def __init__(self, vocab_size, embed_size, time_step, hidden_size, z_dim,
dropout_rate, bos_idx, eos_idx, pad_idx, n_comb):
super(KLPF, self).__init__()
self.time_step = time_step
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.bos_idx = bos_idx
self.eos_idx = eos_idx
self.pad_idx = pad_idx
self.number_combination = n_comb
self.z_dim = z_dim
# encoder
self.encoder = iafEncoder(vocab_size, embed_size,
hidden_size, z_dim, pad_idx)
# iaf
#self.linIAF = linIAF()
#self.combination_L = combination_L()
#self.encoder_y = nn.Linear( hidden_size, self.number_combination )
#self.encoder_L = nn.Linear( hidden_size, (z_dim**2) * self.number_combination )
self.flow = flows.myIAF(z_dim, z_dim, hidden_size,1)
self.softmax = nn.Softmax()
self.classify = nn.Linear(z_dim, 2)
# decoder
self.embedding = nn.Embedding(vocab_size, embed_size,
padding_idx=pad_idx)
self.init_h = nn.Linear(z_dim, hidden_size)
self.init_c = nn.Linear(z_dim, hidden_size)
self.rnn = nn.LSTM(embed_size, hidden_size, batch_first=True)
# output
self.output = nn.Linear(hidden_size, vocab_size)
def __init__(self, config, vocab_size, PAD_token=0):
super(DFVAE, self).__init__()
self.vocab_size = vocab_size
self.maxlen = config['maxlen']
self.clip = config['clip']
self.lambda_gp = config['lambda_gp']
self.temp = config['temp']
self.embedder = nn.Embedding(vocab_size,
config['emb_size'],
padding_idx=PAD_token)
self.utt_encoder = Encoder(self.embedder, config['emb_size'],
config['n_hidden'], True,
config['n_layers'], config['noise_radius'])
self.context_encoder = ContextEncoder(self.utt_encoder,
config['n_hidden'] * 2 + 2,
config['n_hidden'], 1,
config['noise_radius'])
self.prior_net = Variation(config['n_hidden'],
config['z_size']) # p(e|c)
self.post_net = Variation(config['n_hidden'] * 3,
config['z_size']) # q(e|c,x)
#self.prior_highway = nn.Linear(config['n_hidden'], config['n_hidden'])
#self.post_highway = nn.Linear(config['n_hidden'] * 3, config['n_hidden'])
self.postflow1 = flow.myIAF(config['z_size'], config['z_size'] * 2,
config['n_hidden'], 3)
self.postflow2 = flow.myIAF(config['z_size'], config['z_size'] * 2,
config['n_hidden'], 3)
self.postflow3 = flow.myIAF(config['z_size'], config['z_size'] * 2,
config['n_hidden'], 3)
self.priorflow1 = flow.IAF(config['z_size'], config['z_size'] * 2,
config['n_hidden'], 3)
self.priorflow2 = flow.IAF(config['z_size'], config['z_size'] * 2,
config['n_hidden'], 3)
self.priorflow3 = flow.IAF(config['z_size'], config['z_size'] * 2,
config['n_hidden'], 3)
self.post_generator = nn_.SequentialFlow(self.postflow1,
self.postflow2,
self.postflow3)
self.prior_generator = nn_.SequentialFlow(self.priorflow1,
self.priorflow2,
self.priorflow3)
self.decoder = Decoder(self.embedder,
config['emb_size'],
config['n_hidden'] + config['z_size'],
vocab_size,
n_layers=1)
self.optimizer_AE = optim.SGD(
list(self.context_encoder.parameters()) +
list(self.post_net.parameters()) +
list(self.post_generator.parameters()) +
list(self.decoder.parameters()) +
list(self.prior_net.parameters()) +
list(self.prior_generator.parameters())
#+list(self.prior_highway.parameters())
#+list(self.post_highway.parameters())
,
lr=config['lr_ae'])
self.optimizer_G = optim.RMSprop(
list(self.post_net.parameters()) +
list(self.post_generator.parameters()) +
list(self.prior_net.parameters()) +
list(self.prior_generator.parameters())
#+list(self.prior_highway.parameters())
#+list(self.post_highway.parameters())
,
lr=config['lr_gan_g'])
#self.optimizer_D = optim.RMSprop(self.discriminator.parameters(), lr=config['lr_gan_d'])
self.lr_scheduler_AE = optim.lr_scheduler.StepLR(self.optimizer_AE,
step_size=10,
gamma=0.6)
self.criterion_ce = nn.CrossEntropyLoss()