def _get_onnx_sym(model: gluon.HybridBlock, num_inputs: int) -> sym.Symbol:
"""
Returns a symbolic graph for the model
:param model: gluon HybridBlock that constructs the symbolic graph
:param num_inputs: number of inputs to the graph
:return: symbol for the network
"""
var_args = [
sym.Variable('Data{}'.format(i)) for i in range(num_inputs)
]
with ScopedOnnxEnable(model):
return sym.Group(
gluon.block._flatten(model(*var_args), "output")[0])
def test_onnx_export_multi_output():
class MultiOutputBlock(nn.HybridBlock):
def __init__(self):
super(MultiOutputBlock, self).__init__()
with self.name_scope():
self.net = nn.HybridSequential()
for i in range(10):
self.net.add(nn.Dense(100 + i * 10, activation='relu'))
def hybrid_forward(self, F, x):
out = tuple(block(x) for block in self.net._children.values())
return out
net = MultiOutputBlock()
assert len(sym.Group(net(sym.Variable('data'))).list_outputs()) == 10
_check_onnx_export(net, group_outputs=True)
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 _optional_group(symbols, group=False):
if group:
return sym.Group(symbols)
else:
return symbols
