Add detailed error messages for nce layer (#20231)

* Add detailed error messages for nce layer

test=develop

* Fix some problems

test=develop

* Fix unit test coverage

test=develop

paddle/fluid/operators/nce_op.cc

@@ -27,31 +27,52 @@ class NCEOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext *ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("Input"));
-    PADDLE_ENFORCE(ctx->HasInput("Label"));
-    PADDLE_ENFORCE(ctx->HasInput("Weight"));
-    PADDLE_ENFORCE(ctx->HasOutput("Cost"));
-    PADDLE_ENFORCE(ctx->HasOutput("SampleLogits"));
-    PADDLE_ENFORCE(ctx->HasOutput("SampleLabels"));
+    PADDLE_ENFORCE_EQ(ctx->HasInput("Input"), true);
+    PADDLE_ENFORCE_EQ(ctx->HasInput("Label"), true);
+    PADDLE_ENFORCE_EQ(ctx->HasInput("Weight"), true);
+    PADDLE_ENFORCE_EQ(ctx->HasOutput("Cost"), true);
+    PADDLE_ENFORCE_EQ(ctx->HasOutput("SampleLogits"), true);
+    PADDLE_ENFORCE_EQ(ctx->HasOutput("SampleLabels"), true);
 
     auto x_dims = ctx->GetInputDim("Input");
     auto label_dims = ctx->GetInputDim("Label");
     if (ctx->IsRuntime() || (x_dims[0] > 0 && label_dims[0] > 0)) {
-      PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0]);
+      PADDLE_ENFORCE_EQ(
+          x_dims[0], label_dims[0],
+          "ShapeError: the first dimension of Input(Input) and Input(Label) "
+          "should be equal in runtime. But received: Input(Input)'s shape = "
+          "[%s] with 1st dim =  %d, Input(Label)'s shape = [%s] with 1st "
+          "dim = %d.",
+          x_dims, x_dims[0], label_dims, label_dims[0]);
     }
     int num_true_classes = label_dims.size() == 2 ? label_dims[1] : 1;
     if (ctx->HasInput("Bias")) {
-      PADDLE_ENFORCE_EQ(ctx->GetInputDim("Weight")[0],
-                        ctx->GetInputDim("Bias")[0]);
+      PADDLE_ENFORCE_EQ(
+          ctx->GetInputDim("Weight")[0], ctx->GetInputDim("Bias")[0],
+          "ShapeError: the first dimension of Input(Weight) and Input(Bias) "
+          "should be equal. But received: Input(Weight)'s shape = [%s] with "
+          "1st dim = %d, Input(Bias)'s shape = [%s] with 1st dim = %d.",
+          ctx->GetInputDim("Weight"), ctx->GetInputDim("Weight")[0],
+          ctx->GetInputDim("Bias"), ctx->GetInputDim("Bias")[0]);
     }
     auto num_neg_samples = ctx->Attrs().Get<int>("num_neg_samples");
     auto num_total_classes = ctx->Attrs().Get<int>("num_total_classes");
     std::vector<int> custom_neg_classes =
         ctx->Attrs().Get<std::vector<int>>("custom_neg_classes");
-    PADDLE_ENFORCE_EQ(num_total_classes, ctx->GetInputDim("Weight")[0]);
+    PADDLE_ENFORCE_EQ(
+        num_total_classes, ctx->GetInputDim("Weight")[0],
+        "ShapeError: the number of total classes should be equal to the first "
+        "dimension of Input(Weight). But received: Attr(num_total_classes) = "
+        "%d, Input(Weight)'s shape = [%s] with 1st dim = %d.",
+        num_total_classes, ctx->GetInputDim("Weight"),
+        ctx->GetInputDim("Weight")[0]);
     if (custom_neg_classes.size() > 0) {
-      PADDLE_ENFORCE_EQ(custom_neg_classes.size(),
-                        static_cast<size_t>(num_neg_samples));
+      PADDLE_ENFORCE_EQ(
+          custom_neg_classes.size(), static_cast<size_t>(num_neg_samples),
+          "ShapeError: the size of Attr(custom_neg_classes) should be equal "
+          "to the number of negative samples. But received: "
+          "custom_neg_classes.size() = %d, num_neg_samples = %d.",
+          custom_neg_classes.size(), num_neg_samples);
     }
     // set dims of output(Out)
     std::vector<int64_t> out_dims;

paddle/fluid/operators/nce_op.h

@@ -102,9 +102,27 @@ class NCEKernel : public framework::OpKernel<T> {
         auto dist_alias = context.Input<Tensor>("CustomDistAlias");
         auto dist_alias_probs = context.Input<Tensor>("CustomDistAliasProbs");
 
-        PADDLE_ENFORCE_EQ(dist_probs->numel(), num_total_classes);
-        PADDLE_ENFORCE_EQ(dist_alias->numel(), num_total_classes);
-        PADDLE_ENFORCE_EQ(dist_alias_probs->numel(), num_total_classes);
+        PADDLE_ENFORCE_EQ(
+            dist_probs->numel(), num_total_classes,
+            "ShapeError: The number of elements in Input(CustomDistProbs) "
+            "should be equal to the number of total classes. But Received: "
+            "Input(CustomDistProbs).numel() = %d, Attr(num_total_classes) "
+            "= %d.",
+            dist_probs->numel(), num_total_classes);
+        PADDLE_ENFORCE_EQ(
+            dist_alias->numel(), num_total_classes,
+            "ShapeError: The number of elements in Input(CustomDistAlias) "
+            "should be equal to the number of total classes. But Received: "
+            "Input(CustomDistAlias).numel() = %d, Attr(num_total_classes) "
+            "= %d.",
+            dist_alias->numel(), num_total_classes);
+        PADDLE_ENFORCE_EQ(
+            dist_alias_probs->numel(), num_total_classes,
+            "ShapeError: The number of elements in Input(CustomDistAliasProbs) "
+            "should be equal to the number of total classes. But Received: "
+            "Input(CustomDistAliasProbs).numel() = %d, "
+            "Attr(num_total_classes) = %d.",
+            dist_alias_probs->numel(), num_total_classes);
 
         const float *probs_data = dist_probs->data<float>();
         const int *alias_data = dist_alias->data<int>();
@@ -121,7 +139,11 @@ class NCEKernel : public framework::OpKernel<T> {
     const int64_t *sample_labels_data = sample_labels->data<int64_t>();
 
     for (int x = 0; x < sample_labels->numel(); x++) {
-      PADDLE_ENFORCE_GE(sample_labels_data[x], 0, "nce sample label %d", x);
+      PADDLE_ENFORCE_GE(sample_labels_data[x], 0,
+                        "ValueError: Every sample label should be "
+                        "non-negative. But received: "
+                        "Input(SampleLabels)[%d] = %d",
+                        x, sample_labels_data[x]);
     }
 
     auto sample_out = context.Output<Tensor>("SampleLogits");
@@ -289,9 +311,27 @@ class NCEGradKernel : public framework::OpKernel<T> {
         auto dist_alias = context.Input<Tensor>("CustomDistAlias");
         auto dist_alias_probs = context.Input<Tensor>("CustomDistAliasProbs");
 
-        PADDLE_ENFORCE_EQ(dist_probs->numel(), num_total_classes);
-        PADDLE_ENFORCE_EQ(dist_alias->numel(), num_total_classes);
-        PADDLE_ENFORCE_EQ(dist_alias_probs->numel(), num_total_classes);
+        PADDLE_ENFORCE_EQ(
+            dist_probs->numel(), num_total_classes,
+            "ShapeError: The number of elements in Input(CustomDistProbs) "
+            "should be equal to the number of total classes. But Received: "
+            "Input(CustomDistProbs).numel() = %d, Attr(num_total_classes) "
+            "= %d.",
+            dist_probs->numel(), num_total_classes);
+        PADDLE_ENFORCE_EQ(
+            dist_alias->numel(), num_total_classes,
+            "ShapeError: The number of elements in Input(CustomDistAlias) "
+            "should be equal to the number of total classes. But Received: "
+            "Input(CustomDistAlias).numel() = %d, Attr(num_total_classes) "
+            "= %d.",
+            dist_alias->numel(), num_total_classes);
+        PADDLE_ENFORCE_EQ(
+            dist_alias_probs->numel(), num_total_classes,
+            "ShapeError: The number of elements in Input(CustomDistAliasProbs) "
+            "should be equal to the number of total classes. But Received: "
+            "Input(CustomDistAliasProbs).numel() = %d, "
+            "Attr(num_total_classes) = %d.",
+            dist_alias_probs->numel(), num_total_classes);
 
         const float *probs_data = dist_probs->data<float>();
         const int *alias_data = dist_alias->data<int>();

python/paddle/fluid/layers/nn.py

@@ -7033,8 +7033,23 @@ def nce(input,
                        custom_dist=dist)
     """
     helper = LayerHelper('nce', **locals())
-    assert isinstance(input, Variable)
-    assert isinstance(label, Variable)
+
+    if not isinstance(input, Variable):
+        raise TypeError(
+            "The type of 'input' in nce layer must be Variable, but received %s"
+            % (type(input)))
+    if not isinstance(label, Variable):
+        raise TypeError(
+            "The type of 'label' in nce layer must be Variable, but received %s"
+            % (type(label)))
+    if convert_dtype(input.dtype) not in ['float32', 'float64']:
+        raise TypeError(
+            "The data type of 'input' in nce layer must be float32 or float64, but received %s."
+            % (convert_dtype(input.dtype)))
+    if convert_dtype(label.dtype) not in ['int64']:
+        raise TypeError(
+            "The data type of 'label' in nce layer must be int64, but received %s."
+            % (convert_dtype(label.dtype)))
 
     dim = input.shape[1]
     num_true_class = label.shape[1]

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

@@ -19,6 +19,7 @@ import unittest
 
 import paddle.fluid as fluid
 import paddle.fluid.initializer as initializer
+from paddle.fluid import Program, program_guard
 
 from op_test import OpTest
 
@@ -219,5 +220,37 @@ class TestNCECase1SelectedRows(unittest.TestCase):
         self.assertEqual(rets[0], rets[1])
 
 
+class TestNCE_OpError(OpTest):
+    def test_errors(self):
+        with program_guard(Program(), Program()):
+            input1 = fluid.create_lod_tensor(
+                np.array([0.0, 3.0, 2.0, 4.0]), [[1, 1, 2]], fluid.CPUPlace())
+            label1 = fluid.layers.data(
+                name='label1', shape=[-1, 4], dtype="int64")
+            # the input(input) of nce layer must be Variable.
+            self.assertRaises(TypeError, fluid.layers.nce, input1, label1, 5)
+
+            input2 = fluid.layers.data(
+                name='input2', shape=[-1, 4], dtype="float32")
+            label2 = fluid.create_lod_tensor(
+                np.array([0.0, 3.0, 2.0, 4.0]), [[1, 1, 2]], fluid.CPUPlace())
+            # the input(label) of nce layer must be Variable.
+            self.assertRaises(TypeError, fluid.layers.nce, input2, label2, 5)
+
+            input3 = fluid.layers.data(
+                name='input3', shape=[-1, 4], dtype="float16")
+            label3 = fluid.layers.data(
+                name='label3', shape=[-1, 1], dtype="int64")
+            # the data type of input(input) must be float32 or float64.
+            self.assertRaises(TypeError, fluid.layers.nce, input3, label3, 5)
+
+            input4 = fluid.layers.data(
+                name='input4', shape=[-1, 4], dtype="float32")
+            label4 = fluid.layers.data(
+                name='label4', shape=[-1, 1], dtype="int32")
+            # the data type of input(label) must be int64.
+            self.assertRaises(TypeError, fluid.layers.nce, input4, label4, 5)
+
+
 if __name__ == '__main__':
     unittest.main()