def __init__(self, x_size, h_size, opt={}, prefix='answer', dropout=None):
super(SANDecoder, self).__init__()
self.prefix = prefix
self.attn_b = FlatSimilarityWrapper(x_size, h_size, prefix, opt,
dropout)
self.attn_e = FlatSimilarityWrapper(x_size, h_size, prefix, opt,
dropout)
self.rnn_type = opt.get('{}_rnn_type'.format(prefix), 'gru')
self.rnn = RNN_MAP.get(self.rnn_type, nn.GRUCell)(x_size, h_size)
self.num_turn = opt.get('{}_num_turn'.format(prefix), 5)
self.opt = opt
self.mem_random_drop = opt.get('{}_mem_drop_p'.format(prefix), 0)
self.answer_opt = opt.get('{}_opt'.format(prefix), 0)
# 0: std mem; 1: random select step; 2 random selection; voting in pred; 3:sort merge
self.mem_type = opt.get('{}_mem_type'.format(prefix), 0)
self.gamma = opt.get('{}_mem_gamma'.format(prefix), 0.5)
self.alpha = Parameter(torch.zeros(1, 1, 1))
self.proj = nn.Linear(h_size, x_size) if h_size != x_size else None
if dropout is None:
self.dropout = DropoutWrapper(
opt.get('{}_dropout_p'.format(self.prefix), 0))
else:
self.dropout = dropout
self.h2h = nn.Linear(h_size, h_size)
self.a2h = nn.Linear(x_size, h_size, bias=False)
self.luong_output_layer = nn.Linear(h_size + x_size, h_size)
def __init__(self, x_size, h_size, opt={}, prefix='answer', dropout=None):
super(SAN, self).__init__()
self.prefix = prefix
# self.attn_b = FlatSimilarityWrapper(x_size, h_size, prefix, opt, dropout)
# self.attn_e = FlatSimilarityWrapper(x_size, h_size, prefix, opt, dropout)
self.attn = FlatSimilarityWrapper(x_size, h_size, prefix, opt, dropout)
self.rnn_type = '{}{}'.format(
opt.get('{}_rnn_type'.format(prefix), 'gru').upper(), 'Cell')
self.rnn = getattr(nn, self.rnn_type)(x_size, h_size)
self.num_turn = opt.get('{}_num_turn'.format(prefix), 5)
# self.attn_startIdx_over_reasoning_steps = Parameter(torch.ones(1,1,self.num_turn).softmax(2)) #initialize with average
# self.attn_endIdx_over_reasoning_steps = Parameter(torch.ones(1,1,self.num_turn).softmax(2)) #initialize with average
self.opt = opt
self.mem_random_drop = opt.get('{}_mem_drop_p'.format(prefix), 0)
self.answer_opt = opt.get('{}_opt'.format(prefix), 0)
self.mem_type = opt.get('{}_mem_type'.format(prefix), 0)
self.gamma = opt.get('{}_mem_gamma'.format(prefix), 0.5)
self.alpha = Parameter(torch.zeros(1, 1, 1))
self.proj = nn.Linear(h_size, x_size) if h_size != x_size else None
if dropout is None:
self.dropout = DropoutWrapper(
opt.get('{}_dropout_p'.format(self.prefix), 0))
else:
self.dropout = dropout
def __init__(self, x_size, h_size, opt={}, prefix='answer', dropout=None):
super(SAN, self).__init__()
self.prefix = prefix
self.attn_b = FlatSimilarityWrapper(x_size, h_size, prefix, opt,
dropout)
self.attn_e = FlatSimilarityWrapper(x_size, h_size, prefix, opt,
dropout)
self.rnn_type = opt.get('{}_rnn_type'.format(prefix), 'gru')
self.rnn = RNN_MAP.get(self.rnn_type, nn.GRUCell)(x_size, h_size)
self.num_turn = opt.get('{}_num_turn'.format(prefix), 5)
self.opt = opt
self.mem_random_drop = opt.get('{}_mem_drop_p'.format(prefix), 0)
self.answer_opt = opt.get('{}_opt'.format(prefix), 0)
self.mem_type = opt.get('{}_mem_type'.format(prefix), 0)
self.gamma = opt.get('{}_mem_gamma'.format(prefix), 0.5)
self.alpha = Parameter(torch.zeros(1, 1, 1))
self.proj = nn.Linear(h_size, x_size) if h_size != x_size else None
if dropout is None:
self.dropout = DropoutWrapper(
opt.get('{}_dropout_p'.format(self.prefix), 0))
else:
self.dropout = dropout