def __call__(self, net, input, target):
layer = net.layers[0]
if self.adapt:
while True:
self.epochf(None, net, input, target)
for inp, tar in zip(input, target):
out = net.step(inp)
err = tar - out
win = np.argmax(layer.out)
if np.max(err) == 0.0:
layer.np["w"][win] += self.lr * (inp - layer.np["w"][win])
else:
layer.np["w"][win] -= self.lr * (inp - layer.np["w"][win])
else:
while True:
output = []
winners = []
for inp, tar in zip(input, target):
out = net.step(inp)
output.append(out)
winners.append(np.argmax(layer.out))
e = self.error(net, input, target, output)
self.epochf(e, net, input, target)
error = target - output
sign = np.sign((np.max(error, axis=1) == 0) - 0.5)
layer.np["w"][winners] += self.lr * (input - layer.np["w"][winners])
return None
def error(self, net, input):
layer = net.layers[0]
winner_output = np.zeros_like(input)
output = net.sim(input)
winners = np.argmax(output, axis=1)
e = layer.np["w"][winners] - input
return net.errorf(e)
def learn(self, net, input):
layer = net.layers[0]
for inp in input:
out = net.step(inp)
winner = np.argmax(out)
d = layer.last_dist
print("TrainWTA learn here")
layer.np["w"][winner] += self.lr * d[winner] * (inp - layer.np["w"][winner])
return None
def learn(self, net, input):
layer = net.layers[0]
for inp in input:
out = net.step(inp)
winner = np.argmax(out)
d = layer.last_dist # TODO:^^_^^
layer.np["conscience"][winner] += 1
print("TrainCWTA learn here")
layer.np["w"][winner] += self.lr * d[winner] * (inp - layer.np["w"][winner])
layer.np["conscience"].fill(1.0)
return None
