def __init__(self, v_dim, q_dim, num_hid, norm, act, dropout=0.0):
super(Att_3, self).__init__()
norm_layer = get_norm(norm)
self.v_proj = FCNet([v_dim, num_hid], dropout= dropout, norm= norm, act= act)
self.q_proj = FCNet([q_dim, num_hid], dropout= dropout, norm= norm, act= act)
self.nonlinear = FCNet([num_hid, num_hid], dropout= dropout, norm= norm, act= act)
self.linear = norm_layer(nn.Linear(num_hid, 1), dim=None)
def __init__(self, v_dim, q_dim, num_hid, norm, act, dropout=0.0):
super(Att_P, self).__init__()
norm_layer = get_norm(norm)
self.gated_tanh = GTH(in_dim=v_dim + q_dim,
out_dim=num_hid,
dropout=dropout,
norm=norm,
act=act)
self.linear = norm_layer(nn.Linear(num_hid, 1), dim=None)
def __init__(self, v_dim, q_dim, num_hid, norm, act, dropout=0.0):
super(Att_PD, self).__init__()
norm_layer = get_norm(norm)
self.nonlinear = FCNet([v_dim + q_dim, num_hid, num_hid],
dropout=dropout,
norm=norm,
act=act)
self.nonlinear_gate = FCNet([v_dim + q_dim, num_hid, num_hid],
dropout=dropout,
norm=norm,
act='Sigmoid')
self.linear = norm_layer(nn.Linear(num_hid, 1), dim=None)
def __init__(self, in_dim, hid_dim, out_dim, norm, act, dropout=0.5):
super(SimpleClassifier, self).__init__()
norm_layer = get_norm(norm)
act_layer = get_act(act)
layers = [
norm_layer(nn.Linear(in_dim, hid_dim), dim=None),
act_layer(),
nn.Dropout(dropout, inplace=False),
norm_layer(nn.Linear(hid_dim, out_dim), dim=None)
]
self.main = nn.Sequential(*layers)
def __init__(self,
c_dim,
num_hid,
q_dim,
nlayers,
bidirect,
dropout,
rnn_type='LSTM',
v_dim=2048):
"""Module for question embedding
"""
super(CaptionQuestionImageRNN, self).__init__()
assert rnn_type == 'LSTM' or rnn_type == 'GRU'
rnn_cls = nn.LSTM if rnn_type == 'LSTM' else nn.GRU
norm_layer = get_norm('weight')
self.rnn_att = rnn_cls(c_dim,
num_hid,
nlayers,
bidirectional=bidirect,
dropout=dropout,
batch_first=True)
self.rnn_c = rnn_cls(c_dim,
num_hid,
nlayers,
bidirectional=bidirect,
dropout=dropout,
batch_first=True)
self.q_emb_for_c = FCNet([q_dim, num_hid],
dropout=0.2,
norm='weight',
act='LeakyReLU')
self.att_logits = norm_layer(nn.Linear(num_hid, 1), dim=None)
self.v_emb_for_c = FCNet([v_dim, num_hid],
dropout=0.2,
norm='weight',
act='LeakyReLU')
self.v_att_logits = norm_layer(nn.Linear(num_hid, 1), dim=None)
self.Sig = nn.Sigmoid()
self.c_dim = c_dim
self.q_dim = q_dim
self.num_hid = num_hid
self.nlayers = nlayers
self.ndirections = int(bidirect) + 1
self.rnn_type = rnn_type
self.v_dim = v_dim
def __init__(self,
v_dim,
q_dim,
num_hid,
norm,
act,
bidirect=False,
dropout=0.0):
super(Base_Att, self).__init__()
norm_layer = get_norm(norm)
if (bidirect is False):
self.nonlinear = FCNet([v_dim + q_dim, num_hid],
dropout=dropout,
norm=norm,
act=act)
else:
self.nonlinear = FCNet([v_dim + 2 * q_dim, num_hid],
dropout=dropout,
norm=norm,
act=act)
self.linear = norm_layer(nn.Linear(num_hid, 1), dim=None)
