def classify(self, x):
features = utils.to_numpy(self.convnet(x)) # B x 128
features /= np.sum(features**2,axis=1, keepdims=True)**.5
r = np.tensordot(self.features[:,:,None],
features[:,None,:],
axes=([1,2], [2,1])).T # B x 50000
ordered_preds = np.argsort(r, axis=1)
preds = [scipy.stats.mode(self.labels[x[-self.NEAREST:]]).mode for x in ordered_preds]
return np.array([np.eye(10)[x] for x in np.array(preds).flatten()])
def classify(self, x):
return utils.to_numpy(self.convnet(x))
def fn(x):
for layer in model.layers[:-n]:
x = layer(x)
h = utils.to_numpy(x)
h = h.mean(axis)
return h
def detect(self, x):
a, b, c = [
utils.to_numpy(model(x)).argmax(1) for model in self.convnets
]
has_majority = (a == b) | (a == c) | (b == c)
return ~has_majority
def classify(self, x):
a, b, c = [
utils.to_numpy(model(x)).argmax(1) for model in self.convnets
]
majority_vote = ((a == b) * a) | ((a == c) * a) | ((b == c) * c)
return np.array([np.eye(10)[i] for i in majority_vote])
def classify(self, xs, training=False):
xs = self.encode(xs)
return utils.to_numpy(self.convnet(xs))
def classify(self, x):
predictions = [utils.to_numpy(net(x)) for net in self.class_nets]
predictions = np.stack(predictions, axis=1)
predictions = 5 * utils.sigmoid(
5 * (predictions[:, :, 1] - predictions[:, :, 0]))
return utils.softmax(predictions)
def classify(self, x):
preds = [
utils.to_numpy(self.convnet(self.to_tensor(x))) for _ in range(10)
]
return np.mean(preds, axis=0)
def classify(self, x):
x_pad = self.blur(x)
return utils.to_numpy(self.convnet(x_pad))