Spaces:
Runtime error
Runtime error
File size: 12,113 Bytes
c5ca37a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 |
# import torch
import time
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils.rnn import pad_packed_sequence, pack_padded_sequence
import numpy as np
from .decoder import DecoderBase
class LSTMDecoder(DecoderBase):
"""LSTM decoder with constant-length data"""
def __init__(self, args, vocab, model_init, emb_init):
super(LSTMDecoder, self).__init__()
self.ni = args.ni
self.nh = args.dec_nh
self.nz = args.nz
self.vocab = vocab
self.device = args.device
# no padding when setting padding_idx to -1
self.embed = nn.Embedding(len(vocab), args.ni, padding_idx=-1)
self.dropout_in = nn.Dropout(args.dec_dropout_in)
self.dropout_out = nn.Dropout(args.dec_dropout_out)
# for initializing hidden state and cell
self.trans_linear = nn.Linear(args.nz, args.dec_nh, bias=False)
# concatenate z with input
self.lstm = nn.LSTM(input_size=args.ni + args.nz,
hidden_size=args.dec_nh,
num_layers=1,
batch_first=True)
# prediction layer
self.pred_linear = nn.Linear(args.dec_nh, len(vocab), bias=False)
vocab_mask = torch.ones(len(vocab))
# vocab_mask[vocab['<pad>']] = 0
self.loss = nn.CrossEntropyLoss(weight=vocab_mask, reduce=False)
self.reset_parameters(model_init, emb_init)
def reset_parameters(self, model_init, emb_init):
# for name, param in self.lstm.named_parameters():
# # self.initializer(param)
# if 'bias' in name:
# nn.init.constant_(param, 0.0)
# # model_init(param)
# elif 'weight' in name:
# model_init(param)
# model_init(self.trans_linear.weight)
# model_init(self.pred_linear.weight)
for param in self.parameters():
model_init(param)
emb_init(self.embed.weight)
def sample_text(self, input, z, EOS, device):
sentence = [input]
max_index = 0
input_word = input
batch_size, n_sample, _ = z.size()
seq_len = 1
z_ = z.expand(batch_size, seq_len, self.nz)
seq_len = input.size(1)
softmax = torch.nn.Softmax(dim=0)
while max_index != EOS and len(sentence) < 100:
# (batch_size, seq_len, ni)
word_embed = self.embed(input_word)
word_embed = torch.cat((word_embed, z_), -1)
c_init = self.trans_linear(z).unsqueeze(0)
h_init = torch.tanh(c_init)
if len(sentence) == 1:
h_init = h_init.squeeze(dim=1)
c_init = c_init.squeeze(dim=1)
output, hidden = self.lstm.forward(word_embed, (h_init, c_init))
else:
output, hidden = self.lstm.forward(word_embed, hidden)
# (batch_size * n_sample, seq_len, vocab_size)
output_logits = self.pred_linear(output)
output_logits = output_logits.view(-1)
probs = softmax(output_logits)
# max_index = torch.argmax(output_logits)
max_index = torch.multinomial(probs, num_samples=1)
input_word = torch.tensor([[max_index]]).to(device)
sentence.append(max_index)
return sentence
def decode(self, input, z):
"""
Args:
input: (batch_size, seq_len)
z: (batch_size, n_sample, nz)
"""
# not predicting start symbol
# sents_len -= 1
batch_size, n_sample, _ = z.size()
seq_len = input.size(1)
# (batch_size, seq_len, ni)
word_embed = self.embed(input)
word_embed = self.dropout_in(word_embed)
if n_sample == 1:
z_ = z.expand(batch_size, seq_len, self.nz)
else:
word_embed = word_embed.unsqueeze(1).expand(batch_size, n_sample, seq_len, self.ni) \
.contiguous()
# (batch_size * n_sample, seq_len, ni)
word_embed = word_embed.view(batch_size * n_sample, seq_len, self.ni)
z_ = z.unsqueeze(2).expand(batch_size, n_sample, seq_len, self.nz).contiguous()
z_ = z_.view(batch_size * n_sample, seq_len, self.nz)
# (batch_size * n_sample, seq_len, ni + nz)
word_embed = torch.cat((word_embed, z_), -1)
z = z.view(batch_size * n_sample, self.nz)
c_init = self.trans_linear(z).unsqueeze(0)
h_init = torch.tanh(c_init)
# h_init = self.trans_linear(z).unsqueeze(0)
# c_init = h_init.new_zeros(h_init.size())
output, _ = self.lstm(word_embed, (h_init, c_init))
output = self.dropout_out(output)
# (batch_size * n_sample, seq_len, vocab_size)
output_logits = self.pred_linear(output)
return output_logits
def reconstruct_error(self, x, z):
"""Cross Entropy in the language case
Args:
x: (batch_size, seq_len)
z: (batch_size, n_sample, nz)
Returns:
loss: (batch_size, n_sample). Loss
across different sentence and z
"""
#remove end symbol
src = x[:, :-1]
# remove start symbol
tgt = x[:, 1:]
batch_size, seq_len = src.size()
n_sample = z.size(1)
# (batch_size * n_sample, seq_len, vocab_size)
output_logits = self.decode(src, z)
if n_sample == 1:
tgt = tgt.contiguous().view(-1)
else:
# (batch_size * n_sample * seq_len)
tgt = tgt.unsqueeze(1).expand(batch_size, n_sample, seq_len) \
.contiguous().view(-1)
# (batch_size * n_sample * seq_len)
loss = self.loss(output_logits.view(-1, output_logits.size(2)),
tgt)
# (batch_size, n_sample)
return loss.view(batch_size, n_sample, -1).sum(-1)
def log_probability(self, x, z):
"""Cross Entropy in the language case
Args:
x: (batch_size, seq_len)
z: (batch_size, n_sample, nz)
Returns:
log_p: (batch_size, n_sample).
log_p(x|z) across different x and z
"""
return -self.reconstruct_error(x, z)
def greedy_decode(self, z):
return self.sample_decode(z, greedy=True)
def sample_decode(self, z, greedy=False):
"""sample/greedy decoding from z
Args:
z: (batch_size, nz)
Returns: List1
List1: the decoded word sentence list
"""
batch_size = z.size(0)
decoded_batch = [[] for _ in range(batch_size)]
# (batch_size, 1, nz)
c_init = self.trans_linear(z).unsqueeze(0)
h_init = torch.tanh(c_init)
decoder_hidden = (h_init, c_init)
decoder_input = torch.tensor([self.vocab["<s>"]] * batch_size, dtype=torch.long, device=self.device).unsqueeze(1)
end_symbol = torch.tensor([self.vocab["</s>"]] * batch_size, dtype=torch.long, device=self.device)
mask = torch.ones((batch_size), dtype=torch.uint8, device=self.device)
length_c = 1
while mask.sum().item() != 0 and length_c < 100:
# (batch_size, 1, ni) --> (batch_size, 1, ni+nz)
word_embed = self.embed(decoder_input)
word_embed = torch.cat((word_embed, z.unsqueeze(1)), dim=-1)
output, decoder_hidden = self.lstm(word_embed, decoder_hidden)
# (batch_size, 1, vocab_size) --> (batch_size, vocab_size)
decoder_output = self.pred_linear(output)
output_logits = decoder_output.squeeze(1)
# (batch_size)
if greedy:
max_index = torch.argmax(output_logits, dim=1)
else:
probs = F.softmax(output_logits, dim=1)
max_index = torch.multinomial(probs, num_samples=1).squeeze(1)
decoder_input = max_index.unsqueeze(1)
length_c += 1
for i in range(batch_size):
word = self.vocab.id2word(max_index[i].item())
if mask[i].item():
decoded_batch[i].append(self.vocab.id2word(max_index[i].item()))
mask = torch.mul((max_index != end_symbol), mask)
return decoded_batch
class VarLSTMDecoder(LSTMDecoder):
"""LSTM decoder with constant-length data"""
def __init__(self, args, vocab, model_init, emb_init):
super(VarLSTMDecoder, self).__init__(args, vocab, model_init, emb_init)
self.embed = nn.Embedding(len(vocab), args.ni, padding_idx=vocab['<pad>'])
vocab_mask = torch.ones(len(vocab))
vocab_mask[vocab['<pad>']] = 0
self.loss = nn.CrossEntropyLoss(weight=vocab_mask, reduce=False)
self.reset_parameters(model_init, emb_init)
def decode(self, input, z):
"""
Args:
input: tuple which contains x and sents_len
x: (batch_size, seq_len)
sents_len: long tensor of sentence lengths
z: (batch_size, n_sample, nz)
"""
input, sents_len = input
# not predicting start symbol
sents_len = sents_len - 1
batch_size, n_sample, _ = z.size()
seq_len = input.size(1)
# (batch_size, seq_len, ni)
word_embed = self.embed(input)
word_embed = self.dropout_in(word_embed)
if n_sample == 1:
z_ = z.expand(batch_size, seq_len, self.nz)
else:
word_embed = word_embed.unsqueeze(1).expand(batch_size, n_sample, seq_len, self.ni) \
.contiguous()
# (batch_size * n_sample, seq_len, ni)
word_embed = word_embed.view(batch_size * n_sample, seq_len, self.ni)
z_ = z.unsqueeze(2).expand(batch_size, n_sample, seq_len, self.nz).contiguous()
z_ = z_.view(batch_size * n_sample, seq_len, self.nz)
# (batch_size * n_sample, seq_len, ni + nz)
word_embed = torch.cat((word_embed, z_), -1)
sents_len = sents_len.unsqueeze(1).expand(batch_size, n_sample).contiguous().view(-1)
packed_embed = pack_padded_sequence(word_embed, sents_len.tolist(), batch_first=True)
z = z.view(batch_size * n_sample, self.nz)
# h_init = self.trans_linear(z).unsqueeze(0)
# c_init = h_init.new_zeros(h_init.size())
c_init = self.trans_linear(z).unsqueeze(0)
h_init = torch.tanh(c_init)
output, _ = self.lstm(packed_embed, (h_init, c_init))
output, _ = pad_packed_sequence(output, batch_first=True)
output = self.dropout_out(output)
# (batch_size * n_sample, seq_len, vocab_size)
output_logits = self.pred_linear(output)
return output_logits
def reconstruct_error(self, x, z):
"""Cross Entropy in the language case
Args:
x: tuple which contains x_ and sents_len
x_: (batch_size, seq_len)
sents_len: long tensor of sentence lengths
z: (batch_size, n_sample, nz)
Returns:
loss: (batch_size, n_sample). Loss
across different sentence and z
"""
x, sents_len = x
#remove end symbol
src = x[:, :-1]
# remove start symbol
tgt = x[:, 1:]
batch_size, seq_len = src.size()
n_sample = z.size(1)
# (batch_size * n_sample, seq_len, vocab_size)
output_logits = self.decode((src, sents_len), z)
if n_sample == 1:
tgt = tgt.contiguous().view(-1)
else:
# (batch_size * n_sample * seq_len)
tgt = tgt.unsqueeze(1).expand(batch_size, n_sample, seq_len) \
.contiguous().view(-1)
# (batch_size * n_sample * seq_len)
loss = self.loss(output_logits.view(-1, output_logits.size(2)),
tgt)
# (batch_size, n_sample)
return loss.view(batch_size, n_sample, -1).sum(-1) |