def get_op(self, row_cnt):
if row_cnt in self.ops:
return self.ops[row_cnt]
graph = self.model.graph
X = graph.X
# build the graph
h0_probs, hn_sample = graph.get_op_sample_h_given_v(X, row_cnt)
for n in range(1, self.k + 1):
vn_probs, vn_sample, hn_probs, hn_sample = graph.get_op_gibbs_hvh(
hn_sample, row_cnt)
# gradients
grad_W = (sym.dot(sym.transpose(vn_sample), hn_probs) -
sym.dot(sym.transpose(X), h0_probs))
grad_v_bias = sym.sum(vn_sample - X, axis=0)
grad_h_bias = sym.sum(hn_probs - h0_probs, axis=0)
# loss
loss = sym.mean(sym.square(X - vn_probs))
op = sym.Group([loss, grad_W, grad_v_bias, grad_h_bias])
self.ops[row_cnt] = op
return op
def get_op_propgate_h2v(self, h):
v_probs = sym.sigmoid(
sym.broadcast_add(sym.dot(h, sym.transpose(self.W)), self.v_bias))
return v_probs
def hybrid_flatten_pred(sym_pred):
return sym.transpose(sym_pred, (0, 2, 3, 1)).flatten()