def __init__(self,
img_dim,
embed_size,
data_name,
use_abs=False,
no_imgnorm=False):
super(EncoderImagePrecompAttn, self).__init__()
self.embed_size = embed_size
self.no_imgnorm = no_imgnorm
self.data_name = data_name
self.use_abs = use_abs
self.fc = nn.Linear(img_dim, embed_size)
self.init_weights()
# GSR
self.img_rnn = nn.GRU(embed_size, embed_size, 1, batch_first=True)
# GCN reasoning
self.Rs_GCN_1 = Rs_GCN(in_channels=embed_size,
inter_channels=embed_size)
self.Rs_GCN_2 = Rs_GCN(in_channels=embed_size,
inter_channels=embed_size)
self.Rs_GCN_3 = Rs_GCN(in_channels=embed_size,
inter_channels=embed_size)
self.Rs_GCN_4 = Rs_GCN(in_channels=embed_size,
inter_channels=embed_size)
if self.data_name == 'f30k_precomp':
self.bn = nn.BatchNorm1d(embed_size)
def __init__(self, vocab_size, word_dim, embed_size=2048, inter_channels=2048, use_atten=False, word_vec='./vocab/word2vec300d_init_threshold4.npy'):
super(EncoderTextPrecompAttn, self).__init__()
self.use_atten = use_atten
self.embed_size = embed_size
self.embed = nn.Embedding(vocab_size, word_dim)
weight_init = torch.from_numpy(np.load(word_vec))
self.embed.weight.data[:vocab_size] = weight_init
self.fc = nn.Linear(word_dim, inter_channels)
self.init_weights()
# GSR
self.img_rnn = nn.GRU(inter_channels, embed_size, 1, batch_first=True, bidirectional=True)
# GCN reasoning
self.Rs_GCN_1 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
self.Rs_GCN_2 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
self.Rs_GCN_3 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
self.Rs_GCN_4 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
def __init__(self, img_dim, embed_size, inter_channels=2048, use_atten=False, use_box=False, use_label=False):
super(EncoderImagePrecompAttn, self).__init__()
self.embed_size = embed_size
self.use_atten = use_atten
self.use_box = use_box
self.use_label = use_label
if self.use_box:
self.fc_box = nn.Linear(6, embed_size)
if self.use_label:
self.fc_class = nn.Linear(embed_size, 33)
#self.fc = nn.Linear(img_dim, inter_channels)
#self.init_weights()
# GSR
self.img_rnn = nn.GRU(inter_channels, embed_size, 1, batch_first=True, bidirectional=True)
# GCN reasoning
self.Rs_GCN_1 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
self.Rs_GCN_2 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
self.Rs_GCN_3 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)
self.Rs_GCN_4 = Rs_GCN(in_channels=inter_channels, inter_channels=inter_channels)