Polish code

test=develop

paddle/fluid/imperative/layer.cc

@@ -168,12 +168,12 @@ class Autograd {
   }
 };
 
-VarBase* VarBase::NewVarBase(const platform::Place& dst_place,
+std::unique_ptr<VarBase> VarBase::NewVarBase(const platform::Place& dst_place,
                                              const bool blocking) const {
   PADDLE_ENFORCE(var_->IsInitialized(),
                  "Variable must be initialized when getting numpy tensor");
 
-  VarBase* new_var = new VarBase();
+  std::unique_ptr<VarBase> new_var(new VarBase());
   framework::LoDTensor* tensor =
       new_var->var_->GetMutable<framework::LoDTensor>();
   tensor->Resize(var_->Get<framework::LoDTensor>().dims());
@@ -240,9 +240,8 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
     PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
 
     framework::Scope scope;
-    platform::Place place = place_;
-    PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
-    p.op.RuntimeInferShape(scope, place, ctx);
+    PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place_);
+    p.op.RuntimeInferShape(scope, place_, ctx);
     p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
   }
 

paddle/fluid/imperative/layer.h

@@ -21,6 +21,7 @@
 #include <map>     // NOLINT
 #include <string>  // NOLINT
 #include <vector>  // NOLINT
+#include <memory>  // NOLINT
 
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/operator.h"
@@ -153,7 +154,7 @@ class VarBase {
 
   framework::LoDTensor& GradValue();
 
-  VarBase* NewVarBase(const platform::Place& dst_place,
+  std::unique_ptr<VarBase> NewVarBase(const platform::Place& dst_place,
                                       const bool blocking) const;
 
   inline std::string GradName() const {

paddle/fluid/pybind/pybind.cc

@@ -137,13 +137,21 @@ PYBIND11_MODULE(core, m) {
       .def("_grad_ivar",
            [](const imperative::VarBase &self) { return self.grads_; },
            py::return_value_policy::reference)
-      .def("_to",
+      .def("_copy_to",
            [](const imperative::VarBase &self, const platform::CPUPlace &place,
-              bool blocking) { return self.NewVarBase(place, blocking); },
+              bool blocking) {
+             std::unique_ptr<imperative::VarBase> new_var =
+                 self.NewVarBase(place, blocking);
+             return new_var.release();
+           },
            py::return_value_policy::take_ownership)
-      .def("_to",
+      .def("_copy_to",
            [](const imperative::VarBase &self, const platform::CUDAPlace &place,
-              bool blocking) { return self.NewVarBase(place, blocking); },
+              bool blocking) {
+             std::unique_ptr<imperative::VarBase> new_var =
+                 self.NewVarBase(place, blocking);
+             return new_var.release();
+           },
            py::return_value_policy::take_ownership)
       .def("value", [](const imperative::VarBase &self) { return self.var_; },
            py::return_value_policy::reference)

python/paddle/fluid/framework.py

@@ -67,7 +67,7 @@ ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
 CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()
 
 _imperative_tracer_ = None
-_current_expected_place_ = None
+_imperative_current_expected_place_ = None
 
 
 def _in_imperative_mode():
@@ -79,7 +79,7 @@ def _imperative_tracer():
 
 
 def _current_expected_place():
-    return _current_expected_place_
+    return _imperative_current_expected_place_
 
 
 class NameScope(object):
@@ -385,7 +385,7 @@ class Variable(object):
             self._ivar.stop_gradient = stop_gradient
 
     def _numpy(self):
-        new_ivar = self._ivar._to(core.CPUPlace(), True)
+        new_ivar = self._ivar._copy_to(core.CPUPlace(), True)
         return np.array(new_ivar.value().get_tensor())
 
     def _backward(self):
@@ -1313,7 +1313,8 @@ class Block(object):
     def _trace_op(self, op, stop_gradient=False):
         if _in_imperative_mode():
             _imperative_tracer().trace(op.iop, op.inputs, op.outputs, self.desc,
-                                       _current_expected_place_, stop_gradient)
+                                       _imperative_current_expected_place_,
+                                       stop_gradient)
 
     def _insert_op(self, index, *args, **kwargs):
         """
@@ -2338,10 +2339,10 @@ def _imperative_guard(tracer):
 
 @contextlib.contextmanager
 def _imperative_place_guard(place):
-    global _current_expected_place_
-    tmp_place = _current_expected_place_
-    _current_expected_place_ = place
+    global _imperative_current_expected_place_
+    tmp_place = _imperative_current_expected_place_
+    _imperative_current_expected_place_ = place
 
     yield
 
-    _current_expected_place_ = tmp_place
+    _imperative_current_expected_place_ = tmp_place

python/paddle/fluid/imperative/nn.py

@@ -144,7 +144,7 @@ class Conv2D(layers.Layer):
             attrs={'axis': 1})
 
         # Currently, we don't support inplace in imperative mode
-        return self._helper.append_activation(pre_act, force_no_inplace=True)
+        return self._helper.append_activation(pre_act)
 
 
 class Pool2D(layers.Layer):
@@ -286,8 +286,7 @@ class FC(layers.Layer):
         else:
             pre_activation = pre_bias
         # Currently, we don't support inplace in imperative mode
-        return self._helper.append_activation(
-            pre_activation, force_no_inplace=True)
+        return self._helper.append_activation(pre_activation)
 
 
 class BatchNorm(layers.Layer):
@@ -419,5 +418,4 @@ class BatchNorm(layers.Layer):
             })
 
         # Currently, we don't support inplace in imperative mode
-        return self._helper.append_activation(
-            batch_norm_out, force_no_inplace=True)
+        return self._helper.append_activation(batch_norm_out)

python/paddle/fluid/layer_helper.py

@@ -419,7 +419,7 @@ class LayerHelper(object):
             attrs={'axis': dim_start})
         return tmp
 
-    def append_activation(self, input_var, force_no_inplace=False):
+    def append_activation(self, input_var):
         act = self.kwargs.get('act', None)
         if act is None:
             return input_var
@@ -436,7 +436,7 @@ class LayerHelper(object):
         tmp = input_var
         # NOTE(dzhwinter): some activation support inplace compution.
         # NOTE(minqiyang): currently, we don't support inplace in imperative mode
-        if not force_no_inplace and core.IsInplace(act_type):
+        if not imperative_base.enabled() and core.IsInplace(act_type):
             tmp = input_var
         else:
             tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)

python/paddle/fluid/optimizer.py

@@ -388,7 +388,6 @@ class Optimizer(object):
             params_grads = []
             for param in parameters:
                 if param.stop_gradient:
-                    print("parameter:", param.name, "stop gradient, skip it")
                     continue
                 # create gradient variable
                 grad_var = Variable(

python/paddle/fluid/tests/unittests/test_imperative.py

@@ -68,7 +68,7 @@ class MLP(fluid.imperative.Layer):
 
 class TestImperative(unittest.TestCase):
     def test_layer(self):
-        with fluid.imperative.guard(device=None):
+        with fluid.imperative.guard():
             cl = core.Layer()
             cl.forward([])
             l = fluid.imperative.Layer()
@@ -76,7 +76,7 @@ class TestImperative(unittest.TestCase):
 
     def test_pylayer_func_id(self):
 
-        with fluid.imperative.guard(device=None):
+        with fluid.imperative.guard():
 
             class PyLayer1(fluid.imperative.PyLayer):
                 def __init__(self):
@@ -116,7 +116,7 @@ class TestImperative(unittest.TestCase):
 
     def test_pylayer(self):
         np_inp = np.ones([2, 2], np.float32)
-        with fluid.imperative.guard(device=None):
+        with fluid.imperative.guard():
             my_py_layer = MyPyLayer()
             var_inp = fluid.imperative.base.to_variable(np_inp)
             outs = my_py_layer(var_inp)
@@ -133,7 +133,8 @@ class TestImperative(unittest.TestCase):
             x = fluid.layers.reduce_sum(fluid.layers.tanh(x1))
             param_grads = fluid.backward.append_backward(
                 x, parameter_list=[x1.name])[0]
-            exe = fluid.Executor(fluid.CPUPlace())
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
             static_out, static_grad = exe.run(
                 feed={inp.name: np_inp},
@@ -144,7 +145,7 @@ class TestImperative(unittest.TestCase):
 
     def test_layer_in_out(self):
         np_inp = np.array([1.0, 2.0, -1.0], dtype=np.float32)
-        with fluid.imperative.guard(device=None):
+        with fluid.imperative.guard():
             var_inp = fluid.imperative.base.to_variable(np_inp)
             l = MyLayer()
             x = l(var_inp)[0]
@@ -160,7 +161,8 @@ class TestImperative(unittest.TestCase):
             x = l(inp)[0]
             param_grads = fluid.backward.append_backward(
                 x, parameter_list=[l._x_for_debug.name])[0]
-            exe = fluid.Executor(fluid.CPUPlace())
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
             static_out, static_grad = exe.run(
                 feed={inp.name: np_inp},
@@ -171,7 +173,7 @@ class TestImperative(unittest.TestCase):
 
     def test_mlp(self):
         np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
-        with fluid.imperative.guard(device=None):
+        with fluid.imperative.guard():
             var_inp = fluid.imperative.base.to_variable(np_inp)
             mlp = MLP()
             out = mlp(var_inp)
@@ -186,7 +188,8 @@ class TestImperative(unittest.TestCase):
             out = mlp(inp)
             param_grads = fluid.backward.append_backward(
                 out, parameter_list=[mlp._fc1._w.name])[0]
-            exe = fluid.Executor(fluid.CPUPlace())
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
             exe.run(fluid.default_startup_program())
 
             static_out, static_grad = exe.run(

python/paddle/fluid/tests/unittests/test_imperative_gan.py

@@ -20,6 +20,7 @@ import sys
 
 import paddle
 import paddle.fluid as fluid
+import paddle.fluid.core as core
 from paddle.fluid.optimizer import SGDOptimizer
 from paddle.fluid.imperative.nn import Conv2D, Pool2D, FC
 from test_imperative_base import new_program_scope
@@ -58,7 +59,7 @@ class Generator(fluid.imperative.Layer):
 
 
 class TestImperativeMnist(unittest.TestCase):
-    def test_mnist_cpu_float32(self):
+    def test_gan_float32(self):
         seed = 90
 
         startup = fluid.Program()
@@ -115,7 +116,8 @@ class TestImperativeMnist(unittest.TestCase):
             sgd = SGDOptimizer(learning_rate=1e-3)
             sgd.minimize(g_loss)
 
-        exe = fluid.Executor(fluid.CPUPlace())
+        exe = fluid.Executor(fluid.CPUPlace() if not core.is_compiled_with_cuda(
+        ) else fluid.CUDAPlace(0))
         static_params = dict()
         with fluid.scope_guard(scope):
             img = np.ones([2, 1], np.float32)
@@ -135,7 +137,7 @@ class TestImperativeMnist(unittest.TestCase):
                     scope.find_var(param.name).get_tensor())
 
         dy_params = dict()
-        with fluid.imperative.guard(place=fluid.CPUPlace()):
+        with fluid.imperative.guard():
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 

python/paddle/fluid/tests/unittests/test_imperative_optimizer.py

@@ -101,7 +101,7 @@ class TestImperativeMnist(unittest.TestCase):
     def test_mnist_cpu_float32(self):
         seed = 90
 
-        with fluid.imperative.guard(place=fuild.CPUPlace()):
+        with fluid.imperative.guard():
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 
@@ -145,7 +145,8 @@ class TestImperativeMnist(unittest.TestCase):
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 
-            exe = fluid.Executor(fluid.CPUPlace())
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
             mnist = MNIST()
             sgd = SGDOptimizer(learning_rate=1e-3)

python/paddle/fluid/tests/unittests/test_imperative_resnet.py

@@ -143,7 +143,7 @@ class BottleneckBlock(fluid.imperative.Layer):
         y = fluid.layers.elementwise_add(x=short, y=conv2)
 
         layer_helper = LayerHelper('elementwise_add_activation', act='relu')
-        return layer_helper.append_activation(y, force_no_inplace=True)
+        return layer_helper.append_activation(y)
 
 
 class ResNet(fluid.imperative.Layer):
@@ -204,12 +204,9 @@ class ResNet(fluid.imperative.Layer):
 
 
 class TestImperativeResnet(unittest.TestCase):
-    def test_resnet_gpu_float32(self):
+    def test_resnet_float32(self):
         seed = 90
 
-        if not core.is_compiled_with_cuda():
-            return
-
         batch_size = train_parameters["batch_size"]
         batch_num = 1
         with fluid.imperative.guard():
@@ -277,168 +274,8 @@ class TestImperativeResnet(unittest.TestCase):
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 
-            exe = fluid.Executor(fluid.CUDAPlace(0))
-
-            resnet = ResNet()
-            optimizer = optimizer_setting(train_parameters)
-
-            np.random.seed(seed)
-            import random
-            random.seed = seed
-            train_reader = paddle.batch(
-                paddle.dataset.flowers.train(use_xmap=False),
-                batch_size=batch_size)
-
-            img = fluid.layers.data(
-                name='pixel', shape=[3, 224, 224], dtype='float32')
-            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
-            out = resnet(img)
-            loss = fluid.layers.cross_entropy(input=out, label=label)
-            avg_loss = fluid.layers.mean(x=loss)
-            optimizer.minimize(avg_loss)
-
-            # initialize params and fetch them
-            static_param_init_value = {}
-            static_param_name_list = []
-            static_grad_name_list = []
-            for param in fluid.default_startup_program().global_block(
-            ).all_parameters():
-                static_param_name_list.append(param.name)
-            for param in fluid.default_main_program().global_block(
-            ).all_parameters():
-                if not param.stop_gradient:
-                    static_grad_name_list.append(param.name +
-                                                 core.grad_var_suffix())
-
-            out = exe.run(fluid.default_startup_program(),
-                          fetch_list=static_param_name_list)
-
-            for i in range(len(static_param_name_list)):
-                static_param_init_value[static_param_name_list[i]] = out[i]
-
-            for batch_id, data in enumerate(train_reader()):
-                if batch_id >= batch_num:
-                    break
-
-                static_x_data = np.array(
-                    [x[0].reshape(3, 224, 224) for x in data]).astype('float32')
-                y_data = np.array([x[1] for x in data]).astype('int64').reshape(
-                    [batch_size, 1])
-
-                fetch_list = [avg_loss.name]
-                fetch_list.extend(static_param_name_list)
-                fetch_list.extend(static_grad_name_list)
-                out = exe.run(fluid.default_main_program(),
-                              feed={"pixel": static_x_data,
-                                    "label": y_data},
-                              fetch_list=fetch_list)
-
-                static_param_value = {}
-                static_grad_value = {}
-                static_out = out[0]
-                param_start_pos = 1
-                grad_start_pos = len(static_param_name_list) + param_start_pos
-                for i in range(param_start_pos,
-                               len(static_param_name_list) + param_start_pos):
-                    static_param_value[static_param_name_list[
-                        i - param_start_pos]] = out[i]
-                for i in range(grad_start_pos,
-                               len(static_grad_name_list) + grad_start_pos):
-                    static_grad_value[static_grad_name_list[
-                        i - grad_start_pos]] = out[i]
-
-        self.assertTrue(np.allclose(static_out, dy_out))
-
-        self.assertEqual(len(dy_param_init_value), len(static_param_init_value))
-        for key, value in six.iteritems(static_param_init_value):
-            self.assertTrue(np.allclose(value, dy_param_init_value[key]))
-            self.assertTrue(np.isfinite(value.all()))
-            self.assertFalse(np.isnan(value.any()))
-
-        self.assertEqual(len(dy_grad_value), len(static_grad_value))
-        for key, value in six.iteritems(static_grad_value):
-            # TODO(minqiyang): find a way to align the gradient
-            self.assertTrue(np.allclose(value, dy_grad_value[key]))
-            self.assertTrue(np.isfinite(value.all()))
-            self.assertFalse(np.isnan(value.any()))
-
-        self.assertEqual(len(dy_param_value), len(static_param_value))
-        for key, value in six.iteritems(static_param_value):
-            self.assertTrue(np.allclose(value, dy_param_value[key]))
-            self.assertTrue(np.isfinite(value.all()))
-            self.assertFalse(np.isnan(value.any()))
-
-    def test_resnet_cpu_float32(self):
-        seed = 90
-
-        batch_size = train_parameters["batch_size"]
-        batch_num = 1
-        with fluid.imperative.guard(place=fluid.CPUPlace()):
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
-
-            resnet = ResNet()
-            optimizer = optimizer_setting(train_parameters)
-            np.random.seed(seed)
-            import random
-            random.seed = seed
-            train_reader = paddle.batch(
-                paddle.dataset.flowers.train(use_xmap=False),
-                batch_size=batch_size)
-
-            dy_param_init_value = {}
-            for param in fluid.default_main_program().global_block(
-            ).all_parameters():
-                dy_param_init_value[param.name] = param._numpy()
-
-            for batch_id, data in enumerate(train_reader()):
-                if batch_id >= batch_num:
-                    break
-
-                dy_x_data = np.array(
-                    [x[0].reshape(3, 224, 224) for x in data]).astype('float32')
-                y_data = np.array([x[1] for x in data]).astype('int64').reshape(
-                    batch_size, 1)
-
-                img = to_variable(dy_x_data)
-                label = to_variable(y_data)
-                label._stop_gradient = True
-
-                out = resnet(img)
-                loss = fluid.layers.cross_entropy(input=out, label=label)
-                avg_loss = fluid.layers.mean(x=loss)
-
-                dy_out = avg_loss._numpy()
-
-                if batch_id == 0:
-                    for param in fluid.default_main_program().global_block(
-                    ).all_parameters():
-                        if param.name not in dy_param_init_value:
-                            dy_param_init_value[param.name] = param._numpy()
-
-                avg_loss._backward()
-
-                dy_grad_value = {}
-                for param in fluid.default_main_program().global_block(
-                ).all_parameters():
-                    if not param.stop_gradient:
-                        np_array = np.array(param._ivar._grad_ivar().value()
-                                            .get_tensor())
-                        dy_grad_value[param.name + core.grad_var_suffix(
-                        )] = np_array
-
-                optimizer.minimize(avg_loss)
-
-                dy_param_value = {}
-                for param in fluid.default_main_program().global_block(
-                ).all_parameters():
-                    dy_param_value[param.name] = param._numpy()
-
-        with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
-
-            exe = fluid.Executor(fluid.CPUPlace())
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
             resnet = ResNet()
             optimizer = optimizer_setting(train_parameters)