def __init__(self, nblock=4):
super(E2ETransformer, self).__init__()
self.fvextractor = extractor.Resnet(nblock)
self.fvdim = 512 // (2**(4 - nblock))
# self.transformer = Transformer(self.fvdim)
self.transformer = Transformer(self.fvdim, NUM_HEADS=3, NUM_LAYERS=2)
def __init__(self, k=10, agg='l1', nblock=4, feat='resnet'):
super(NAggN, self).__init__()
if feat == 'resnet':
self.fvextractor = extractor.Resnet(nblock)
fvdim = 512 // (2**(4 - nblock))
elif feat == 'small':
self.fvextractor = extractor.SmallFeat()
fvdim = 512
else:
raise Exception("Not Implemented")
if agg == 'l1':
self.aggregator = L1Aggregator(fvdim)
elif agg == 'l2':
self.aggregator = L2Aggregator(fvdim)
elif agg == 'avg':
self.aggregator = AvgAggregator()
else:
raise Exception("unknown aggregator")
self.proj = nn.Linear(fvdim, k)
def __init__(self, nblock=4, k=10):
super(Resnet50, self).__init__()
self.fvextractor = extractor.Resnet(nblock, basic='resnet50')
fvdim = 2048 // (2**(4 - nblock))
self.proj = nn.Linear(fvdim, k)
def __init__(self, nblock=4, k=10):
super(SiNet, self).__init__()
self.fvextractor = extractor.Resnet(nblock)
fvdim = 512 // (2**(4 - nblock))
self.proj = nn.Linear(fvdim, k)