v1.4.py

#!/usr/bin/env python                                                                                                                                                                

import chainer
from chainer import functions as F
from chainer import optimizers as O
import numpy as np


print(chainer.__version__)


class Encoder(chainer.FunctionSet):
    def __call__(self, x):
        return self.l1(x)


class Comparator(chainer.FunctionSet):
    def __call__(self, xs):
        x1, x2 = xs
        x1 = comp.enc1(x1)
        x2 = comp.enc2(x2)
        return x1 - x2


# model setup                                                                                                                                                                        
enc1 = Encoder(l1=F.Linear(10, 3))
enc2 = Encoder(l1=F.Linear(10, 3))
comp = Comparator(enc1=enc1, enc2=enc2)

o = O.Adam()
o.setup(comp)

# forward                                                                                                                                                                            
xs = (chainer.Variable(
          np.random.uniform(-1, 1, (5, 10)).astype(np.float32)),
      chainer.Variable(
          np.random.uniform(-1, 1, (5, 10)).astype(np.float32)))
y = comp(xs)

# backward                                                                                                                                                                           
o.zero_grads()
y.grad = np.ones_like(y.data, dtype=np.float32)
y.backward()
print('before update, W=', comp.enc1.l1.W)
o.update()
print('after update, W=', comp.enc1.l1.W)

v1.5.py

#!/usr/bin/env python                                                                                                                                                                

import chainer
from chainer import links as L
from chainer import optimizers as O
import numpy as np


print(chainer.__version__)


class Encoder(chainer.Chain):
    def __call__(self, x):
        return self.l1(x)


class Comparator(chainer.Chain):
    def __call__(self, xs):
        x1, x2 = xs
        x1 = self.enc1(x1)
        x2 = self.enc2(x2)
        return x1 - x2


# model setup                                                                                                                                                                        
enc1 = Encoder(l1=L.Linear(10, 3))
enc2 = Encoder(l1=L.Linear(10, 3))
comp = Comparator(enc1=enc1, enc2=enc2)

o = O.Adam()
o.setup(comp)

# forward                                                                                                                                                                            
xs = (chainer.Variable(
          np.random.uniform(-1, 1, (5, 10)).astype(np.float32)),
      chainer.Variable(
          np.random.uniform(-1, 1, (5, 10)).astype(np.float32)))
y = comp(xs)

# backward                                                                                                                                                                           
comp.zerograds()
y.grad = np.ones_like(y.data, dtype=np.float32)
y.backward()
print('before update, W=', comp.enc1.l1.W.data)
o.update()
print('after update, W=', comp.enc1.l1.W.data)