def test_embed():
""" Test that embedding is treated like a Variable"""
embed_dense = L.EmbedID(5, 10)
embed_sparse = L.EmbedID(5, 10)
embed_dense.W.data[:] = np.random.randn(5, 10).astype('float32')
embed_sparse.W.data[:] = np.random.randn(5, 10).astype('float32')
embed_sparse.W.data[:, 1:] /= 1e5
dhl_dense_01 = dirichlet_likelihood(embed_dense, alpha=0.1).data
dhl_sparse_01 = dirichlet_likelihood(embed_sparse, alpha=0.1).data
msg = "Sparse vector has higher likelihood than dense with alpha=0.1"
assert dhl_sparse_01 > dhl_dense_01, msg
def forward(self, doc, wrd, window=5):
doc, wrd = utils.move(self.xp, doc, wrd)
proportions = self.proportions(doc)
ld = dirichlet_likelihood(self.proportions.W)
context = F.matmul(F.softmax(proportions), self.factors())
loss = self.loss_func(context, wrd)
return loss, ld
def test_concentration():
""" Test that alpha > 1.0 on a dense vector has a higher likelihood
than alpha < 1.0 on a dense vector, and test that a sparse vector
has the opposite character. """
dense = np.random.randn(5, 10).astype('float32')
sparse = np.random.randn(5, 10).astype('float32')
sparse[:, 1:] /= 1e5
weights = Variable(dense)
dhl_dense_10 = dirichlet_likelihood(weights, alpha=10.0).data
dhl_dense_01 = dirichlet_likelihood(weights, alpha=0.1).data
weights = Variable(sparse)
dhl_sparse_10 = dirichlet_likelihood(weights, alpha=10.0).data
dhl_sparse_01 = dirichlet_likelihood(weights, alpha=0.1).data
msg = "Sparse vector has higher likelihood than dense with alpha=0.1"
assert dhl_sparse_01 > dhl_dense_01, msg
msg = "Dense vector has higher likelihood than sparse with alpha=10.0"
assert dhl_dense_10 > dhl_sparse_10, msg
def forward(self, ids, bow):
bow, ids = utils.move(self.xp, bow, ids)
proportions = self.proportions(ids)
ld = dirichlet_likelihood(proportions)
doc = F.matmul(F.softmax(proportions), self.factors())
logp = F.dropout(self.embedding(doc))
# loss = -F.sum(bow * F.log_softmax(logp))
sources, targets, counts = [], [], []
lpi = F.sum(bow * F.log_softmax(logp), axis=1)
loss = -F.sum(lpi)
return loss, ld
def forward(self, ids, bow):
bow, ids = utils.move(self.xp, bow, ids)
proportions = self.proportions(ids)
ld = dirichlet_likelihood(proportions)
doc = F.matmul(F.softmax(proportions), self.factors())
logp = F.dropout(self.embedding(doc))
# loss = -F.sum(bow * F.log_softmax(logp))
sources, targets, counts = [], [], []
lpi = F.sum(bow * F.log_softmax(logp), axis=1)
loss = -F.sum(lpi)
return loss, ld
def prior(self):
dl1 = dirichlet_likelihood(self.mixture.weights)
return dl1
def prior(self):
dl1 = dirichlet_likelihood(self.mixture.weights)
return dl1
def prior(self):
""" Returns the log likelihood of the observed topic proportions."""
dl1 = dirichlet_likelihood(self.mixture.weights)
return dl1
def prior(self):
dl1 = dirichlet_likelihood(self.mixture_sty.weights)
dl2 = dirichlet_likelihood(self.mixture_aut.weights)
return dl1 + dl2
def prior(self):
dl1 = dirichlet_likelihood(self.mixture_sty.weights)
# dl2 = dirichlet_likelihood(self.mixture_aut.weights)
return dl1# + dl2
def prior(self):
# defaults to inialization with uniform prior (1/n_topics)
return dirichlet_likelihood(self.mixture.W, alpha=self.alpha)
