x_in = concat([input_tensor, h_cur], axis=1)
update = sigmoid(self.update_gate(x_in))
reset = sigmoid(self.reset_gate(x_in))
x_out = tanh(
self.out_gate(concat([input_tensor, h_cur * reset], axis=1)))
h_new = h_cur * (1 - update) + x_out * update
return h_new
def build_graph(self, mode='train'):
self._build_data()
pred = self._build_net()
if mode == 'train':
loss, no_grad_set = self._compute_loss(pred)
pred = layers.sigmoid(pred)
acc = self._compute_acc(pred)
return loss, acc, pred, no_grad_set
else:
pred = layers.sigmoid(pred)
return pred
def forward(self, q, k, v, lengths, speaker_embed, start_index,
force_monotonic=False, prev_coeffs=None, window=None):
# add position encoding as an inductive bias
if self.has_bias: # multi-speaker model
omega_q = 2 * F.sigmoid(
F.squeeze(self.q_pos_affine(speaker_embed), axes=[-1]))
omega_k = 2 * self.omega_initial * F.sigmoid(F.squeeze(
self.k_pos_affine(speaker_embed), axes=[-1]))
else: # single-speaker case
batch_size = q.shape[0]
omega_q = F.ones((batch_size, ), dtype="float32")
omega_k = F.ones((batch_size, ), dtype="float32") * self.omega_default
q += self.position_encoding_weight * positional_encoding(q, start_index, omega_q)
k += self.position_encoding_weight * positional_encoding(k, 0, omega_k)
q, k, v = self.q_affine(q), self.k_affine(k), self.v_affine(v)
activations = F.matmul(q, k, transpose_y=True)
activations /= np.sqrt(self.attention_dim)
if self.training:
# mask the <pad> parts from the encoder
mask = F.sequence_mask(lengths, dtype="float32")
attn_bias = F.scale(1. - mask, -1000)
activations += F.unsqueeze(attn_bias, [1])
elif force_monotonic:
assert window is not None
backward_step, forward_step = window
T_enc = k.shape[1]
batch_size, T_dec, _ = q.shape
# actually T_dec = 1 here
alpha = F.fill_constant((batch_size, T_dec), value=0, dtype="int64") \
if prev_coeffs is None \
else F.argmax(prev_coeffs, axis=-1)
backward = F.sequence_mask(alpha - backward_step, maxlen=T_enc, dtype="bool")
forward = F.sequence_mask(alpha + forward_step, maxlen=T_enc, dtype="bool")
mask = F.cast(F.logical_xor(backward, forward), "float32")
# print("mask's shape:", mask.shape)
attn_bias = F.scale(1. - mask, -1000)
activations += attn_bias
# softmax
coefficients = F.softmax(activations, axis=-1)
# context vector
coefficients = F.dropout(coefficients, 1. - self.keep_prob,
dropout_implementation='upscale_in_train')
contexts = F.matmul(coefficients, v)
# context normalization
enc_lengths = F.cast(F.unsqueeze(lengths, axes=[1, 2]), "float32")
contexts *= F.sqrt(enc_lengths)
# out affine
contexts = self.out_affine(contexts)
return contexts, coefficients
def forward(self, input, pre_hidden, pre_cell):
concat_input_hidden = layers.concat([input, pre_hidden], 1)
gate_input = layers.matmul(x=concat_input_hidden, y=self._weight)
gate_input = layers.elementwise_add(gate_input, self._bias)
i, j, f, o = layers.split(gate_input, num_or_sections=4, dim=-1)
new_cell = layers.elementwise_add(
layers.elementwise_mul(
pre_cell,
layers.sigmoid(layers.elementwise_add(f, self._forget_bias))),
layers.elementwise_mul(layers.sigmoid(i), layers.tanh(j)))
new_hidden = layers.tanh(new_cell) * layers.sigmoid(o)
return new_hidden, new_cell
def build_program(self, dtype):
with fluid.program_guard(self.main_program, self.startup_program):
self.feed_vars = self._prepare_feed_vars([32, 128], dtype, 5)
tmp_0 = layers.assign(self.feed_vars[0])
# subgraph with 9 op nodes
tmp_1 = tmp_0 * layers.sigmoid(self.feed_vars[1]) + layers.sigmoid(
self.feed_vars[2]) * layers.tanh(self.feed_vars[3])
tmp_2 = layers.tanh(tmp_1) + layers.sigmoid(self.feed_vars[4])
self.append_gradients(tmp_2)
self.num_fused_ops = 2
self.fetch_list = [tmp_2, self.grad(tmp_0)]
def build_program(self, dtype):
with fluid.program_guard(self.main_program, self.startup_program):
self.feed_vars = self._prepare_feed_vars([32, 64], dtype, 5)
one = layers.fill_constant(shape=[1], dtype=dtype, value=1.0)
tmp_0 = one * self.feed_vars[0]
# subgraph with 9 op nodes
tmp_1 = tmp_0 * layers.sigmoid(self.feed_vars[1]) + layers.sigmoid(
self.feed_vars[2]) * layers.tanh(self.feed_vars[3])
tmp_2 = layers.tanh(tmp_1) + layers.sigmoid(self.feed_vars[4])
self.append_gradients(tmp_2)
self.num_fused_ops = 2
self.fetch_list = [tmp_2, self.grad(tmp_0)]
def _decode(self,
x,
y,
w,
h,
anchors,
stride,
scale_x_y,
eps,
is_gt=False):
conv_shape = x.shape # (8, 13, 13, 3)
batch_size = conv_shape[0]
n_grid = conv_shape[1]
anchor_per_scale = conv_shape[3]
_x = L.unsqueeze(x, 4)
_y = L.unsqueeze(y, 4)
conv_raw_dxdy = L.concat([_x, _y], -1) # (8, 13, 13, 3, 2)
_w = L.unsqueeze(w, 4)
_h = L.unsqueeze(h, 4)
conv_raw_dwdh = L.concat([_w, _h], -1) # (8, 13, 13, 3, 2)
rows = L.range(0, n_grid, 1, 'float32')
cols = L.range(0, n_grid, 1, 'float32')
rows = L.expand(L.reshape(rows, (1, -1, 1)), [n_grid, 1, 1])
cols = L.expand(L.reshape(cols, (-1, 1, 1)), [1, n_grid, 1])
offset = L.concat([rows, cols], axis=-1)
offset = L.reshape(offset, (1, n_grid, n_grid, 1, 2))
offset = L.expand(offset, [batch_size, 1, 1, anchor_per_scale, 1])
if is_gt:
decode_xy = (conv_raw_dxdy + offset) / n_grid
else:
if (abs(scale_x_y - 1.0) < eps):
decode_xy = L.sigmoid(conv_raw_dxdy)
decode_xy = (decode_xy + offset) / n_grid
else:
# Grid Sensitive
decode_xy = scale_x_y * L.sigmoid(conv_raw_dxdy) - 0.5 * (
scale_x_y - 1.0)
decode_xy = (decode_xy + offset) / n_grid
anchor_t = fluid.layers.assign(np.copy(anchors).astype(np.float32))
decode_wh = (L.exp(conv_raw_dwdh) * anchor_t) / (n_grid * stride)
decode_xywh = L.concat([decode_xy, decode_wh], axis=-1)
if is_gt:
decode_xywh.stop_gradient = True
return decode_xywh # (8, 13, 13, 3, 4)
def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
h_boot = layers.data(shape=[self.input_dim],
dtype='float32',
name='h_boot')
h_boot.stop_gradient = False
rnn = layers.StaticRNN()
with rnn.step():
h_pre = rnn.memory(init=h_boot)
x_t = rnn.step_input(x)
temp_l = layers.fc(input=x_t,
size=self.input_dim,
param_attr='W',
bias_attr=False)
temp_r = layers.fc(input=h_pre,
size=self.input_dim,
param_attr='U',
bias_attr=False)
h = layers.sigmoid(x=layers.elementwise_add(x=temp_l, y=temp_r))
rnn.update_memory(h_pre, h)
rnn.output(h)
return rnn()
def add_input(self, x, condition=None):
"""Add a step input. This method works similarily with `forward` but in a `step-in-step-out` fashion.
Args:
x (Variable): shape(B, C_res, T=1), input for a step, dtype float32.
condition (Variable, optional): shape(B, C_cond, T=1). condition for a step, dtype float32. Defaults to None.
Returns:
(residual, skip_connection)
residual (Variable): shape(B, C_res, T=1), the residual for a step, which is used as the input to the next layer of ResidualBlock.
skip_connection (Variable): shape(B, C_res, T=1), the skip connection for a step. This output is accumulated with that of other ResidualBlocks.
"""
h = x
# dilated conv
h = self.conv.add_input(h)
# condition
if condition is not None:
h += self.condition_proj(condition)
# gated tanh
content, gate = F.split(h, 2, dim=1)
z = F.sigmoid(gate) * F.tanh(content)
# projection
residual = F.scale(z + x, np.sqrt(0.5))
skip_connection = z
return residual, skip_connection
def forward(self, x, speaker_embed=None):
"""
Args:
x (Variable): shape(B, C_in, T), dtype float32, the input of Conv1DGLU layer, where B means batch_size, C_in means the input channels T means input time steps.
speaker_embed (Variable): shape(B, C_sp), dtype float32, speaker embed, where C_sp means speaker embedding size.
Returns:
x (Variable): shape(B, C_out, T), the output of Conv1DGLU, where
C_out means the `num_filters`.
"""
residual = x
x = F.dropout(x,
self.dropout,
dropout_implementation="upscale_in_train")
x = self.conv(x)
content, gate = F.split(x, num_or_sections=2, dim=1)
if speaker_embed is not None:
sp = F.softsign(self.fc(speaker_embed))
content = F.elementwise_add(content, sp, axis=0)
# glu
x = F.sigmoid(gate) * content
if self.residual:
x = F.scale(x + residual, np.sqrt(0.5))
return x
def add_input(self, x_t, speaker_embed=None):
"""
Takes a step of inputs and return a step of outputs. It works similarily with the `forward` method, but in a `step-in-step-out` fashion.
Args:
x_t (Variable): shape(B, C_in, T=1), dtype float32, the input of Conv1DGLU layer, where B means batch_size, C_in means the input channels.
speaker_embed (Variable): Shape(B, C_sp), dtype float32, speaker embed, where C_sp means speaker embedding size.
Returns:
x (Variable): shape(B, C_out), the output of Conv1DGLU, where C_out means the `num_filter`.
"""
residual = x_t
x_t = F.dropout(x_t,
self.dropout,
dropout_implementation="upscale_in_train")
x_t = self.conv.add_input(x_t)
content_t, gate_t = F.split(x_t, num_or_sections=2, dim=1)
if speaker_embed is not None:
sp = F.softsign(self.fc(speaker_embed))
content_t = F.elementwise_add(content_t, sp, axis=0)
# glu
x_t = F.sigmoid(gate_t) * content_t
if self.residual:
x_t = F.scale(x_t + residual, np.sqrt(0.5))
return x_t
def forward(self, input, bias=None, padding=None):
"""
input: input feature (B, T, C)
padding: only used when using causal conv, we pad mannually
"""
input_dropped = F.dropout(input,
1. - self.keep_prob,
dropout_implementation="upscale_in_train")
if self.causal:
assert padding is not None
input_dropped = F.concat([padding, input_dropped], axis=1)
hidden = self.conv(input_dropped)
if self.has_bias:
assert bias is not None
transformed_bias = F.softsign(self.bias_affine(bias))
hidden_embedded = hidden + F.unsqueeze(transformed_bias, [1])
else:
hidden_embedded = hidden
# glu
content, gate = F.split(hidden, num_or_sections=2, dim=-1)
content = hidden_embedded[:, :, :self.in_channel]
hidden = F.sigmoid(gate) * content
# # residual
hidden = F.scale(input + hidden, math.sqrt(0.5))
return hidden
def link_predict_model(num_nodes,
hidden_size=16,
name='link_predict_task',
binary_op_type="Weighted-L2"):
pyreader = l.py_reader(capacity=70,
shapes=[[-1, 1], [-1, 1], [-1, 1]],
dtypes=['int64', 'int64', 'int64'],
lod_levels=[0, 0, 0],
name=name + '_pyreader',
use_double_buffer=True)
u, v, label = l.read_file(pyreader)
u_embed = l.embedding(input=u,
size=[num_nodes, hidden_size],
param_attr=fluid.ParamAttr(name='content'))
v_embed = l.embedding(input=v,
size=[num_nodes, hidden_size],
param_attr=fluid.ParamAttr(name='content'))
u_embed.stop_gradient = True
v_embed.stop_gradient = True
edge_embed = binary_op(u_embed, v_embed, binary_op_type)
logit = l.fc(input=edge_embed, size=1)
loss = l.sigmoid_cross_entropy_with_logits(logit, l.cast(label, 'float32'))
loss = l.reduce_mean(loss)
prob = l.sigmoid(logit)
return pyreader, loss, prob, label
def test_sigmoid(self):
program = Program()
with program_guard(program):
input = layers.data(name="input", shape=[16], dtype="float32")
out = layers.sigmoid(input, name='sigmoid')
self.assertIsNotNone(out)
print(str(program))
def forward(self, x, condition=None):
"""Conv1D gated-tanh Block.
Args:
x (Variable): shape(B, C_res, T), the input. (B stands for batch_size, C_res stands for residual channels, T stands for time steps.) dtype float32.
condition (Variable, optional): shape(B, C_cond, T), the condition, it has been upsampled in time steps, so it has the same time steps as the input does.(C_cond stands for the condition's channels). Defaults to None.
Returns:
(residual, skip_connection)
residual (Variable): shape(B, C_res, T), the residual, which is used as the input to the next layer of ResidualBlock.
skip_connection (Variable): shape(B, C_res, T), the skip connection. This output is accumulated with that of other ResidualBlocks.
"""
time_steps = x.shape[-1]
h = x
# dilated conv
h = self.conv(h)
if h.shape[-1] != time_steps:
h = h[:, :, :time_steps]
# condition
if condition is not None:
h += self.condition_proj(condition)
# gated tanh
content, gate = F.split(h, 2, dim=1)
z = F.sigmoid(gate) * F.tanh(content)
# projection
residual = F.scale(z + x, math.sqrt(.5))
skip_connection = z
return residual, skip_connection
def create_model(args, config, graph_label):
"""Create model for given model configuration."""
logging.info('building model')
graph_wrapper = GraphWrapper(name="graph",
node_feat=[('atom_type', [None, 1], "int64"),
('chirality_tag', [None,
1], "int64")],
edge_feat=[('bond_type', [None, 1], "int64"),
('bond_direction', [None,
1], "int64")])
encoder = GINEncoder(config)
global_repr, patch_summary = encoder.forward(graph_wrapper)
hid = L.fc(global_repr,
config['hidden_size'],
act='relu',
name='finetune_fc1')
hid = L.fc(hid, config['hidden_size'], act='relu', name='finetune_fc2')
logits = L.fc(global_repr, args.num_tasks, name="finetune_fc3")
loss = L.sigmoid_cross_entropy_with_logits(x=logits, label=graph_label)
loss = L.reduce_mean(loss)
pred = L.sigmoid(logits)
keys = ['loss', 'graph_wrapper', 'encoder', 'graph_emb', 'pred']
Agent = namedtuple('Agent', keys)
return Agent(loss=loss,
graph_wrapper=graph_wrapper,
encoder=encoder,
graph_emb=global_repr,
pred=pred)
def decoder_step(gru_unit,
cue_gru_unit,
step_in,
hidden,
input_size,
hidden_size,
memory,
memory_mask,
knowledge,
mask=None):
""" decoder step """
# get attention out
# get hidden top layers
top_hidden = layers.slice(hidden, axes=[0], starts=[0], ends=[1])
top_hidden = layers.squeeze(top_hidden, axes=[0])
top_hidden = layers.unsqueeze(top_hidden, axes=[1])
weight_memory, attn = dot_attention(top_hidden, memory, memory_mask)
step_in = layers.unsqueeze(step_in, axes=[1])
rnn_input_list = [step_in, weight_memory]
if weight_memory.shape[0] == -1:
knowledge_1 = layers.reshape(knowledge, shape=weight_memory.shape)
else:
knowledge_1 = knowledge
cue_input_list = [knowledge_1, weight_memory]
output_list = [weight_memory]
rnn_input = layers.concat(rnn_input_list, axis=2)
rnn_input = layers.squeeze(rnn_input, axes=[1])
rnn_output, rnn_last_hidden = gru_unit(rnn_input, hidden, mask)
cue_input = layers.concat(cue_input_list, axis=2)
cue_input = layers.squeeze(cue_input, axes=[1])
cue_rnn_out, cue_rnn_last_hidden = cue_gru_unit(cue_input, hidden, mask)
h_y = layers.tanh(
fc(rnn_last_hidden, hidden_size, hidden_size, name="dec_fc1"))
h_cue = layers.tanh(
fc(cue_rnn_last_hidden, hidden_size, hidden_size, name="dec_fc2"))
concate_y_cue = layers.concat([h_y, h_cue], axis=2)
k = layers.sigmoid(fc(concate_y_cue, hidden_size * 2, 1, name='dec_fc3'))
new_hidden = h_y * k - h_cue * (k - 1.0)
new_hidden_tmp = layers.transpose(new_hidden, perm=[1, 0, 2])
output_list.append(new_hidden_tmp)
real_out = layers.concat(output_list, axis=2)
if mask:
mask_tmp = layers.unsqueeze(mask, axes=[0])
new_hidden = layers.elementwise_mul((new_hidden - hidden),
mask_tmp,
axis=0)
new_hidden += hidden
return real_out, new_hidden
def forward(self, features, im_info):
# features: p6 -> p2
pred_objectness_logits, pred_anchor_deltas = self.rpn_head(features)
rpn_rois = []
rpn_roi_probs = [] # p2 -> p6
for lvl in range(self.k_min, self.k_max + 1): # 2 -> 6
lvl_anchors = generate_anchors(
stride=2.**lvl,
sizes=[self.anchor_size[0] * 2.**(lvl - self.k_min)],
aspect_ratios=self.anchor_aspect_ratios)
lvl_cls_logits = pred_objectness_logits[self.k_max - lvl]
lvl_cls_logits = L.sigmoid(lvl_cls_logits).numpy()
lvl_bbox_deltas = pred_anchor_deltas[self.k_max - lvl].numpy()
lvl_rois, lvl_roi_probs, lvl_anchors = self.generate_proposal_op(
lvl_cls_logits, lvl_bbox_deltas, lvl_anchors, 1 / 2**lvl,
im_info)
rpn_rois.append(lvl_rois)
rpn_roi_probs.append(lvl_roi_probs)
rois = self.collect_and_distribute_op(rpn_rois, rpn_roi_probs)
# list of ndarray(size: (n, 5)), p2 -> p5
return rois
def _compute_pc(self, x, mask):
if mask is not None:
x -= (1 - mask) * 1e10
x = layers.reduce_max(x, dim=1, keep_dim=True)
x = layers.relu(self.pc_fc1(x))
x = layers.sigmoid(self.pc_fc2(x))
return x
def appnp(gw, feature, alpha=0.2, k_hop=10, name=""):
"""Implementation of APPNP of "Predict then Propagate: Graph Neural Networks
meet Personalized PageRank" (ICLR 2019).
Args:
gw: Graph wrapper object (:code:`StaticGraphWrapper` or :code:`GraphWrapper`)
feature: A tensor with shape (num_nodes, feature_size).
edge_dropout: Edge dropout rate.
k_hop: K Steps for Propagation
Return:
A tensor with shape (num_nodes, hidden_size)
"""
def send_src_copy(src_feat, dst_feat, edge_feat):
feature = src_feat["h"]
return feature
def get_norm(indegree):
float_degree = L.cast(indegree, dtype="float32")
float_degree = L.clamp(float_degree, min=1.0)
norm = L.pow(float_degree, factor=-0.5)
return norm
cks = []
h0 = feature
ngw = gw
norm = get_norm(ngw.indegree())
for i in range(k_hop):
feature = feature * norm
msg = gw.send(send_src_copy, nfeat_list=[("h", feature)])
feature = gw.recv(msg, "sum")
feature = feature * norm
#feature = feature * (1 - alpha) + h0 * alpha
fan_in = feature.shape[-1] * 3
bias_bound = 1.0 / math.sqrt(fan_in)
fc_bias_attr = F.ParamAttr(
initializer=F.initializer.UniformInitializer(low=-bias_bound,
high=bias_bound))
negative_slope = math.sqrt(5)
gain = math.sqrt(2.0 / (1 + negative_slope**2))
std = gain / math.sqrt(fan_in)
weight_bound = math.sqrt(3.0) * std
fc_w_attr = F.ParamAttr(initializer=F.initializer.UniformInitializer(
low=-weight_bound, high=weight_bound))
gate_f = L.fc([feature, h0, feature - h0],
1,
param_attr=fc_w_attr,
name=name + 'appnp_gate_' + str(i),
bias_attr=fc_bias_attr)
alpha = L.sigmoid(gate_f)
feature = feature * (1 - alpha) + h0 * alpha
if (i + 1) % 3 == 0:
cks.append(feature)
return feature, cks
def gru_step(self, input, hidden, mask=None):
""" gru step """
hidden_array = []
for i in range(self.num_layers):
hidden_temp = layers.slice(hidden,
axes=[0],
starts=[i],
ends=[i + 1])
hidden_temp = layers.reshape(hidden_temp,
shape=[-1, self.hidden_size])
hidden_array.append(hidden_temp)
last_hidden_array = []
for k in range(self.num_layers):
trans_input = layers.matmul(input, self.weight_input_array[k])
trans_input += self.bias_input_array[k]
trans_hidden = layers.matmul(hidden_array[k],
self.weight_hidden_array[k])
trans_hidden += self.bias_hidden_array[k]
input_array = layers.split(trans_input, num_or_sections=3, dim=-1)
trans_array = layers.split(trans_hidden, num_or_sections=3, dim=-1)
reset_gate = layers.sigmoid(input_array[0] + trans_array[0])
input_gate = layers.sigmoid(input_array[1] + trans_array[1])
new_gate = layers.tanh(input_array[2] +
reset_gate * trans_array[2])
new_hidden = new_gate + input_gate * (hidden_array[k] - new_gate)
if mask:
neg_mask = layers.fill_constant_batch_size_like(
input=mask, shape=[1], value=1.0, dtype='float32') - mask
new_hidden = new_hidden * mask + hidden_array[k] * neg_mask
last_hidden_array.append(new_hidden)
input = new_hidden
if self.dropout and self.dropout > 0.0:
input = layers.dropout(input, dropout_prob=self.dropout)
last_hidden = layers.concat(last_hidden_array, 0)
last_hidden = layers.reshape(
last_hidden, shape=[self.num_layers, -1, self.hidden_size])
return input, last_hidden
def inference(self):
"""
Used for inference with labels.
"""
graph_wrapper, logits = self.forward(is_test=True)
pred = layers.sigmoid(logits)
self.graph_wrapper = graph_wrapper
self.pred = pred
def forward(self, input, state):
#logging.info("input shape: {}".format(input.shape))
pre_hidden, pre_cell = state
#logging.info("pre hidden shape: {}".format(pre_hidden.shape))
#logging.info("pre cell shape: {}".format(pre_cell.shape))
# i,f,c,o 四个值均有Wx+Wh+b 即W(x+h)+b
# 因此：
# 实际相乘为[x, b]·W+b
# x,b 横向相连, shape为[batch_size, input_size+hidden_size]
# W的shape为[input_size+hidden_size, 4*hidden_size]
# b的shape为[4*hidden_size,]
# 横向连接
# shape: [batch_size, input_size+hidden_size]
concat_input_hidden = L.concat([input, pre_hidden], axis=1)
#logging.info("x concat h shape: {}".format(concat_input_hidden.shape))
# 计算Wx+Wh+b
# shape: [batch_size, 4*hidden_size]
gate_input = L.matmul(x=concat_input_hidden, y=self._weight)
#logging.info("[x, b]·W shape: {}".format(gate_input.shape))
# shape: [batch_size, 4*hidden_size]
gate_input = L.elementwise_add(gate_input, self._bias)
#logging.info("[x, b]·W+b shape: {}".format(gate_input.shape))
# i,f,c,o四值按最后一维分开 因此每个的最后一维都是hidden_size
i, f, c, o = L.split(gate_input, num_or_sections=4, dim=-1)
# new_c = pre_c·sigmoid(f+forget_bias) + sigmoid(i)·tanh(c)
# shape: [batch_size, hidden_size]
new_cell = L.elementwise_add(
L.elementwise_mul(
pre_cell,
L.sigmoid(L.elementwise_add(f, self._forget_bias))),
L.elementwise_mul(L.sigmoid(i), L.tanh(c))
)
#logging.info("new_cell shape: {}".format(new_cell.shape))
# new_h = tanh(new_c)*sigmoid(o)
# shape: [batch_size, hidden_size]
new_hidden = L.tanh(new_cell) * L.sigmoid(o)
#logging.info("new_hidden shape: {}".format(new_hidden.shape))
return new_hidden, [new_hidden, new_cell]
def act(a, act='tanh'):
if act == 'tanh':
return layers.tanh(a)
elif act == 'sigmoid':
return layers.sigmoid(a)
elif act == 'relu':
return layers.relu(a)
else:
return a
def forward(self, t, k, pn1, pn2, mask=None):
pc = self._compute_pc(t, mask)
k = get_k_inter(t, k)
k = layers.expand(k, [1, t.shape[1], 1])
h = self.multi_head_attn(t, mask=mask)
h = layers.concat([t, h, k], axis=-1)
h = self.conv1d(h)
po = layers.sigmoid(self.po_fc(h))
po1 = layers.sigmoid(self.po1_fc(h))
po2 = layers.sigmoid(self.po2_fc(h))
po1 = po * po1 * pc * pn1
po2 = po * po2 * pc * pn2
return po1, po2
def _split_ioup(self, output, an_num, num_classes):
"""
Split output feature map to output, predicted iou
along channel dimension
"""
ioup = output[:, :an_num, :, :]
ioup = L.sigmoid(ioup)
oriout = output[:, an_num:, :, :]
return (ioup, oriout)
def lstm_step(x_t, hidden_t_prev, cell_t_prev, size): def linear(inputs): return layers.fc(input=inputs, size=size, bias_attr=True) forget_gate = layers.sigmoid(x=linear([hidden_t_prev, x_t])) input_gate = layers.sigmoid(x=linear([hidden_t_prev, x_t])) output_gate = layers.sigmoid(x=linear([hidden_t_prev, x_t])) cell_tilde = layers.tanh(x=linear([hidden_t_prev, x_t])) cell_t = layers.sums(input=[ layers.elementwise_mul( x=forget_gate, y=cell_t_prev), layers.elementwise_mul( x=input_gate, y=cell_tilde) ]) hidden_t = layers.elementwise_mul( x=output_gate, y=layers.tanh(x=cell_t)) return hidden_t, cell_t
def build_model(self):
node_features = self.graph_wrapper.node_feat["feat"]
output = self.gcn(gw=self.graph_wrapper,
feature=node_features,
hidden_size=self.hidden_size,
activation="relu",
norm=self.graph_wrapper.node_feat["norm"],
name="gcn_layer_1")
output1 = output
output = self.gcn(gw=self.graph_wrapper,
feature=output,
hidden_size=self.hidden_size,
activation="relu",
norm=self.graph_wrapper.node_feat["norm"],
name="gcn_layer_2")
output2 = output
output = self.gcn(gw=self.graph_wrapper,
feature=output,
hidden_size=self.hidden_size,
activation="relu",
norm=self.graph_wrapper.node_feat["norm"],
name="gcn_layer_3")
output = L.concat(input=[output1, output2, output], axis=-1)
output, ratio_length = sag_pool(gw=self.graph_wrapper,
feature=output,
ratio=self.pooling_ratio,
graph_id=self.graph_id,
dataset=self.args.dataset_name,
name="sag_pool_1")
output = L.lod_reset(output, self.graph_wrapper.graph_lod)
cat1 = L.sequence_pool(output, "sum")
ratio_length = L.cast(ratio_length, dtype="float32")
cat1 = L.elementwise_div(cat1, ratio_length, axis=-1)
cat2 = L.sequence_pool(output, "max")
output = L.concat(input=[cat2, cat1], axis=-1)
output = L.fc(output, size=self.hidden_size, act="relu")
output = L.dropout(output, dropout_prob=self.dropout_ratio)
output = L.fc(output, size=self.hidden_size // 2, act="relu")
output = L.fc(output,
size=self.num_classes,
act=None,
param_attr=fluid.ParamAttr(name="final_fc"))
self.labels = L.cast(self.labels, dtype="float32")
loss = L.sigmoid_cross_entropy_with_logits(x=output, label=self.labels)
self.loss = L.mean(loss)
pred = L.sigmoid(output)
self.pred = L.argmax(x=pred, axis=-1)
correct = L.equal(self.pred, self.labels_1dim)
correct = L.cast(correct, dtype="int32")
self.correct = L.reduce_sum(correct)
def create_rnn_op(self):
x = layers.data(
shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False,
**self.p_info)
x.stop_gradient = False
h_boot = layers.data(
shape=[self.input_dim],
dtype='float32',
name='h_boot',
**self.p_info)
h_boot.stop_gradient = False
rnn = layers.StaticRNN(main_program=self.main_program)
with rnn.step():
h_pre = rnn.memory(init=h_boot)
x_t = rnn.step_input(x)
temp_l = layers.fc(input=x_t,
size=self.input_dim,
param_attr='W',
bias_attr=False,
**self.p_info)
temp_r = layers.fc(input=h_pre,
size=self.input_dim,
param_attr='U',
bias_attr=False,
**self.p_info)
h = layers.sigmoid(
x=layers.elementwise_add(
x=temp_l, y=temp_r, **self.p_info),
**self.p_info)
rnn.update_memory(h_pre, h)
rnn.output(h)
return rnn()