wehub-resource-sync
389 lines
16 KiB
Python
389 lines
16 KiB
Python
# Adopted from https://github.com/KiroSummer/A_Syntax-aware_MTL_Framework_for_Chinese_SRL
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
from torch.autograd import Variable
|
|
import numpy as np
|
|
import torch.nn.functional as F
|
|
|
|
from hanlp.components.srl.span_rank.util import block_orth_normal_initializer
|
|
|
|
|
|
def get_tensor_np(t):
|
|
return t.data.cpu().numpy()
|
|
|
|
|
|
def orthonormal_initializer(output_size, input_size):
|
|
"""adopted from Timothy Dozat https://github.com/tdozat/Parser/blob/master/lib/linalg.py
|
|
|
|
Args:
|
|
output_size:
|
|
input_size:
|
|
|
|
Returns:
|
|
|
|
|
|
"""
|
|
print((output_size, input_size))
|
|
I = np.eye(output_size)
|
|
lr = .1
|
|
eps = .05 / (output_size + input_size)
|
|
success = False
|
|
tries = 0
|
|
while not success and tries < 10:
|
|
Q = np.random.randn(input_size, output_size) / np.sqrt(output_size)
|
|
for i in range(100):
|
|
QTQmI = Q.T.dot(Q) - I
|
|
loss = np.sum(QTQmI ** 2 / 2)
|
|
Q2 = Q ** 2
|
|
Q -= lr * Q.dot(QTQmI) / (
|
|
np.abs(Q2 + Q2.sum(axis=0, keepdims=True) + Q2.sum(axis=1, keepdims=True) - 1) + eps)
|
|
if np.max(Q) > 1e6 or loss > 1e6 or not np.isfinite(loss):
|
|
tries += 1
|
|
lr /= 2
|
|
break
|
|
success = True
|
|
if success:
|
|
print(('Orthogonal pretrainer loss: %.2e' % loss))
|
|
else:
|
|
print('Orthogonal pretrainer failed, using non-orthogonal random matrix')
|
|
Q = np.random.randn(input_size, output_size) / np.sqrt(output_size)
|
|
return np.transpose(Q.astype(np.float32))
|
|
|
|
|
|
class LayerNorm(nn.Module):
|
|
def __init__(self, features, eps=1e-8):
|
|
super(LayerNorm, self).__init__()
|
|
self.gamma = nn.Parameter(torch.ones(features))
|
|
self.beta = nn.Parameter(torch.zeros(features))
|
|
self.eps = eps
|
|
|
|
def forward(self, x):
|
|
mean = x.mean(-1, keepdim=True)
|
|
std = x.std(-1, keepdim=True)
|
|
return self.gamma * (x - mean) / (std + self.eps) + self.beta
|
|
|
|
|
|
class DropoutLayer3D(nn.Module):
|
|
def __init__(self, input_size, dropout_rate=0.0):
|
|
super(DropoutLayer3D, self).__init__()
|
|
self.dropout_rate = dropout_rate
|
|
self.input_size = input_size
|
|
self.drop_mask = torch.FloatTensor(self.input_size).fill_(1 - self.dropout_rate)
|
|
self.drop_mask = Variable(torch.bernoulli(self.drop_mask), requires_grad=False)
|
|
if torch.cuda.is_available():
|
|
self.drop_mask = self.drop_mask.cuda()
|
|
|
|
def reset_dropout_mask(self, batch_size, length):
|
|
self.drop_mask = torch.FloatTensor(batch_size, length, self.input_size).fill_(1 - self.dropout_rate)
|
|
self.drop_mask = Variable(torch.bernoulli(self.drop_mask), requires_grad=False)
|
|
if torch.cuda.is_available():
|
|
self.drop_mask = self.drop_mask.cuda()
|
|
|
|
def forward(self, x):
|
|
if self.training:
|
|
return torch.mul(x, self.drop_mask)
|
|
else: # eval
|
|
return x * (1.0 - self.dropout_rate)
|
|
|
|
|
|
class DropoutLayer(nn.Module):
|
|
def __init__(self, input_size, dropout_rate=0.0):
|
|
super(DropoutLayer, self).__init__()
|
|
self.dropout_rate = dropout_rate
|
|
self.input_size = input_size
|
|
self.drop_mask = torch.Tensor(self.input_size).fill_(1 - self.dropout_rate)
|
|
self.drop_mask = torch.bernoulli(self.drop_mask)
|
|
|
|
def reset_dropout_mask(self, batch_size):
|
|
self.drop_mask = torch.Tensor(batch_size, self.input_size).fill_(1 - self.dropout_rate)
|
|
self.drop_mask = torch.bernoulli(self.drop_mask)
|
|
|
|
def forward(self, x):
|
|
if self.training:
|
|
return torch.mul(x, self.drop_mask.to(x.device))
|
|
else: # eval
|
|
return x * (1.0 - self.dropout_rate)
|
|
|
|
|
|
class NonLinear(nn.Module):
|
|
def __init__(self, input_size, hidden_size, activation=None):
|
|
super(NonLinear, self).__init__()
|
|
self.input_size = input_size
|
|
self.hidden_size = hidden_size
|
|
self.linear = nn.Linear(in_features=input_size, out_features=hidden_size)
|
|
if activation is None:
|
|
self._activate = lambda x: x
|
|
else:
|
|
if not callable(activation):
|
|
raise ValueError("activation must be callable: type={}".format(type(activation)))
|
|
self._activate = activation
|
|
|
|
self.reset_parameters()
|
|
|
|
def forward(self, x):
|
|
y = self.linear(x)
|
|
return self._activate(y)
|
|
|
|
def reset_parameters(self):
|
|
nn.init.xavier_uniform_(self.linear.weight)
|
|
nn.init.zeros_(self.linear.bias)
|
|
|
|
|
|
class Biaffine(nn.Module):
|
|
def __init__(self, in1_features, in2_features, out_features,
|
|
bias=(True, True)):
|
|
super(Biaffine, self).__init__()
|
|
self.in1_features = in1_features
|
|
self.in2_features = in2_features
|
|
self.out_features = out_features
|
|
self.bias = bias
|
|
self.linear_input_size = in1_features + int(bias[0])
|
|
self.linear_output_size = out_features * (in2_features + int(bias[1]))
|
|
self.linear = nn.Linear(in_features=self.linear_input_size,
|
|
out_features=self.linear_output_size,
|
|
bias=False)
|
|
|
|
self.reset_parameters()
|
|
|
|
def reset_parameters(self):
|
|
torch.nn.init.xavier_uniform_(self.linear.weight)
|
|
|
|
def forward(self, input1, input2):
|
|
batch_size, len1, dim1 = input1.size()
|
|
batch_size, len2, dim2 = input2.size()
|
|
if self.bias[0]:
|
|
ones = input1.data.new(batch_size, len1, 1).zero_().fill_(1) # this kind of implementation is too tedious
|
|
input1 = torch.cat((input1, Variable(ones)), dim=2)
|
|
dim1 += 1
|
|
if self.bias[1]:
|
|
ones = input2.data.new(batch_size, len2, 1).zero_().fill_(1)
|
|
input2 = torch.cat((input2, Variable(ones)), dim=2)
|
|
dim2 += 1
|
|
|
|
affine = self.linear(input1)
|
|
|
|
affine = affine.view(batch_size, len1 * self.out_features, dim2)
|
|
input2 = torch.transpose(input2, 1, 2)
|
|
# torch.bmm: Performs a batch matrix-matrix product of matrices stored in batch1 and batch2.
|
|
biaffine = torch.transpose(torch.bmm(affine, input2), 1, 2)
|
|
# view: Returns a new tensor with the same data as the self tensor but of a different size.
|
|
biaffine = biaffine.contiguous().view(batch_size, len2, len1, self.out_features)
|
|
|
|
return biaffine
|
|
|
|
def __repr__(self):
|
|
return self.__class__.__name__ + ' (' \
|
|
+ 'in1_features=' + str(self.in1_features) \
|
|
+ ', in2_features=' + str(self.in2_features) \
|
|
+ ', out_features=' + str(self.out_features) + ')'
|
|
|
|
|
|
class HighwayLSTMCell(nn.Module):
|
|
def __init__(self, input_size, hidden_size):
|
|
super(HighwayLSTMCell, self).__init__()
|
|
self.input_size = input_size
|
|
self.hidden_size = hidden_size
|
|
self.linear_ih = nn.Linear(in_features=input_size,
|
|
out_features=6 * hidden_size)
|
|
self.linear_hh = nn.Linear(in_features=hidden_size,
|
|
out_features=5 * hidden_size,
|
|
bias=False)
|
|
self.reset_parameters() # reset all the param in the MyLSTMCell
|
|
|
|
def reset_parameters(self):
|
|
weight_ih = block_orth_normal_initializer([self.input_size, ], [self.hidden_size] * 6)
|
|
self.linear_ih.weight.data.copy_(weight_ih)
|
|
|
|
weight_hh = block_orth_normal_initializer([self.hidden_size, ], [self.hidden_size] * 5)
|
|
self.linear_hh.weight.data.copy_(weight_hh)
|
|
# nn.init.constant(self.linear_hh.weight, 1.0)
|
|
# nn.init.constant(self.linear_ih.weight, 1.0)
|
|
|
|
nn.init.constant(self.linear_ih.bias, 0.0)
|
|
|
|
def forward(self, x, mask=None, hx=None, dropout=None):
|
|
assert mask is not None and hx is not None
|
|
_h, _c = hx
|
|
_x = self.linear_ih(x) # compute the x
|
|
preact = self.linear_hh(_h) + _x[:, :self.hidden_size * 5]
|
|
|
|
i, f, o, t, j = preact.chunk(chunks=5, dim=1)
|
|
i, f, o, t, j = F.sigmoid(i), F.sigmoid(f + 1.0), F.sigmoid(o), F.sigmoid(t), F.tanh(j)
|
|
k = _x[:, self.hidden_size * 5:]
|
|
|
|
c = f * _c + i * j
|
|
c = mask * c + (1.0 - mask) * _c
|
|
|
|
h = t * o * F.tanh(c) + (1.0 - t) * k
|
|
if dropout is not None:
|
|
h = dropout(h)
|
|
h = mask * h + (1.0 - mask) * _h
|
|
return h, c
|
|
|
|
|
|
class VariationalLSTMCell(nn.Module):
|
|
def __init__(self, input_size, hidden_size):
|
|
super(VariationalLSTMCell, self).__init__()
|
|
self.input_size = input_size
|
|
self.hidden_size = hidden_size
|
|
self.linear = nn.Linear(in_features=input_size + self.hidden_size, out_features=3 * hidden_size)
|
|
self.reset_parameters() # reset all the param in the MyLSTMCell
|
|
|
|
def reset_parameters(self):
|
|
weight = block_orth_normal_initializer([self.input_size + self.hidden_size, ], [self.hidden_size] * 3)
|
|
self.linear.weight.data.copy_(weight)
|
|
nn.init.constant_(self.linear.bias, 0.0)
|
|
|
|
def forward(self, x, mask=None, hx=None, dropout=None):
|
|
assert mask is not None and hx is not None
|
|
_h, _c = hx
|
|
_h = dropout(_h)
|
|
_x = self.linear(torch.cat([x, _h], 1)) # compute the x
|
|
i, j, o = _x.chunk(3, dim=1)
|
|
i = torch.sigmoid(i)
|
|
c = (1.0 - i) * _c + i * torch.tanh(j)
|
|
c = mask * c # + (1.0 - mask) * _c
|
|
h = torch.tanh(c) * torch.sigmoid(o)
|
|
h = mask * h # + (1.0 - mask) * _h
|
|
|
|
return h, c
|
|
|
|
|
|
class VariationalLSTM(nn.Module):
|
|
"""A module that runs multiple steps of LSTM."""
|
|
|
|
def __init__(self, input_size, hidden_size, num_layers=1, batch_first=False, \
|
|
bidirectional=False, dropout_in=0, dropout_out=0):
|
|
super(VariationalLSTM, self).__init__()
|
|
self.input_size = input_size
|
|
self.hidden_size = hidden_size
|
|
self.num_layers = num_layers
|
|
self.batch_first = batch_first
|
|
self.bidirectional = bidirectional
|
|
self.dropout_in = dropout_in
|
|
self.dropout_out = dropout_out
|
|
self.num_directions = 2 if bidirectional else 1
|
|
|
|
self.fcells = []
|
|
self.bcells = []
|
|
for layer in range(num_layers):
|
|
layer_input_size = input_size if layer == 0 else hidden_size * self.num_directions
|
|
self.fcells.append(nn.LSTMCell(input_size=layer_input_size, hidden_size=hidden_size))
|
|
if self.bidirectional:
|
|
self.bcells.append(nn.LSTMCell(input_size=layer_input_size, hidden_size=hidden_size))
|
|
|
|
self._all_weights = []
|
|
for layer in range(num_layers):
|
|
layer_params = (self.fcells[layer].weight_ih, self.fcells[layer].weight_hh, \
|
|
self.fcells[layer].bias_ih, self.fcells[layer].bias_hh)
|
|
suffix = ''
|
|
param_names = ['weight_ih_l{}{}', 'weight_hh_l{}{}']
|
|
param_names += ['bias_ih_l{}{}', 'bias_hh_l{}{}']
|
|
param_names = [x.format(layer, suffix) for x in param_names]
|
|
for name, param in zip(param_names, layer_params):
|
|
setattr(self, name, param)
|
|
self._all_weights.append(param_names)
|
|
|
|
if self.bidirectional:
|
|
layer_params = (self.bcells[layer].weight_ih, self.bcells[layer].weight_hh, \
|
|
self.bcells[layer].bias_ih, self.bcells[layer].bias_hh)
|
|
suffix = '_reverse'
|
|
param_names = ['weight_ih_l{}{}', 'weight_hh_l{}{}']
|
|
param_names += ['bias_ih_l{}{}', 'bias_hh_l{}{}']
|
|
param_names = [x.format(layer, suffix) for x in param_names]
|
|
for name, param in zip(param_names, layer_params):
|
|
setattr(self, name, param)
|
|
self._all_weights.append(param_names)
|
|
|
|
self.reset_parameters()
|
|
|
|
def reset_parameters(self): # modified by kiro
|
|
for name, param in self.named_parameters():
|
|
print(name)
|
|
if "weight" in name:
|
|
# for i in range(4):
|
|
# nn.init.orthogonal(self.__getattr__(name)[self.hidden_size*i:self.hidden_size*(i+1),:])
|
|
nn.init.orthogonal(self.__getattr__(name))
|
|
if "bias" in name:
|
|
nn.init.normal(self.__getattr__(name), 0.0, 0.01)
|
|
# nn.init.constant(self.__getattr__(name), 1.0) # different from zhang's 0
|
|
|
|
@staticmethod
|
|
def _forward_rnn(cell, input, masks, initial, drop_masks):
|
|
max_time = input.size(0)
|
|
output = []
|
|
hx = initial
|
|
for time in range(max_time):
|
|
h_next, c_next = cell(input=input[time], hx=hx)
|
|
h_next = h_next * masks[time] + initial[0] * (1 - masks[time])
|
|
c_next = c_next * masks[time] + initial[1] * (1 - masks[time])
|
|
output.append(h_next)
|
|
if drop_masks is not None: h_next = h_next * drop_masks
|
|
hx = (h_next, c_next)
|
|
output = torch.stack(output, 0)
|
|
return output, hx
|
|
|
|
@staticmethod
|
|
def _forward_brnn(cell, input, masks, initial, drop_masks):
|
|
max_time = input.size(0)
|
|
output = []
|
|
hx = initial
|
|
for time in reversed(list(range(max_time))):
|
|
h_next, c_next = cell(input=input[time], hx=hx)
|
|
h_next = h_next * masks[time] + initial[0] * (1 - masks[time])
|
|
c_next = c_next * masks[time] + initial[1] * (1 - masks[time])
|
|
output.append(h_next)
|
|
if drop_masks is not None: h_next = h_next * drop_masks
|
|
hx = (h_next, c_next)
|
|
output.reverse()
|
|
output = torch.stack(output, 0)
|
|
return output, hx
|
|
|
|
def forward(self, input, masks, initial=None):
|
|
if self.batch_first:
|
|
input = input.transpose(0, 1) # transpose: return the transpose matrix
|
|
masks = torch.unsqueeze(masks.transpose(0, 1), dim=2)
|
|
max_time, batch_size, _ = input.size()
|
|
masks = masks.expand(-1, -1, self.hidden_size) # expand: -1 means not expand that dimension
|
|
if initial is None:
|
|
initial = Variable(input.data.new(batch_size, self.hidden_size).zero_())
|
|
initial = (initial, initial) # h0, c0
|
|
h_n = []
|
|
c_n = []
|
|
|
|
for layer in range(self.num_layers):
|
|
max_time, batch_size, input_size = input.size()
|
|
input_mask, hidden_mask = None, None
|
|
if self.training: # when training, use the dropout
|
|
input_mask = input.data.new(batch_size, input_size).fill_(1 - self.dropout_in)
|
|
input_mask = Variable(torch.bernoulli(input_mask), requires_grad=False)
|
|
input_mask = input_mask / (1 - self.dropout_in)
|
|
# permute: exchange the dimension
|
|
input_mask = torch.unsqueeze(input_mask, dim=2).expand(-1, -1, max_time).permute(2, 0, 1)
|
|
input = input * input_mask
|
|
|
|
hidden_mask = input.data.new(batch_size, self.hidden_size).fill_(1 - self.dropout_out)
|
|
hidden_mask = Variable(torch.bernoulli(hidden_mask), requires_grad=False)
|
|
hidden_mask = hidden_mask / (1 - self.dropout_out)
|
|
|
|
layer_output, (layer_h_n, layer_c_n) = VariationalLSTM._forward_rnn(cell=self.fcells[layer], \
|
|
input=input, masks=masks,
|
|
initial=initial,
|
|
drop_masks=hidden_mask)
|
|
if self.bidirectional:
|
|
blayer_output, (blayer_h_n, blayer_c_n) = VariationalLSTM._forward_brnn(cell=self.bcells[layer], \
|
|
input=input, masks=masks,
|
|
initial=initial,
|
|
drop_masks=hidden_mask)
|
|
|
|
h_n.append(torch.cat([layer_h_n, blayer_h_n], 1) if self.bidirectional else layer_h_n)
|
|
c_n.append(torch.cat([layer_c_n, blayer_c_n], 1) if self.bidirectional else layer_c_n)
|
|
input = torch.cat([layer_output, blayer_output], 2) if self.bidirectional else layer_output
|
|
|
|
h_n = torch.stack(h_n, 0)
|
|
c_n = torch.stack(c_n, 0)
|
|
if self.batch_first:
|
|
input = input.transpose(1, 0) # transpose: return the transpose matrix
|
|
return input, (h_n, c_n)
|