def test_backprop():
graph = chainer_compiler_core.load(
'out/ch2o_node_Linear_backprop/model.onnx')
params = graph.params()
input_names = graph.input_names()
output_names = graph.output_names()
assert len(input_names) == 1
assert len(output_names) == 1
fwd_graph, bwd_graph = graph.backward()
assert len(fwd_graph.input_names()) == 1
assert len(fwd_graph.output_names()) == 3
assert len(bwd_graph.input_names()) == 3
assert len(bwd_graph.output_names()) == 2
fwd = fwd_graph.compile()
bwd = bwd_graph.compile()
fwd_inputs = dict(params)
t1 = aranges(5, 7)
fwd_inputs[input_names[0]] = chainer_compiler_core.value(t1)
loss = (chainerx.dot(t1, params['/l1/W'].array().T) +
params['/l1/b'].array())
fwd_outputs = fwd.run(fwd_inputs)
assert len(fwd_outputs) == 3
chainerx.testing.assert_allclose(loss,
fwd_outputs[output_names[0]].array())
grad_loss = aranges(*loss.shape) + 4.2
bwd_inputs = {}
for name in fwd_graph.output_names():
iname = name
value = fwd_outputs[name]
if name in output_names:
iname = 'grad_in@' + name
value = chainer_compiler_core.value(grad_loss)
bwd_inputs[iname] = value
bwd_outputs = bwd.run(bwd_inputs)
grad_w = chainerx.dot(grad_loss.T, t1)
chainerx.testing.assert_allclose(grad_w,
bwd_outputs['grad_out@/l1/W'].array())
grad_b = chainerx.sum(grad_loss, axis=0)
chainerx.testing.assert_allclose(grad_b,
bwd_outputs['grad_out@/l1/b'].array())
def forward_chainerx(self, x):
a, b = x
# TODO(sonots): Support transa and transb in ChainerX
if self.transa or self.transb or self.transc:
return chainer.Fallback
# TODO(sonots): Support dtype promotion in ChainerX
if a.dtype != b.dtype:
return chainer.Fallback
# TODO(sonots): Support ndim > 2 in ChainerX
if a.ndim != 2 or b.ndim != 2:
return chainer.Fallback
# TODO(niboshi): Support it
if self.dtype is not None and self.dtype != a.dtype:
return chainer.Fallback
return chainerx.dot(a, b),
def forward_chainerx(self, x):
a, b = x
# TODO(sonots): Support transa and transb in ChainerX
if self.transa or self.transb or self.transc:
return chainer.Fallback
# TODO(sonots): Support dtype promotion in ChainerX
if a.dtype != b.dtype:
return chainer.Fallback
# TODO(sonots): Support ndim > 2 in ChainerX
if a.ndim != 2 or b.ndim != 2:
return chainer.Fallback
# TODO(niboshi): Support it
if self.dtype is not None and self.dtype != a.dtype:
return chainer.Fallback
return chainerx.dot(a, b),
def test_inference():
graph = chainer_compiler_core.load('out/ch2o_node_Linear/model.onnx')
params = graph.params()
input_names = graph.input_names()
output_names = graph.output_names()
assert len(input_names) == 1
assert len(output_names) == 2
xcvm = graph.compile()
inputs = dict(params)
t1 = aranges(5, 7)
inputs[input_names[0]] = chainer_compiler_core.value(t1)
y1 = chainerx.dot(t1, params['/l1/W'].array().T) + params['/l1/b'].array()
y2 = chainerx.dot(t1, params['/l2/W'].array().T)
outputs = xcvm.run(inputs)
assert len(outputs) == 2
chainerx.testing.assert_allclose(y1, outputs[output_names[0]].array())
chainerx.testing.assert_allclose(y2, outputs[output_names[1]].array())
assert 'op_type: "ChainerLinear"' in graph.dump()