supprt high rank; test=develop

paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc

@@ -34,15 +34,22 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel {
 
     auto x_dims = ctx->GetInputDim("X");
     auto labels_dims = ctx->GetInputDim("Label");
-    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(labels_dims.size(), 2,
-                      "Input(Label)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(X) and Input(Label) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Label) should "
-                      "be equal.");
+
+    int rank = x_dims.size();
+    PADDLE_ENFORCE_EQ(rank, labels_dims.size(),
+                      "Input(X) and Input(Label) shall have the same rank.");
+    bool check = true;
+    if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
+                                framework::product(labels_dims) <= 0)) {
+      check = false;
+    }
+
+    if (check) {
+      PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank),
+                        framework::slice_ddim(labels_dims, 0, rank),
+                        "Input(X) and Input(Label) shall have the same shape "
+                        "except the last dimension.");
+    }
 
     ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
@@ -65,23 +72,24 @@ class SigmoidCrossEntropyWithLogitsGradOp
     auto x_dims = ctx->GetInputDim("X");
     auto labels_dims = ctx->GetInputDim("Label");
     auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
-    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(labels_dims.size(), 2,
-                      "Input(Label)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(dout_dims.size(), 2,
-                      "Input(Out@Grad)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(X) and Input(Label) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Label) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[0], dout_dims[0],
-                      "The 1st dimension of Input(X) and Input(Out@Grad) "
-                      "should be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[1], dout_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Out@Grad) "
-                      "should be equal.");
+
+    int rank = x_dims.size();
+    bool check = true;
+    if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
+                                framework::product(labels_dims) <= 0)) {
+      check = false;
+    }
+
+    if (check) {
+      PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank),
+                        framework::slice_ddim(labels_dims, 0, rank),
+                        "Input(X) and Input(Label) shall have the same shape.");
+
+      PADDLE_ENFORCE_EQ(
+          framework::slice_ddim(x_dims, 0, rank),
+          framework::slice_ddim(dout_dims, 0, rank),
+          "Input(X) and Input(Out@Grad) shall have the same shape.");
+    }
 
     ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
   }

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

@@ -149,5 +149,98 @@ class TestSigmoidCrossEntropyWithNorm(OpTest):
         self.check_grad(['X'], 'Out')
 
 
+class TestSigmoidCrossEntropyWithLogitsOp5(OpTest):
+    """Test sigmoid_cross_entropy_with_logit_op with probabalistic label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        batch_size = [10, 10]
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype("float32")),
+            'Label': np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+            .astype("float32")
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestSigmoidCrossEntropyWithNorm2(OpTest):
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        batch_size = [10, 10]
+        num_classes = 20
+        ignore_index = -1
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype("float32")),
+            'Label': np.random.randint(-1, 2, tuple(batch_size + [num_classes]))
+            .astype("float32")
+        }
+        self.attrs = {'ignore_index': ignore_index, 'normalize': True}
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        out = -term1 - term2
+        out[np.where(self.inputs['Label'] == ignore_index)] = 0
+        if self.attrs['normalize']:
+            out = out / float(
+                np.where(self.inputs['Label'] != ignore_index)[0].size)
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestSigmoidCrossEntropyWithLogitsOp6(OpTest):
+    """Test sigmoid_cross_entropy_with_logit_op with binary label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        batch_size = [10, 10]
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype("float32")),
+            'Label': np.random.randint(0, 2, tuple(batch_size + [num_classes]))
+            .astype("float32")
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
 if __name__ == '__main__':
     unittest.main()