def verify_elemwise_sum(num_args):
s = [sym.Variable("input" + str(i)) for i in range(num_args)]
y = sym.elemwise_sum(*s, num_args=num_args)
def forward(**inputs):
return np.sum(np.array(list(inputs.values())), axis=0)
def backward(head_grads, **inputs):
return [head_grads] * num_args
shape = {s[i]: (3, 4, 5) for i in range(num_args)}
check_function(y, forward, backward, shape=shape)
def verify_elemwise_sum(num_args):
s = [sym.Variable("input" + str(i)) for i in range(num_args)]
y = sym.elemwise_sum(*s, num_args=num_args)
def forward(**inputs):
return np.sum(np.array(list(inputs.values())), axis=0)
def backward(head_grads, **inputs):
return [head_grads] * num_args
dtype = "float32"
inputs = [("input" + str(i), (3, 4, 5), s[i]) for i in range(num_args)]
helper(y, inputs, dtype, forward, backward, need_input=False)
def test_multi_loss_graph_gradients():
# input data
shape1 = (1000, 100)
data1 = sym.Variable('data1', shape=(1000, 100), dtype=0)
# fake non-sparse label
label = sym.full(fill_value=3)
# square loss
sub1 = sym.elemwise_sub(data1, label, name="sub1")
square_loss = sym.sum(data=sub1**2, axis=1, name="square_loss")
# fake loss1
shape2 = (1000, )
data2 = sym.Variable('data2', shape=shape2, dtype=0)
loss1 = sym.sqrt(data2, name="loss1")
# fake loss2
loss2 = sym.relu(data1, name='loss2')
# block loss1
total_loss = sym.elemwise_sum(sym.block_grad(loss1),
square_loss,
num_args=2,
name="total_loss")
# grad_g.symbol.list_output_names()
# >> ['loss1_grad_0_output', 'grad_sum_output']
grad_g = graph_util.get_gradient_graph([total_loss, loss2],
total_loss.list_input_variables())
# infer shape
in_shapes, out_shapes = graph_util.infer_shape(grad_g)
assert out_shapes == [list(shape2), list(shape1)]
# grad_data1 is elemwise_sum of grad_loss2, grad_square_loss
grad_data1 = grad_g.symbol[1]
assert grad_data1.list_attr()['num_args'] == '2'
# block grad should return zero grad
grad_data2 = grad_g.symbol[0]
assert 'zeros_like' in grad_g.ir()
# test reverse infer shape for label
assert grad_g.apply('InferShape').json_attr('shape_num_unknown_nodes') == 0
# infer type
in_dtypes, out_dtypes = graph_util.infer_dtype(grad_g)
assert out_dtypes == ['float32', 'float32']
# test reverse infer type for label
assert grad_g.apply('InferType').json_attr('dtype_num_unknown_nodes') == 0
