提交 610a2584 编写于 作者: Y Yu Yang

Fix all unit tests in Python

上级 d6d46412
...@@ -60,7 +60,12 @@ void ExposeOperator(ClassType &m) { ...@@ -60,7 +60,12 @@ void ExposeOperator(ClassType &m) {
-> std::unordered_map<std::string, std::vector<std::string>> { -> std::unordered_map<std::string, std::vector<std::string>> {
return op.outputs_; return op.outputs_;
}) })
.def("__str__", &ClassType::type::DebugString); .def("inputs",
[](const typename ClassType::type &op) { return op.inputs_; })
.def("__str__", &ClassType::type::DebugString)
.def("no_intermediate_outputs", [](const typename ClassType::type &op) {
return op.OutputVars(false);
});
} }
static size_t UniqueIntegerGenerator() { static size_t UniqueIntegerGenerator() {
......
...@@ -53,15 +53,18 @@ def get_numeric_gradient(op, ...@@ -53,15 +53,18 @@ def get_numeric_gradient(op,
tensor.set(input_values[var_name], core.CPUPlace()) tensor.set(input_values[var_name], core.CPUPlace())
# Create all output variable in local_scope # Create all output variable in local_scope
for output in op.outputs(): opts = op.outputs()
if local_scope.find_var(output) is None: for key in opts:
local_scope.new_var(output).get_tensor() for output in opts[key]:
if local_scope.find_var(output) is None:
local_scope.new_var(output).get_tensor()
op.infer_shape(local_scope) op.infer_shape(local_scope)
# allocate output memory # allocate output memory
for output in op.outputs(): for key in opts:
local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace()) for output in opts[key]:
local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace(
))
# TODO(yuyang18): Only CPU is support now. # TODO(yuyang18): Only CPU is support now.
cpu_ctx = core.DeviceContext.create(core.CPUPlace()) cpu_ctx = core.DeviceContext.create(core.CPUPlace())
...@@ -150,19 +153,24 @@ class GradientChecker(unittest.TestCase): ...@@ -150,19 +153,24 @@ class GradientChecker(unittest.TestCase):
if no_grad_set is None: if no_grad_set is None:
no_grad_set = set() no_grad_set = set()
tmp_outs = forward_op.temp_outputs() no_tmp_out = forward_op.no_intermediate_outputs()
no_tmp_out = filter(lambda name: name not in tmp_outs,
forward_op.outputs())
if len(no_tmp_out) != 1: if len(no_tmp_out) != 1:
raise ValueError("non temp out_names should be 1") raise ValueError("non temp out_names should be 1")
in_names = forward_op.inputs() inputs = forward_op.inputs()
in_names = [item for k in inputs for item in inputs[k]]
outputs = forward_op.outputs()
out_names = [item for k in outputs for item in outputs[k]]
for no_grad in no_grad_set: for no_grad in no_grad_set:
if no_grad not in in_names: if no_grad not in in_names:
raise ValueError("no_grad should be in in_names") raise ValueError("no_grad should be in in_names")
backward_op = core.Operator.backward(forward_op, no_grad_set) backward_op = core.Operator.backward(forward_op, no_grad_set)
bwd_outputs = backward_op.outputs()
bwd_out_names = [item for k in bwd_outputs for item in bwd_outputs[k]]
places = [core.CPUPlace()] places = [core.CPUPlace()]
if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu(): if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
places.append(core.GPUPlace(0)) places.append(core.GPUPlace(0))
...@@ -188,7 +196,7 @@ class GradientChecker(unittest.TestCase): ...@@ -188,7 +196,7 @@ class GradientChecker(unittest.TestCase):
var.set(value, place) var.set(value, place)
# create output var # create output var
for out_name in forward_op.outputs(): for out_name in out_names:
scope.new_var(out_name).get_tensor() scope.new_var(out_name).get_tensor()
# infer the shape of output var and compute/set value of output var # infer the shape of output var and compute/set value of output var
...@@ -198,7 +206,7 @@ class GradientChecker(unittest.TestCase): ...@@ -198,7 +206,7 @@ class GradientChecker(unittest.TestCase):
# create output grad var # create output grad var
# set shape as the output var # set shape as the output var
# set value of this grad to ones # set value of this grad to ones
for name in forward_op.outputs(): for name in out_names:
out_tensor = scope.find_var(name).get_tensor() out_tensor = scope.find_var(name).get_tensor()
grad_tensor = scope.new_var(grad_var_name(name)).get_tensor() grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
grad_tensor.set_dims(out_tensor.shape()) grad_tensor.set_dims(out_tensor.shape())
...@@ -206,7 +214,7 @@ class GradientChecker(unittest.TestCase): ...@@ -206,7 +214,7 @@ class GradientChecker(unittest.TestCase):
grad_tensor.set(data, place) grad_tensor.set(data, place)
# create input grad var # create input grad var
for name in backward_op.outputs(): for name in bwd_out_names:
scope.new_var(name).get_tensor() scope.new_var(name).get_tensor()
# infer the shape of input gradient var and compute/set it's value # infer the shape of input gradient var and compute/set it's value
......
...@@ -21,17 +21,18 @@ class TestCrossEntropy(unittest.TestCase): ...@@ -21,17 +21,18 @@ class TestCrossEntropy(unittest.TestCase):
self.outputs = {'Y': numpy.array(Y).astype("float32")} self.outputs = {'Y': numpy.array(Y).astype("float32")}
# class CrossEntropyGradOpTest(GradientChecker): class CrossEntropyGradOpTest(GradientChecker):
# def test_softmax_grad(self): def test_softmax_grad(self):
# op = create_op("onehot_cross_entropy") op = create_op("onehot_cross_entropy")
# batch_size = 100 batch_size = 100
# class_num = 10 class_num = 10
# inputs = { inputs = {
# "X": numpy.random.uniform( "X": numpy.random.uniform(
# 0.1, 1.0, [batch_size, class_num]).astype("float32"), 0.1, 1.0, [batch_size, class_num]).astype("float32"),
# "label": (class_num / 2) * numpy.ones(batch_size).astype("int32") "label": (class_num / 2) * numpy.ones(batch_size).astype("int32")
# } }
# self.check_grad(op, inputs, set("X"), "Y") self.check_grad(op, inputs, set("X"), "Y")
if __name__ == "__main__": if __name__ == "__main__":
unittest.main() unittest.main()
...@@ -25,12 +25,12 @@ class TestNet(unittest.TestCase): ...@@ -25,12 +25,12 @@ class TestNet(unittest.TestCase):
net.complete_add_op(True) net.complete_add_op(True)
expected = ''' expected = '''
Op(plain_net), inputs:(W, X, Y), outputs:(Out, fc.out, pre_activation). Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}.
Op(add_two), inputs:(X, Y), outputs:(Out). Op(add_two), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}.
Op(plain_net), inputs:(W, X), outputs:(fc.out, pre_activation). Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
Op(plain_net), inputs:(W, X), outputs:(fc.out, pre_activation). Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
Op(mul), inputs:(X, W), outputs:(pre_activation). Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}.
Op(sigmoid), inputs:(pre_activation), outputs:(fc.out). Op(sigmoid), inputs:{X[pre_activation]}, outputs:{Y[fc.out]}.
''' '''
self.assertEqual(expected, "\n" + str(net)) self.assertEqual(expected, "\n" + str(net))
......
import paddle.v2.framework.proto.op_proto_pb2 as op_proto_lib import paddle.v2.framework.proto.framework_pb2 as framework_pb2
import paddle.v2.framework.proto.attribute_pb2 as attr_type_lib
import unittest import unittest
class TestFrameworkProto(unittest.TestCase): class TestFrameworkProto(unittest.TestCase):
def test_all(self): def test_all(self):
op_proto = op_proto_lib.OpProto() op_proto = framework_pb2.OpProto()
ipt0 = op_proto.inputs.add() ipt0 = op_proto.inputs.add()
ipt0.name = "a" ipt0.name = "a"
ipt0.comment = "the input of cosine op" ipt0.comment = "the input of cosine op"
...@@ -19,7 +18,7 @@ class TestFrameworkProto(unittest.TestCase): ...@@ -19,7 +18,7 @@ class TestFrameworkProto(unittest.TestCase):
attr = op_proto.attrs.add() attr = op_proto.attrs.add()
attr.name = "scale" attr.name = "scale"
attr.comment = "scale of cosine op" attr.comment = "scale of cosine op"
attr.type = attr_type_lib.FLOAT attr.type = framework_pb2.FLOAT
op_proto.type = "cos" op_proto.type = "cos"
self.assertTrue(op_proto.IsInitialized()) self.assertTrue(op_proto.IsInitialized())
......
...@@ -24,11 +24,12 @@ class TestSoftmaxOp(unittest.TestCase): ...@@ -24,11 +24,12 @@ class TestSoftmaxOp(unittest.TestCase):
} }
# class SoftmaxGradOpTest(GradientChecker): class SoftmaxGradOpTest(GradientChecker):
# def test_softmax(self): def test_softmax(self):
# op = create_op("softmax") op = create_op("softmax")
# inputs = {"X": np.random.uniform(0.1, 1, [10, 10]).astype("float32")} inputs = {"X": np.random.uniform(0.1, 1, [10, 10]).astype("float32")}
# self.check_grad(op, inputs, set("X"), "Y") self.check_grad(op, inputs, set("X"), "Y")
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册