y = np.random.choice([0, 1], size=(N,), p=[1./3, 2./3])
test.append((x, y))
model = RRBM(N, train)
# check partition function
assert np.allclose(model.Z(model.W, model.b, train[0][0]), model.enum_Z(model.W, model.b, train[0][0]), atol = 0.01)
gW_fd, gb_fd = fd(model)
gW, gb = model.g(model.W, model.b)
# check gradient
assert np.allclose(gW, gW_fd, atol=0.01)
assert np.allclose(gb, gb_fd, atol=0.01)
# check conversion
W2, b2 = model.vec2theta(model.theta2vec(model.W, model.b))
assert (W2 == model.W).all()
assert (b2 == model.b).all()
# check optimization
err, ham = model.eval(test)
print err
print ham
model.learn()
err, ham = model.eval(test)
print err
print ham
characters[i] = 1.0
test.append((lex_col[lang1][l1], lex_col[lang2][l2], embedding, characters))
model.reset()
model.train = train
before = model.eval(test)[0:2]
#print size, experiment
C_best = 0
err, hmm, mistakes = float("inf"), float("inf"), None
# grid search
for C in [1.0]: #0.0, 0.1, 0.5, 1.0, 2.0, 3.0, 4.0]:
model.C = C
model.learn(disp=0)
err_tmp, hmm_tmp, mistakes_tmp = model.eval(test)
if hmm_tmp < hmm:
err, hmm, mistakes = err_tmp, hmm_tmp, mistakes_tmp
C_best = C
after = (err, hmm)
print size, experiment, before, after, C_best
for iter in xrange(100):
train2 = []
for datum in train:
train2.append(datum)