] + itos #+ itos_lemmas[:vocab_size] + itos_morph
assert len(itos_total) == outVocabSize
# could also provide per-word subcategorization frames from the treebank as input???
#baseline = nn.Linear(args.emb_dim, 1).cuda()
dropout = nn.Dropout(args.dropout_rate).cuda()
rnn_both = nn.GRU(2 * args.emb_dim, args.rnn_dim, args.rnn_layers).cuda()
for name, param in rnn_both.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.xavier_normal(param)
decoder = nn.Linear(args.rnn_dim, outVocabSize).cuda()
#pos_ptb_decoder = nn.Linear(128,len(posFine)+3).cuda()
startHidden = nn.Linear(1, args.rnn_dim).cuda()
startHidden.bias.data.fill_(0)
components = [rnn_both, decoder, word_pos_morph_embeddings, startHidden]
hiddenToLogSDHidden = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
cellToMean = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
sampleToHidden = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
hiddenToLogSDHidden.bias.data.fill_(0)
cellToMean.bias.data.fill_(0)
sampleToHidden.bias.data.fill_(0)
] + itos #+ itos_lemmas[:vocab_size] + itos_morph
assert len(itos_total) == outVocabSize
# could also provide per-word subcategorization frames from the treebank as input???
#baseline = nn.Linear(args.emb_dim, 1).cuda()
dropout = nn.Dropout(args.dropout_rate).cuda()
rnn_both = nn.GRU(2 * args.emb_dim, args.rnn_dim, args.rnn_layers).cuda()
for name, param in rnn_both.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.xavier_normal(param)
decoder = nn.Linear(args.rnn_dim, outVocabSize).cuda()
#pos_ptb_decoder = nn.Linear(128,len(posFine)+3).cuda()
startHidden = nn.Linear(1, args.rnn_dim).cuda()
components = [rnn_both, decoder, word_pos_morph_embeddings, startHidden]
import torchkit.nn as nn_
character_embeddings = torch.nn.Embedding(
num_embeddings=len(itos_chars_total) + 3,
embedding_dim=args.char_emb_dim).cuda()
char_composition = torch.nn.LSTM(args.char_emb_dim,
args.char_enc_hidden_dim,
1,
embedding_dim=args.emb_dim).cuda()
outVocabSize = 2 * (3 + vocab_size)
itos_total = ["EOS", "OOV", "SOS"] + itos_1 + ["EOS", "OOV", "SOS"] + itos_2
assert len(itos_total) == outVocabSize
dropout = nn.Dropout(args.dropout_rate).cuda()
rnn_both = nn.GRU(2 * args.emb_dim, args.rnn_dim, args.rnn_layers).cuda()
for name, param in rnn_both.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.xavier_normal(param)
decoder_1 = nn.Linear(args.rnn_dim, 50003).cuda()
decoder_2 = nn.Linear(args.rnn_dim, 50003).cuda()
#pos_ptb_decoder = nn.Linear(128,len(posFine)+3).cuda()
startHidden = nn.Linear(1, args.rnn_dim).cuda()
components = [
rnn_both, decoder_1, decoder_2, word_pos_morph_embeddings, startHidden
]
hiddenToLogSDHidden = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
cellToMean = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
sampleToHidden = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
import torchkit.nn as nn_
assert len(itos_total) == outVocabSize
# could also provide per-word subcategorization frames from the treebank as input???
#baseline = nn.Linear(args.emb_dim, 1).cuda()
dropout = nn.Dropout(args.dropout_rate).cuda()
rnn_both = nn.GRU(args.emb_dim, args.rnn_dim, args.rnn_layers).cuda()
for name, param in rnn_both.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.xavier_normal(param)
decoder = nn.Linear(args.rnn_dim,outVocabSize).cuda()
#pos_ptb_decoder = nn.Linear(128,len(posFine)+3).cuda()
components = [rnn_both, decoder, word_pos_morph_embeddings]
# klLoss = [None for _ in inputEmbeddings]
# logStandardDeviationHidden = hiddenToLogSDHidden(hidden[1][0])
# sampled = torch.normal(hiddenMean, torch.exp(logStandardDeviationHidden))
# klLoss = 0.5 * (-1 - 2 * (logStandardDeviationHidden) + torch.pow(meanHidden, 2) + torch.exp(2*logStandardDeviationHidden))
# hiddenNew = sampleToHidden(sampled)
# cellNew = sampleToCell(sampled)
] + itos #+ itos_lemmas[:vocab_size] + itos_morph
assert len(itos_total) == outVocabSize
# could also provide per-word subcategorization frames from the treebank as input???
#baseline = nn.Linear(args.emb_dim, 1).cuda()
dropout = nn.Dropout(args.dropout_rate).cuda()
rnn_both = nn.GRU(2 * args.emb_dim, args.rnn_dim, args.rnn_layers).cuda()
for name, param in rnn_both.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.xavier_normal(param)
decoder = nn.Linear(args.rnn_dim, outVocabSize).cuda()
#pos_ptb_decoder = nn.Linear(128,len(posFine)+3).cuda()
startHidden = nn.Linear(1, args.rnn_dim).cuda()
startHidden.bias.data.fill_(0)
components = [rnn_both, decoder, word_pos_morph_embeddings, startHidden]
char_embeddings = torch.nn.Embedding(num_embeddings=len(itos_chars_total) + 3,
embedding_dim=args.char_emb_dim)
char_composition = nn.GRU(args.char_emb_dim,
args.char_rnn_dim,
1,
bidirectional=True).cuda()
char_composition_output = nn.Linear(2 * args.char_rnn_dim, args.emb_dim)
] + itos #+ itos_lemmas[:vocab_size] + itos_morph
assert len(itos_total) == outVocabSize
# could also provide per-word subcategorization frames from the treebank as input???
#baseline = nn.Linear(args.emb_dim, 1).cuda()
dropout = nn.Dropout(args.dropout_rate).cuda()
rnn_both = nn.GRU(args.emb_dim, args.rnn_dim, args.rnn_layers).cuda()
for name, param in rnn_both.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.xavier_normal(param)
decoder = nn.Linear(args.rnn_dim, outVocabSize).cuda()
#pos_ptb_decoder = nn.Linear(128,len(posFine)+3).cuda()
components = [rnn_both, decoder, word_pos_morph_embeddings]
# klLoss = [None for _ in inputEmbeddings]
# logStandardDeviationHidden = hiddenToLogSDHidden(hidden[1][0])
# sampled = torch.normal(hiddenMean, torch.exp(logStandardDeviationHidden))
# klLoss = 0.5 * (-1 - 2 * (logStandardDeviationHidden) + torch.pow(meanHidden, 2) + torch.exp(2*logStandardDeviationHidden))
# hiddenNew = sampleToHidden(sampled)
# cellNew = sampleToCell(sampled)
hiddenToLogSDHidden = nn.Linear(args.rnn_dim, args.rnn_dim).cuda()
hiddenToLogSDHidden.weight.data.fill_(0)
hiddenToLogSDHidden.bias.data.fill_(0)
