p04896132
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix bprop cache caused error with variable params

This commit is contained in:
义峰潘 2020-04-10 10:54:31 +08:00 committed by panyifeng
parent 268d358a1d
commit 2bef22d8a3
2 changed files with 19 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
mindspore/ccsrc/optimizer/ad/kprim.cc
View File

@ -92,9 +92,11 @@ FuncGraphPtr KPrim::KPrimitive(const ValueNodePtr &value_node, const pipeline::R
return nullptr; return nullptr;
} }
bool is_faked_bprop = false;
auto bprop_fg = GetBprop(prim); auto bprop_fg = GetBprop(prim);
if (bprop_fg == nullptr) { if (bprop_fg == nullptr) {
bprop_fg = FakeBprop(value_node, resources); bprop_fg = FakeBprop(value_node, resources);
is_faked_bprop = true;
} }
auto expanded_fg = BpropToK(prim, bprop_fg); auto expanded_fg = BpropToK(prim, bprop_fg);
@ -104,8 +106,11 @@ FuncGraphPtr KPrim::KPrimitive(const ValueNodePtr &value_node, const pipeline::R
<< trace::GetDebugInfo(bprop_fg->debug_info()); << trace::GetDebugInfo(bprop_fg->debug_info());
} }
// Set bprop_g graph cache // To support primitives with variable params, do not cache faked bprop
bprop_registry_[prim] = expanded_fg; if (!is_faked_bprop) {
// Set bprop_g graph cache
bprop_registry_[prim] = expanded_fg;
}
return expanded_fg; return expanded_fg;
} }

12
tests/ut/python/pynative_mode/test_stop_gradient.py
View File

@ -366,3 +366,15 @@ def test_stop_gradient_11():
with pytest.raises(RuntimeError): with pytest.raises(RuntimeError):
bprop(PrimWithNoBprop_(), Tensor(np.ones([2]).astype(np.float32)), bprop(PrimWithNoBprop_(), Tensor(np.ones([2]).astype(np.float32)),
Tensor(np.ones([2]).astype(np.float32))) Tensor(np.ones([2]).astype(np.float32)))
def test_stop_print():
class StopPrint(nn.Cell):
def __init__(self):
super(StopPrint, self).__init__()
self.printm = P.Print()
def construct(self, x, y):
self.printm("StopPrint", x)
self.printm(y)
return x, y
C.grad_all(StopPrint())(Tensor(np.ones([2]).astype(np.float32)),
Tensor(np.ones([2]).astype(np.float32)))