ont hot support tensor depth (#16972)

phlrain · web-flow · commit 82358bfdc127 · 2019-06-04T10:41:04.000+08:00
* support some input tensor remain on cpu; test=develop

* fix input = none; test=develop

* fix unfound bug; test=develop

* fix proto None case; test=develop

* fix bug; test=develop

* fix proto null bug; test=develop

* remove conv check; test=develop

* fix test bug; test=develop

* move fill constant; test=develop

* no change in proto; test=develop

* fix bug; test=develop

* change attr detph name; test=develop

* remove remain cpu; test=develop

* fix bug; test=develop

* merge develop; test=develop

* fix one_hot bug; test=develop

* fix bug; test=develop

* fix bug; test=develop

* fix bug; test=develop

* fix python api bug; test=develop
diff --git a/paddle/fluid/operators/one_hot_op.cc b/paddle/fluid/operators/one_hot_op.cc
@@ -13,6 +13,8 @@
 // limitations under the License.
 
 #include "paddle/fluid/operators/one_hot_op.h"
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/framework.pb.h"
 
 namespace paddle {
@@ -34,15 +36,34 @@ class OneHotOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE_GE(x_dims[x_dims.size() - 1], 1U,
                         "Last dimension of Input(X) should be 1.");
     }
-    int depth = ctx->Attrs().Get<int>("depth");
-
-    PADDLE_ENFORCE_GT(depth, 0, "Should provide a positive depth (%d).", depth);
 
     framework::DDim out_dims(x_dims);
+    int depth = ctx->Attrs().Get<int>("depth");
+    if (ctx->HasInput("depth_tensor")) {
+      depth = -1;
+    }
+
     out_dims[out_dims.size() - 1] = depth;
     ctx->SetOutputDim("Out", out_dims);
     ctx->ShareLoD("X", /* --> */ "Out");
   }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
+                                   ctx.device_context());
+  }
+
+  framework::OpKernelType GetKernelTypeForVar(
+      const std::string& var_name, const Tensor& tensor,
+      const framework::OpKernelType& expected_kernel_type) const override {
+    if (var_name == "depth_tensor") {
+      return expected_kernel_type;
+    }
+    return framework::OpKernelType(expected_kernel_type.data_type_,
+                                   tensor.place(), tensor.layout());
+  }
 };
 
 class OneHotOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -52,11 +73,15 @@ class OneHotOpMaker : public framework::OpProtoAndCheckerMaker {
              "(LoDTensor, LoDTensor<int>) Input variable with rank at least 2. "
              "The last dimension of X should be 1. Each value of X is an index "
              "to indicate the position.");
+    AddInput("depth_tensor", "(Tensor, Tensor<int>), Length of one-hot vector")
+        .AsDispensable();
     AddOutput("Out",
               "(Tensor, Tensor<float>) Output tensor with same rank as X. "
               "The tensor consists of one-hot representations of values in X.");
+
     AddAttr<int>("depth",
-                 "A positive integer to specify the length of one-hot vector.");
+                 "A positive integer to specify the length of one-hot vector.")
+        .SetDefault(-1);
     AddAttr<int>("dtype",
                  "An integer to specify the data type of one-hot "
                  "vector. The default value is FP32.")
diff --git a/paddle/fluid/operators/one_hot_op.cu b/paddle/fluid/operators/one_hot_op.cu
@@ -62,8 +62,25 @@ class OneHotCUDAKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* in = context.Input<LoDTensor>("X");
     auto* out = context.Output<LoDTensor>("Out");
-    int depth = context.Attr<int>("depth");
 
+    int depth = -1;
+    if (context.HasInput("depth_tensor")) {
+      auto* depth_tensor = context.Input<framework::Tensor>("depth_tensor");
+      if (platform::is_gpu_place(depth_tensor->place())) {
+        framework::Tensor temp;
+        TensorCopySync(*depth_tensor, platform::CPUPlace(), &temp);
+        depth = *temp.data<int32_t>();
+      } else {
+        depth = *depth_tensor->data<int32_t>();
+      }
+
+      auto in_dims = in->dims();
+      framework::DDim out_dims(in_dims);
+      out_dims[out_dims.size() - 1] = depth;
+      out->Resize(out_dims);
+    } else {
+      depth = context.Attr<int>("depth");
+    }
     framework::VisitDataType(
         static_cast<framework::proto::VarType::Type>(
             context.Attr<int>("dtype")),
diff --git a/paddle/fluid/operators/one_hot_op.h b/paddle/fluid/operators/one_hot_op.h
@@ -49,13 +49,23 @@ struct OneHotOpFunctor {
 };
 
 using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
 template <typename DeviceContext, typename T>
 class OneHotKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* in = context.Input<LoDTensor>("X");
     auto* out = context.Output<LoDTensor>("Out");
     int depth = context.Attr<int>("depth");
+    if (context.HasInput("depth_tensor")) {
+      auto* depth_tensor = context.Input<Tensor>("depth_tensor");
+      auto* depth_data = depth_tensor->data<int32_t>();
+      depth = depth_data[0];
+      auto in_dims = in->dims();
+      framework::DDim out_dims(in_dims);
+      out_dims[out_dims.size() - 1] = depth;
+      out->Resize(out_dims);
+    }
 
     framework::VisitDataType(
         static_cast<framework::proto::VarType::Type>(
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
@@ -1013,7 +1013,7 @@ def __init__(self,
                 return
             if type is None:
                 raise ValueError(
-                    "`type` to initilized an Operator can not be None.")
+                    "`type` to initialized an Operator can not be None.")
             else:
                 callstack_var_name = op_maker.kOpCreationCallstackAttrName()
                 op_attrs[callstack_var_name] = list(
@@ -1036,7 +1036,6 @@ def find_name(var_list, name):
                     found = find_name(inputs, in_proto.name)
                     assert found or in_proto.dispensable, "Input {} not found".format(
                         in_proto.name)
-
                     if found:
                         in_args = inputs[in_proto.name]
                         if not isinstance(in_args, list):
@@ -1046,13 +1045,17 @@ def find_name(var_list, name):
                                 "Input %s expects only one input, but %d are given."
                                 % (in_proto.name, len(in_args)))
                         in_arg_names = []
-                        for arg in in_args:
+                        for index, arg in enumerate(in_args):
                             if isinstance(arg, six.string_types):
                                 in_arg_names.append(arg)
                             elif isinstance(arg, six.binary_type):
                                 in_arg_names.append(arg.decode())
-                            else:
+                            elif isinstance(arg, Variable):
                                 in_arg_names.append(cpt.to_text(arg.name))
+                            else:
+                                raise ValueError(
+                                    "not suprt args type , should be[ string_type, binary_type, Varibale]"
+                                )
                         self.desc.set_input(in_proto.name, in_arg_names)
                     else:
                         self.desc.set_input(in_proto.name, [])
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
@@ -6564,11 +6564,24 @@ def one_hot(input, depth):
             one_hot_label = fluid.layers.one_hot(input=label, depth=10)
     """
     helper = LayerHelper("one_hot", **locals())
+
     one_hot_out = helper.create_variable_for_type_inference(dtype='float32')
+
+    if in_dygraph_mode():
+        inputs = {'X': input}
+        attrs = {'depth': depth}
+    else:
+        if not isinstance(depth, Variable):
+            # user attribute 
+            inputs = {'X': input}
+            attrs = {'depth': depth}
+        else:
+            inputs = {'X': input, 'depth_tensor': depth}
+            attrs = {}
     helper.append_op(
         type="one_hot",
-        inputs={'X': input},
-        attrs={'depth': depth},
+        inputs=inputs,
+        attrs=attrs,
         outputs={'Out': one_hot_out},
         stop_gradient=True)
     return one_hot_out
diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py
@@ -1267,6 +1267,12 @@ def make_scatter(self):
             out = layers.scatter(input=x, index=idx, updates=updates)
             return (out)
 
+    def make_one_hot(self):
+        with fluid.framework._dygraph_place_guard(place=fluid.CPUPlace()):
+            label = self._get_data(name="label", shape=[1], dtype="int32")
+            one_hot_label = layers.one_hot(input=label, depth=10)
+            return (one_hot_label)
+
     def make_label_smooth(self):
         # TODO(minqiyang): support gpu ut
         self._force_to_use_cpu = True
diff --git a/python/paddle/fluid/tests/unittests/test_one_hot_op.py b/python/paddle/fluid/tests/unittests/test_one_hot_op.py
@@ -28,10 +28,34 @@ class TestOneHotOp(OpTest):
     def setUp(self):
         self.op_type = 'one_hot'
         depth = 10
+        depth_np = np.array(10).astype('int32')
         dimension = 12
         x_lod = [[4, 1, 3, 3]]
         x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
-        x = np.array(x).astype('int').reshape([sum(x_lod[0]), 1])
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod), 'depth_tensor': depth_np}
+        self.attrs = {'dtype': int(core.VarDesc.VarType.FP32)}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestOneHotOp_attr(OpTest):
+    def setUp(self):
+        self.op_type = 'one_hot'
+        depth = 10
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
 
         out = np.zeros(shape=(np.product(x.shape[:-1]),
                               depth)).astype('float32')
@@ -40,21 +64,45 @@ def setUp(self):
             out[i, x[i]] = 1.0
 
         self.inputs = {'X': (x, x_lod)}
-        self.attrs = {'depth': depth, 'dtype': int(core.VarDesc.VarType.FP32)}
+        self.attrs = {'dtype': int(core.VarDesc.VarType.FP32), 'depth': depth}
         self.outputs = {'Out': (out, x_lod)}
 
     def test_check_output(self):
         self.check_output()
 
 
 class TestOneHotOp_default_dtype(OpTest):
+    def setUp(self):
+        self.op_type = 'one_hot'
+        depth = 10
+        depth_np = np.array(10).astype('int32')
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod), 'depth_tensor': depth_np}
+        self.attrs = {}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestOneHotOp_default_dtype_attr(OpTest):
     def setUp(self):
         self.op_type = 'one_hot'
         depth = 10
         dimension = 12
         x_lod = [[4, 1, 3, 3]]
         x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
-        x = np.array(x).astype('int').reshape([sum(x_lod[0]), 1])
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
 
         out = np.zeros(shape=(np.product(x.shape[:-1]),
                               depth)).astype('float32')
diff --git a/python/paddle/fluid/tests/unittests/test_operator_desc.py b/python/paddle/fluid/tests/unittests/test_operator_desc.py
@@ -33,7 +33,7 @@ def test_error_type(self):
         except ValueError as v_err:
             self.assertEqual(
                 cpt.get_exception_message(v_err),
-                "`type` to initilized an Operator can not be None.")
+                "`type` to initialized an Operator can not be None.")
         try:
             block.append_op(type="no_such_op")
             self.assertFail()
diff --git a/python/paddle/fluid/tests/unittests/test_recordio_reader.py b/python/paddle/fluid/tests/unittests/test_recordio_reader.py
@@ -86,3 +86,7 @@ def test_shuffle_reader(self):
     def test_double_buffer_reader(self):
         self.test_main(decorator_callback=lambda reader: fluid.layers.io.double_buffer(reader,
                                                                                        place='cuda:0' if fluid.core.is_compiled_with_cuda() else 'cpu'))
+
+
+if __name__ == '__main__':
+    unittest.main()
