def __init__(self, vocab_size, embed_dim, embed_init, max_nsent, max_npara,
experiment, *args, **kwargs):
super(quantize_pos_model, self).__init__(vocab_size, embed_dim,
embed_init, experiment)
if self.expe.config.encoder_type.lower() == "lstm" or \
self.expe.config.encoder_type.lower() == "gru":
ensize = 2 * self.expe.config.ensize
else:
ensize = embed_dim
self.sent_id_pred = model_utils.get_mlp(
input_size=ensize,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.nb,
n_layer=self.expe.config.mlplayer,
dropout=self.expe.config.dp)
self.para_id_pred = model_utils.get_mlp(
input_size=ensize,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.nb,
n_layer=self.expe.config.mlplayer,
dropout=self.expe.config.dp)
self.bins = np.arange(self.expe.config.nb) / self.expe.config.nb
def __init__(self, vocab_size, embed_dim, embed_init, max_nsent, max_npara,
max_nlv, doc_title_vocab_size, sec_title_vocab_size,
experiment, *args, **kwargs):
super(quantize_pos_regression_model,
self).__init__(vocab_size, embed_dim, embed_init, experiment)
if self.expe.config.encoder_type.lower() in [
"lstm", "gru", "gru_attn"
]:
ensize = 2 * self.expe.config.ensize
else:
ensize = embed_dim
self.sent_id_pred = model_utils.get_mlp(
input_size=ensize,
hidden_size=self.expe.config.mhsize,
output_size=1,
n_layer=self.expe.config.mlplayer,
dropout=self.expe.config.dp)
self.para_id_pred = model_utils.get_mlp(
input_size=ensize,
hidden_size=self.expe.config.mhsize,
output_size=1,
n_layer=self.expe.config.mlplayer,
dropout=self.expe.config.dp)
self.lv_pred = model_utils.get_mlp(input_size=ensize,
hidden_size=self.expe.config.mhsize,
output_size=max_nlv,
n_layer=self.expe.config.mlplayer,
dropout=self.expe.config.dp)
self.doc_title_decode = decoders.bag_of_words(
input_size=ensize,
mlp_hidden_size=self.expe.config.mhsize,
mlp_layer=self.expe.config.mlplayer,
hidden_size=self.expe.config.desize,
dropout=self.expe.config.dp,
embed_dim=embed_dim,
tie_weight=self.expe.config.tw,
word_dropout=self.expe.config.wd,
embed_init=embed_init,
vocab_size=doc_title_vocab_size,
log=experiment.log)
self.sec_title_decode = decoders.bag_of_words(
input_size=ensize,
mlp_hidden_size=self.expe.config.mhsize,
mlp_layer=self.expe.config.mlplayer,
hidden_size=self.expe.config.desize,
dropout=self.expe.config.dp,
embed_dim=embed_dim,
tie_weight=self.expe.config.tw,
word_dropout=self.expe.config.wd,
embed_init=embed_init,
vocab_size=sec_title_vocab_size,
log=experiment.log)
self.bins = np.arange(self.expe.config.nb) / self.expe.config.nb
def __init__(self, vocab_size, input_size, mlp_hidden_size, mlp_layer,
hidden_size, embed_dim, dropout, tie_weight, word_dropout,
log, embed_init, *args, **kwargs):
super(bag_of_words, self).__init__()
self.func = model_utils.get_mlp(input_size, mlp_hidden_size,
hidden_size, mlp_layer, dropout)
self.hid2vocab = nn.Linear(hidden_size, vocab_size)
def __init__(self, ysize, zsize, vocab_size, mlp_hidden_size, mlp_layer,
hidden_size, dropout, *args, **kwargs):
super(lstm, self).__init__()
self.cell = nn.LSTM(zsize,
hidden_size,
bidirectional=False,
batch_first=True)
self.hid2vocab = model_utils.get_mlp(hidden_size + ysize, hidden_size,
vocab_size, mlp_layer, dropout)
def __init__(self, vocab_size, embed_dim, embed_init, hidden_size, dropout,
nclass, ncode, mlp_hidden_size, mlp_layer, log, *args,
**kwargs):
super(bilstm_lc, self).__init__(vocab_size, embed_dim, embed_init,
dropout, log)
self.lstm = nn.LSTM(embed_dim // ncode * ncode,
hidden_size,
bidirectional=True,
batch_first=True)
self.lc = nn.ModuleList([
model_utils.get_mlp(embed_dim, hidden_size, nclass, mlp_layer,
dropout) for _ in range(ncode)
])
self.lc_embed = nn.ModuleList(
[nn.Embedding(nclass, embed_dim // ncode) for _ in range(ncode)])
def __init__(self, vocab_size, input_size, mlp_hidden_size, mlp_layer,
hidden_size, embed_dim, dropout, tie_weight, word_dropout,
log, embed_init, *args, **kwargs):
super(lstm, self).__init__(vocab_size, input_size, mlp_hidden_size,
mlp_layer, hidden_size, embed_dim, dropout,
tie_weight, word_dropout, log, embed_init,
*args, **kwargs)
self.latent2hid = model_utils.get_mlp(input_size, mlp_hidden_size,
hidden_size * 2, mlp_layer,
dropout)
self.cell = nn.LSTM(embed_dim,
hidden_size,
bidirectional=False,
batch_first=True)
def __init__(self, vocab_size, input_size, mlp_hidden_size, mlp_layer,
hidden_size, embed_dim, dropout, tie_weight, word_dropout,
log, embed_init, *args, **kwargs):
super(gru_cat, self).__init__(vocab_size, input_size, mlp_hidden_size,
mlp_layer, hidden_size, embed_dim,
dropout, tie_weight, word_dropout, log,
embed_init, *args, **kwargs)
self.latent2hid = model_utils.get_mlp(input_size, mlp_hidden_size,
hidden_size, mlp_layer, dropout)
self.cell = nn.GRU(embed_dim + input_size,
hidden_size,
bidirectional=False,
batch_first=True)
if not tie_weight:
self.hid2vocab = nn.Linear(hidden_size, vocab_size)
def __init__(self, vocab_size, embed_dim, embed_init, experiment):
super(vgvae, self).__init__(vocab_size, embed_dim, embed_init,
experiment)
self.yencode = getattr(encoders, self.expe.config.yencoder_type)(
embed_dim=embed_dim,
embed_init=embed_init,
hidden_size=self.expe.config.ensize,
vocab_size=vocab_size,
dropout=self.expe.config.dp,
log=experiment.log)
self.zencode = getattr(encoders, self.expe.config.zencoder_type)(
embed_dim=embed_dim,
embed_init=embed_init,
hidden_size=self.expe.config.ensize,
vocab_size=vocab_size,
dropout=self.expe.config.dp,
mlp_hidden_size=self.expe.config.mhsize,
mlp_layer=self.expe.config.mlplayer,
ncode=self.expe.config.ncode,
nclass=self.expe.config.nclass,
log=experiment.log)
if "lstm" in self.expe.config.yencoder_type.lower():
y_out_size = 2 * self.expe.config.ensize
elif self.expe.config.yencoder_type.lower() == "word_avg":
y_out_size = embed_dim
if "lstm" in self.expe.config.zencoder_type.lower():
z_out_size = 2 * self.expe.config.ensize
elif self.expe.config.zencoder_type.lower() == "word_avg":
z_out_size = embed_dim
self.mean1 = model_utils.get_mlp(input_size=y_out_size,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.ysize,
n_layer=self.expe.config.ymlplayer,
dropout=self.expe.config.dp)
self.logvar1 = model_utils.get_mlp(input_size=y_out_size,
hidden_size=self.expe.config.mhsize,
output_size=1,
n_layer=self.expe.config.ymlplayer,
dropout=self.expe.config.dp)
self.mean2 = model_utils.get_mlp(input_size=z_out_size,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.zsize,
n_layer=self.expe.config.zmlplayer,
dropout=self.expe.config.dp)
self.logvar2 = model_utils.get_mlp(input_size=z_out_size,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.zsize,
n_layer=self.expe.config.zmlplayer,
dropout=self.expe.config.dp)
self.decode = getattr(decoders, self.expe.config.decoder_type)(
embed_init=embed_init,
embed_dim=embed_dim,
ysize=self.expe.config.ysize,
zsize=self.expe.config.zsize,
mlp_hidden_size=self.expe.config.mhsize,
mlp_layer=self.expe.config.mlplayer,
hidden_size=self.expe.config.desize,
dropout=self.expe.config.dp,
vocab_size=vocab_size,
log=experiment.log)
if "lc" in self.expe.config.zencoder_type.lower():
enc_embed_dim = embed_dim // self.expe.config.ncode *\
self.expe.config.ncode
else:
enc_embed_dim = embed_dim
self.enc_pos_decode = model_utils.get_mlp(
input_size=self.expe.config.zsize + enc_embed_dim,
hidden_size=self.expe.config.mhsize,
n_layer=self.expe.config.mlplayer,
output_size=MAX_LEN,
dropout=self.expe.config.dp)
self.dec_pos_decode = model_utils.get_mlp(
input_size=self.expe.config.zsize + embed_dim,
hidden_size=self.expe.config.mhsize,
n_layer=self.expe.config.mlplayer,
output_size=MAX_LEN,
dropout=self.expe.config.dp)
def __init__(self, vocab_size, embed_dim, embed_init, experiment):
super(base, self).__init__()
self.expe = experiment
self.eps = self.expe.config.eps
self.margin = self.expe.config.m
self.use_cuda = self.expe.config.use_cuda
self.yencode = getattr(encoders, self.expe.config.yencoder_type)(
embed_dim=embed_dim,
embed_init=embed_init,
hidden_size=self.expe.config.ensize,
vocab_size=vocab_size,
log=experiment.log)
self.zencode = getattr(encoders, self.expe.config.zencoder_type)(
embed_dim=embed_dim,
embed_init=embed_init,
hidden_size=self.expe.config.ensize,
vocab_size=vocab_size,
log=experiment.log)
if "lstm" in self.expe.config.yencoder_type.lower():
y_out_size = 2 * self.expe.config.ensize
elif self.expe.config.yencoder_type.lower() == "word_avg":
y_out_size = embed_dim
elif self.expe.config.yencoder_type.lower() == "bert":
y_out_size = embed_dim
if "lstm" in self.expe.config.zencoder_type.lower():
z_out_size = 2 * self.expe.config.ensize
elif self.expe.config.zencoder_type.lower() == "word_avg":
z_out_size = embed_dim
elif self.expe.config.zencoder_type.lower() == "bert":
z_out_size = embed_dim
self.mean1 = model_utils.get_mlp(input_size=y_out_size,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.ysize,
n_layer=self.expe.config.ymlplayer,
dropout=self.expe.config.dp)
self.logvar1 = model_utils.get_mlp(input_size=y_out_size,
hidden_size=self.expe.config.mhsize,
output_size=1,
n_layer=self.expe.config.ymlplayer,
dropout=self.expe.config.dp)
self.mean2 = model_utils.get_mlp(input_size=z_out_size,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.zsize,
n_layer=self.expe.config.zmlplayer,
dropout=self.expe.config.dp)
self.logvar2 = model_utils.get_mlp(input_size=z_out_size,
hidden_size=self.expe.config.mhsize,
output_size=self.expe.config.zsize,
n_layer=self.expe.config.zmlplayer,
dropout=self.expe.config.dp)
if self.expe.config.zencoder_type.lower() == "word_avg":
assert self.expe.config.decoder_type.lower() == "bag_of_words"
self.decode = getattr(decoders, self.expe.config.decoder_type)(
ysize=self.expe.config.ysize,
zsize=self.expe.config.zsize,
mlp_hidden_size=self.expe.config.mhsize,
mlp_layer=self.expe.config.mlplayer,
hidden_size=self.expe.config.desize,
dropout=self.expe.config.dp,
vocab_size=vocab_size)
self.pos_decode = model_utils.get_mlp(
input_size=self.expe.config.zsize + embed_dim,
hidden_size=self.expe.config.mhsize,
n_layer=self.expe.config.mlplayer,
output_size=MAX_LEN,
dropout=self.expe.config.dp)
def __init__(self, ysize, zsize, mlp_layer, hidden_size, vocab_size,
dropout, *args, **kwargs):
super(bag_of_words, self).__init__()
self.hid2vocab = model_utils.get_mlp(ysize + zsize, hidden_size,
vocab_size, mlp_layer, dropout)