488 lines
18 KiB
Python
488 lines
18 KiB
Python
#
|
|
# Copyright (c) 2018, Salesforce, Inc.
|
|
# The Board of Trustees of the Leland Stanford Junior University
|
|
# All rights reserved.
|
|
#
|
|
# Redistribution and use in source and binary forms, with or without
|
|
# modification, are permitted provided that the following conditions are met:
|
|
#
|
|
# * Redistributions of source code must retain the above copyright notice, this
|
|
# list of conditions and the following disclaimer.
|
|
#
|
|
# * Redistributions in binary form must reproduce the above copyright notice,
|
|
# this list of conditions and the following disclaimer in the documentation
|
|
# and/or other materials provided with the distribution.
|
|
#
|
|
# * Neither the name of the copyright holder nor the names of its
|
|
# contributors may be used to endorse or promote products derived from
|
|
# this software without specific prior written permission.
|
|
#
|
|
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
|
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
|
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
|
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
|
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
|
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
|
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
|
|
import torch
|
|
from torch import nn
|
|
from torch.nn import functional as F
|
|
from torch.autograd import Variable
|
|
import math
|
|
import os
|
|
import sys
|
|
import numpy as np
|
|
import torch.nn as nn
|
|
|
|
from torch.nn.utils.rnn import pad_packed_sequence as unpack
|
|
from torch.nn.utils.rnn import pack_padded_sequence as pack
|
|
|
|
|
|
INF = 1e10
|
|
EPSILON = 1e-10
|
|
|
|
class LSTMDecoder(nn.Module):
|
|
|
|
def __init__(self, num_layers, input_size, rnn_size, dropout):
|
|
super(LSTMDecoder, self).__init__()
|
|
self.dropout = nn.Dropout(dropout)
|
|
self.num_layers = num_layers
|
|
self.layers = nn.ModuleList()
|
|
|
|
for i in range(num_layers):
|
|
self.layers.append(nn.LSTMCell(input_size, rnn_size))
|
|
input_size = rnn_size
|
|
|
|
def forward(self, input, hidden):
|
|
h_0, c_0 = hidden
|
|
h_1, c_1 = [], []
|
|
for i, layer in enumerate(self.layers):
|
|
input = self.dropout(input)
|
|
h_1_i, c_1_i = layer(input, (h_0[i], c_0[i]))
|
|
input = h_1_i
|
|
h_1 += [h_1_i]
|
|
c_1 += [c_1_i]
|
|
|
|
h_1 = torch.stack(h_1)
|
|
c_1 = torch.stack(c_1)
|
|
|
|
return input, (h_1, c_1)
|
|
|
|
def max_margin_loss(probs, targets, pad_idx=1):
|
|
|
|
batch_size, max_length, depth = probs.size()
|
|
targets_mask = (targets != pad_idx).float()
|
|
flat_mask = targets_mask.view(batch_size*max_length,)
|
|
flat_preds = probs.view(batch_size*max_length, depth)
|
|
|
|
one_hot = torch.zeros_like(probs)
|
|
one_hot_gold = one_hot.scatter_(2, targets.unsqueeze(2), 1)
|
|
|
|
marginal_scores = probs - one_hot_gold + 1
|
|
marginal_scores = marginal_scores.view(batch_size*max_length, depth)
|
|
max_margin = torch.max(marginal_scores, dim=1)[0]
|
|
|
|
gold_score = torch.masked_select(flat_preds, one_hot_gold.view(batch_size*max_length, depth).byte())
|
|
margin = max_margin - gold_score
|
|
|
|
return torch.sum(margin*flat_mask) + 1e-8
|
|
|
|
|
|
def positional_encodings_like(x, t=None):
|
|
if t is None:
|
|
positions = torch.arange(0., x.size(1))
|
|
if x.is_cuda:
|
|
positions = positions.cuda(x.get_device())
|
|
else:
|
|
positions = t
|
|
encodings = torch.zeros(*x.size()[1:])
|
|
if x.is_cuda:
|
|
encodings = encodings.cuda(x.get_device())
|
|
for channel in range(x.size(-1)):
|
|
if channel % 2 == 0:
|
|
encodings[:, channel] = torch.sin(
|
|
positions / 10000 ** (channel / x.size(2)))
|
|
else:
|
|
encodings[:, channel] = torch.cos(
|
|
positions / 10000 ** ((channel - 1) / x.size(2)))
|
|
return Variable(encodings)
|
|
|
|
|
|
# torch.matmul can't do (4, 3, 2) @ (4, 2) -> (4, 3)
|
|
def matmul(x, y):
|
|
if x.dim() == y.dim():
|
|
return x @ y
|
|
if x.dim() == y.dim() - 1:
|
|
return (x.unsqueeze(-2) @ y).squeeze(-2)
|
|
return (x @ y.unsqueeze(-2)).squeeze(-2)
|
|
|
|
|
|
def pad_to_match(x, y):
|
|
x_len, y_len = x.size(1), y.size(1)
|
|
if x_len == y_len:
|
|
return x, y
|
|
extra = x.new_ones((x.size(0), abs(y_len - x_len)))
|
|
if x_len < y_len:
|
|
return torch.cat((x, extra), 1), y
|
|
return x, torch.cat((y, extra), 1)
|
|
|
|
|
|
class LayerNorm(nn.Module):
|
|
|
|
def __init__(self, d_model, eps=1e-6):
|
|
super().__init__()
|
|
self.gamma = nn.Parameter(torch.ones(d_model))
|
|
self.beta = nn.Parameter(torch.zeros(d_model))
|
|
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 ResidualBlock(nn.Module):
|
|
|
|
def __init__(self, layer, d_model, dropout_ratio):
|
|
super().__init__()
|
|
self.layer = layer
|
|
self.dropout = nn.Dropout(dropout_ratio)
|
|
self.layernorm = LayerNorm(d_model)
|
|
|
|
def forward(self, *x, padding=None):
|
|
return self.layernorm(x[0] + self.dropout(self.layer(*x, padding=padding)))
|
|
|
|
|
|
class Attention(nn.Module):
|
|
|
|
def __init__(self, d_key, dropout_ratio, causal):
|
|
super().__init__()
|
|
self.scale = math.sqrt(d_key)
|
|
self.dropout = nn.Dropout(dropout_ratio)
|
|
self.causal = causal
|
|
|
|
def forward(self, query, key, value, padding=None):
|
|
dot_products = matmul(query, key.transpose(1, 2))
|
|
if query.dim() == 3 and self.causal:
|
|
tri = key.new_ones((key.size(1), key.size(1))).triu(1) * INF
|
|
dot_products.sub_(tri.unsqueeze(0))
|
|
if not padding is None:
|
|
dot_products.masked_fill_(padding.unsqueeze(1).expand_as(dot_products), -INF)
|
|
return matmul(self.dropout(F.softmax(dot_products / self.scale, dim=-1)), value)
|
|
|
|
|
|
class MultiHead(nn.Module):
|
|
|
|
def __init__(self, d_key, d_value, n_heads, dropout_ratio, causal=False):
|
|
super().__init__()
|
|
self.attention = Attention(d_key, dropout_ratio, causal=causal)
|
|
self.wq = Linear(d_key, d_key, bias=False)
|
|
self.wk = Linear(d_key, d_key, bias=False)
|
|
self.wv = Linear(d_value, d_value, bias=False)
|
|
self.n_heads = n_heads
|
|
|
|
def forward(self, query, key, value, padding=None):
|
|
query, key, value = self.wq(query), self.wk(key), self.wv(value)
|
|
query, key, value = (
|
|
x.chunk(self.n_heads, -1) for x in (query, key, value))
|
|
return torch.cat([self.attention(q, k, v, padding=padding)
|
|
for q, k, v in zip(query, key, value)], -1)
|
|
|
|
|
|
class LinearReLU(nn.Module):
|
|
|
|
def __init__(self, d_model, d_hidden):
|
|
super().__init__()
|
|
self.feedforward = Feedforward(d_model, d_hidden, activation='relu')
|
|
self.linear = Linear(d_hidden, d_model)
|
|
|
|
def forward(self, x, padding=None):
|
|
return self.linear(self.feedforward(x))
|
|
|
|
|
|
class TransformerEncoderLayer(nn.Module):
|
|
|
|
def __init__(self, dimension, n_heads, hidden, dropout):
|
|
super().__init__()
|
|
self.selfattn = ResidualBlock(
|
|
MultiHead(
|
|
dimension, dimension, n_heads, dropout),
|
|
dimension, dropout)
|
|
self.feedforward = ResidualBlock(
|
|
LinearReLU(dimension, hidden),
|
|
dimension, dropout)
|
|
|
|
def forward(self, x, padding=None):
|
|
return self.feedforward(self.selfattn(x, x, x, padding=padding))
|
|
|
|
|
|
class TransformerEncoder(nn.Module):
|
|
|
|
def __init__(self, dimension, n_heads, hidden, num_layers, dropout):
|
|
super().__init__()
|
|
self.layers = nn.ModuleList(
|
|
[TransformerEncoderLayer(dimension, n_heads, hidden, dropout) for i in range(num_layers)])
|
|
self.dropout = nn.Dropout(dropout)
|
|
|
|
def forward(self, x, padding=None):
|
|
x = self.dropout(x)
|
|
encoding = [x]
|
|
for layer in self.layers:
|
|
x = layer(x, padding=padding)
|
|
encoding.append(x)
|
|
return encoding
|
|
|
|
|
|
class TransformerDecoderLayer(nn.Module):
|
|
|
|
def __init__(self, dimension, n_heads, hidden, dropout, causal=True):
|
|
super().__init__()
|
|
self.selfattn = ResidualBlock(
|
|
MultiHead(dimension, dimension, n_heads,
|
|
dropout, causal),
|
|
dimension, dropout)
|
|
self.attention = ResidualBlock(
|
|
MultiHead(dimension, dimension, n_heads,
|
|
dropout),
|
|
dimension, dropout)
|
|
self.feedforward = ResidualBlock(
|
|
LinearReLU(dimension, hidden),
|
|
dimension, dropout)
|
|
|
|
def forward(self, x, encoding, context_padding=None, answer_padding=None):
|
|
x = self.selfattn(x, x, x, padding=answer_padding)
|
|
return self.feedforward(self.attention(x, encoding, encoding, padding=context_padding))
|
|
|
|
|
|
class TransformerDecoder(nn.Module):
|
|
|
|
def __init__(self, dimension, n_heads, hidden, num_layers, dropout, causal=True):
|
|
super().__init__()
|
|
self.layers = nn.ModuleList(
|
|
[TransformerDecoderLayer(dimension, n_heads, hidden, dropout, causal=causal) for i in range(num_layers)])
|
|
self.dropout = nn.Dropout(dropout)
|
|
self.d_model = dimension
|
|
|
|
def forward(self, x, encoding, context_padding=None, positional_encodings=True, answer_padding=None):
|
|
if positional_encodings:
|
|
x = x + positional_encodings_like(x)
|
|
x = self.dropout(x)
|
|
for layer, enc in zip(self.layers, encoding[1:]):
|
|
x = layer(x, enc, context_padding=context_padding, answer_padding=answer_padding)
|
|
return x
|
|
|
|
|
|
def mask(targets, out, squash=True, pad_idx=1):
|
|
mask = (targets != pad_idx)
|
|
out_mask = mask.unsqueeze(-1).expand_as(out).contiguous()
|
|
if squash:
|
|
out_after = out[out_mask].contiguous().view(-1, out.size(-1))
|
|
else:
|
|
out_after = out * out_mask.float()
|
|
targets_after = targets[mask]
|
|
return out_after, targets_after
|
|
|
|
|
|
|
|
|
|
|
|
class Highway(torch.nn.Module):
|
|
def __init__(self, d_in, activation='relu', n_layers=1):
|
|
super(Highway, self).__init__()
|
|
self.d_in = d_in
|
|
self._layers = torch.nn.ModuleList([Linear(d_in, 2 * d_in) for _ in range(n_layers)])
|
|
for layer in self._layers:
|
|
layer.bias[d_in:].fill_(1)
|
|
self.activation = getattr(F, activation)
|
|
|
|
def forward(self, inputs):
|
|
current_input = inputs
|
|
for layer in self._layers:
|
|
projected_input = layer(current_input)
|
|
linear_part = current_input
|
|
nonlinear_part = projected_input[:, :self.d_in] if projected_input.dim() == 2 else projected_input[:, :, :self.d_in]
|
|
nonlinear_part = self.activation(nonlinear_part)
|
|
gate = projected_input[:, self.d_in:(2 * self.d_in)] if projected_input.dim() == 2 else projected_input[:, :, self.d_in:(2 * self.d_in)]
|
|
gate = F.sigmoid(gate)
|
|
current_input = gate * linear_part + (1 - gate) * nonlinear_part
|
|
return current_input
|
|
|
|
|
|
class LinearFeedforward(nn.Module):
|
|
|
|
def __init__(self, d_in, d_hid, d_out, activation='relu'):
|
|
super().__init__()
|
|
self.feedforward = Feedforward(d_in, d_hid, activation=activation)
|
|
self.linear = Linear(d_hid, d_out)
|
|
self.dropout = nn.Dropout(0.2)
|
|
|
|
def forward(self, x):
|
|
return self.dropout(self.linear(self.feedforward(x)))
|
|
|
|
class PackedLSTM(nn.Module):
|
|
|
|
def __init__(self, d_in, d_out, bidirectional=False, num_layers=1,
|
|
dropout=0.0, batch_first=True):
|
|
"""A wrapper class that packs input sequences and unpacks output sequences"""
|
|
super().__init__()
|
|
if bidirectional:
|
|
d_out = d_out // 2
|
|
self.rnn = nn.LSTM(d_in, d_out,
|
|
num_layers=num_layers,
|
|
dropout=dropout,
|
|
bidirectional=bidirectional,
|
|
batch_first=batch_first)
|
|
self.batch_first = batch_first
|
|
|
|
def forward(self, inputs, lengths, hidden=None):
|
|
lens, indices = torch.sort(inputs.new_tensor(lengths, dtype=torch.long), 0, True)
|
|
inputs = inputs[indices] if self.batch_first else inputs[:, indices]
|
|
outputs, (h, c) = self.rnn(pack(inputs, lens.tolist(),
|
|
batch_first=self.batch_first), hidden)
|
|
outputs = unpack(outputs, batch_first=self.batch_first)[0]
|
|
_, _indices = torch.sort(indices, 0)
|
|
outputs = outputs[_indices] if self.batch_first else outputs[:, _indices]
|
|
h, c = h[:, _indices, :], c[:, _indices, :]
|
|
return outputs, (h, c)
|
|
|
|
|
|
class Linear(nn.Linear):
|
|
|
|
def forward(self, x):
|
|
size = x.size()
|
|
return super().forward(
|
|
x.contiguous().view(-1, size[-1])).view(*size[:-1], -1)
|
|
|
|
|
|
class Feedforward(nn.Module):
|
|
|
|
def __init__(self, d_in, d_out, activation=None, bias=True, dropout=0.2):
|
|
super().__init__()
|
|
if activation is not None:
|
|
self.activation = getattr(torch, activation)
|
|
else:
|
|
self.activation = lambda x: x
|
|
self.linear = Linear(d_in, d_out, bias=bias)
|
|
self.dropout = nn.Dropout(dropout)
|
|
|
|
def forward(self, x):
|
|
return self.activation(self.linear(self.dropout(x)))
|
|
|
|
|
|
class Embedding(nn.Module):
|
|
|
|
def __init__(self, field, trained_dimension, dropout=0.0, project=True):
|
|
super().__init__()
|
|
self.field = field
|
|
self.project = project
|
|
dimension = 0
|
|
pretrained_dimension = field.vocab.vectors.size(-1)
|
|
self.pretrained_embeddings = [nn.Embedding(len(field.vocab), pretrained_dimension)]
|
|
self.pretrained_embeddings[0].weight.data = field.vocab.vectors
|
|
self.pretrained_embeddings[0].weight.requires_grad = False
|
|
dimension += pretrained_dimension
|
|
if self.project:
|
|
self.projection = Feedforward(dimension, trained_dimension)
|
|
dimension = trained_dimension
|
|
self.dropout = nn.Dropout(dropout)
|
|
self.dimension = dimension
|
|
|
|
def forward(self, x, lengths=None, device=-1):
|
|
pretrained_embeddings = self.pretrained_embeddings[0](x.cpu()).to(x.device).detach()
|
|
return self.projection(pretrained_embeddings) if self.project else pretrained_embeddings
|
|
|
|
def set_embeddings(self, w):
|
|
self.pretrained_embeddings[0].weight.data = w
|
|
self.pretrained_embeddings[0].weight.requires_grad = False
|
|
|
|
|
|
class SemanticFusionUnit(nn.Module):
|
|
|
|
def __init__(self, d, l):
|
|
super().__init__()
|
|
self.r_hat = Feedforward(d*l, d, 'tanh')
|
|
self.g = Feedforward(d*l, d, 'sigmoid')
|
|
self.dropout = nn.Dropout(0.2)
|
|
|
|
def forward(self, x):
|
|
c = self.dropout(torch.cat(x, -1))
|
|
r_hat = self.r_hat(c)
|
|
g = self.g(c)
|
|
o = g * r_hat + (1 - g) * x[0]
|
|
return o
|
|
|
|
|
|
class LSTMDecoderAttention(nn.Module):
|
|
def __init__(self, dim, dot=False):
|
|
super().__init__()
|
|
self.linear_in = nn.Linear(dim, dim, bias=False)
|
|
self.linear_out = nn.Linear(2 * dim, dim, bias=False)
|
|
self.tanh = nn.Tanh()
|
|
self.mask = None
|
|
self.dot = dot
|
|
|
|
def applyMasks(self, context_mask):
|
|
self.context_mask = context_mask
|
|
|
|
def forward(self, input, context):
|
|
if not self.dot:
|
|
targetT = self.linear_in(input).unsqueeze(2) # batch x dim x 1
|
|
else:
|
|
targetT = input.unsqueeze(2)
|
|
|
|
context_scores = torch.bmm(context, targetT).squeeze(2)
|
|
context_scores.masked_fill_(self.context_mask, -float('inf'))
|
|
context_attention = F.softmax(context_scores, dim=-1) + EPSILON
|
|
context_alignment = torch.bmm(context_attention.unsqueeze(1), context).squeeze(1)
|
|
|
|
combined_representation = torch.cat([input, context_alignment], 1)
|
|
output = self.tanh(self.linear_out(combined_representation))
|
|
|
|
return output, context_attention, context_alignment
|
|
|
|
|
|
class CoattentiveLayer(nn.Module):
|
|
|
|
def __init__(self, d, dropout=0.2):
|
|
super().__init__()
|
|
self.proj = Feedforward(d, d, dropout=0.0)
|
|
self.embed_sentinel = nn.Embedding(2, d)
|
|
self.dropout = nn.Dropout(dropout)
|
|
|
|
def forward(self, context, question, context_padding, question_padding):
|
|
context_padding = torch.cat([context.new_zeros((context.size(0), 1), dtype=torch.long)==1, context_padding], 1)
|
|
question_padding = torch.cat([question.new_zeros((question.size(0), 1), dtype=torch.long)==1, question_padding], 1)
|
|
|
|
context_sentinel = self.embed_sentinel(context.new_zeros((context.size(0), 1), dtype=torch.long))
|
|
context = torch.cat([context_sentinel, self.dropout(context)], 1) # batch_size x (context_length + 1) x features
|
|
|
|
question_sentinel = self.embed_sentinel(question.new_ones((question.size(0), 1), dtype=torch.long))
|
|
question = torch.cat([question_sentinel, question], 1) # batch_size x (question_length + 1) x features
|
|
question = torch.tanh(self.proj(question)) # batch_size x (question_length + 1) x features
|
|
|
|
affinity = context.bmm(question.transpose(1,2)) # batch_size x (context_length + 1) x (question_length + 1)
|
|
attn_over_context = self.normalize(affinity, context_padding) # batch_size x (context_length + 1) x 1
|
|
attn_over_question = self.normalize(affinity.transpose(1,2), question_padding) # batch_size x (question_length + 1) x 1
|
|
sum_of_context = self.attn(attn_over_context, context) # batch_size x (question_length + 1) x features
|
|
sum_of_question = self.attn(attn_over_question, question) # batch_size x (context_length + 1) x features
|
|
coattn_context = self.attn(attn_over_question, sum_of_context) # batch_size x (context_length + 1) x features
|
|
coattn_question = self.attn(attn_over_context, sum_of_question) # batch_size x (question_length + 1) x features
|
|
return torch.cat([coattn_context, sum_of_question], 2)[:, 1:], torch.cat([coattn_question, sum_of_context], 2)[:, 1:]
|
|
|
|
@staticmethod
|
|
def attn(weights, candidates):
|
|
w1, w2, w3 = weights.size()
|
|
c1, c2, c3 = candidates.size()
|
|
return weights.unsqueeze(3).expand(w1, w2, w3, c3).mul(candidates.unsqueeze(2).expand(c1, c2, w3, c3)).sum(1).squeeze(1)
|
|
|
|
@staticmethod
|
|
def normalize(original, padding):
|
|
raw_scores = original.clone()
|
|
raw_scores.masked_fill_(padding.unsqueeze(-1).expand_as(raw_scores), -INF)
|
|
return F.softmax(raw_scores, dim=1)
|
|
|