Fix clip gradient with empty input (#14709)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14709

As titled

Reviewed By: Wakeupbuddy

Differential Revision: D13305554

fbshipit-source-id: 380062d4b0e4f9dc0207a27766cac7b8d05384d5

diff --git a/caffe2/operators/clip_op.cc b/caffe2/operators/clip_op.cc
index b056005..80f4451 100644 (file)
--- a/caffe2/operators/clip_op.cc
+++ b/caffe2/operators/clip_op.cc
@@ -20,7 +20,7 @@ bool ClipGradientOp<float, CPUContext>::RunOnDevice() {
   auto& Y = Input(0);
   auto& dY = Input(1);
   auto* dX = Output(0);
-  CAFFE_ENFORCE_GT(Y.numel(), 0);
+  CAFFE_ENFORCE_GE(Y.numel(), 0);
   CAFFE_ENFORCE_EQ(dY.numel(), Y.numel());
   dX->ResizeLike(Y);
   const float* Ydata = Y.data<float>();

diff --git a/caffe2/operators/clip_op.cu b/caffe2/operators/clip_op.cu
index 167ef21..46422a9 100644 (file)
--- a/caffe2/operators/clip_op.cu
+++ b/caffe2/operators/clip_op.cu
@@ -44,7 +44,7 @@ template <>
 bool ClipOp<float, CUDAContext>::RunOnDevice() {
   auto& X = Input(0);
   auto* Y = Output(0);
-  CAFFE_ENFORCE_GT(X.size(), 0);
+  CAFFE_ENFORCE_GE(X.size(), 0);
   Y->ResizeLike(X);
   ClipKernel<<<
       CAFFE_GET_BLOCKS(X.size()),
@@ -60,7 +60,7 @@ bool ClipGradientOp<float, CUDAContext>::RunOnDevice() {
   auto& Y = Input(0);
   auto& dY = Input(1);
   auto* dX = Output(0);
-  CAFFE_ENFORCE_GT(Y.size(), 0);
+  CAFFE_ENFORCE_GE(Y.size(), 0);
   CAFFE_ENFORCE_EQ(dY.size(), Y.size());
   dX->ResizeLike(Y);
   ClipGradientKernel<<<

diff --git a/caffe2/python/operator_test/clip_op_test.py b/caffe2/python/operator_test/clip_op_test.py
index 46163d3..0c981ce 100644 (file)
--- a/caffe2/python/operator_test/clip_op_test.py
+++ b/caffe2/python/operator_test/clip_op_test.py
@@ -14,16 +14,18 @@ import caffe2.python.serialized_test.serialized_test_util as serial
 
 
 class TestClip(serial.SerializedTestCase):
-    @serial.given(X=hu.tensor(),
+    @serial.given(X=hu.tensor(min_dim=0),
            min_=st.floats(min_value=-2, max_value=0),
            max_=st.floats(min_value=0, max_value=2),
            inplace=st.booleans(),
            **hu.gcs)
     def test_clip(self, X, min_, max_, inplace, gc, dc):
         # go away from the origin point to avoid kink problems
-
-        X[np.abs(X - min_) < 0.05] += 0.1
-        X[np.abs(X - max_) < 0.05] += 0.1
+        if np.isscalar(X):
+            X = np.array([], dtype=np.float32)
+        else:
+            X[np.abs(X - min_) < 0.05] += 0.1
+            X[np.abs(X - max_) < 0.05] += 0.1
 
         def clip_ref(X):
             X = X.clip(min_, max_)
@@ -40,13 +42,15 @@ class TestClip(serial.SerializedTestCase):
         # Gradient check wrt X
         self.assertGradientChecks(gc, op, [X], 0, [0])
 
-    @given(X=hu.tensor(),
+    @given(X=hu.tensor(min_dim=0),
            inplace=st.booleans(),
            **hu.gcs)
     def test_clip_default(self, X, inplace, gc, dc):
         # go away from the origin point to avoid kink problems
-        X += 0.04 * np.sign(X)
-
+        if np.isscalar(X):
+            X = np.array([], dtype=np.float32)
+        else:
+            X += 0.04 * np.sign(X)
         def clip_ref(X):
             return (X,)