def __init__(self, S, L, means):
self.S = S
self.L = L
self.means = means
self.l2norm = L2Normalization()
# RoI max pooling
self.r_mac_pool = RoIPool(1, 1, 0.03125)
self.pca_shift = Shift(2048)
self.pca_shift.bias.data = torch.Tensor(np.load('centered2.npy'))
self.pca_fc = nn.Linear(2048, 2048, bias=True)
# Load the PCA weights learned with off-the-shelf Resnet101
self.pca_fc.weight.data = torch.Tensor(np.load('weights/pca_weight.npy'))
self.pca_fc.bias.data = torch.Tensor(np.load('weights/pca_bias.npy'))
def __init__(self):
super(OurNet, self).__init__()
self.l2norm = L2Normalization()
#RoI max pooling
self.r_mac_pool = RoIPool(1, 1, 0.03125)
#Define regional attention network
self.region_attention = ContextAwareRegionalAttentionNetwork(spatial_scale = 0.03125)
self.pca_shift = Shift(2048)
self.pca_fc = nn.Linear(2048, 2048, bias=False)
#Load the PCA weights learned with off-the-shelf Resnet101
self.pca_fc.weight.data = torch.Tensor(np.load('weights/pca_components.npy'))
self.pca_shift.bias.data = torch.Tensor(np.load('weights/pca_mean.npy'))
self.resnet = Resnet.resnet101(pretrained=False)
#Load off-the-shelf Resnet101 of caffe version.
dic = torch.load('weights/resnet101_caffeProto.pth', map_location=lambda storage, loc: storage)
self.resnet.load_state_dict(dic, strict=False)