Document and test explicit padding for separable_conv2d.

Explicit padding already worked for separable conv2d once support was added to depthwise conv2d, but it was not documented or tested.

PiperOrigin-RevId: 308897692
Change-Id: Ic8a58eb26c814f4d9fefdaec4471cfff20a8c971

tensorflow/python/kernel_tests/conv_ops_test.py

@@ -2662,6 +2662,8 @@ class SeparableConv2DTest(test.TestCase):
       if data_format == "NCHW":
         real_t1 = array_ops.transpose(t1, [0, 3, 1, 2])
         strides = [1, 1, stride, stride]
+        if isinstance(padding, list):
+          padding = [padding[0], padding[3], padding[1], padding[2]]
 
       conv = nn_impl.separable_conv2d(
           real_t1,
@@ -2756,6 +2758,45 @@ class SeparableConv2DTest(test.TestCase):
       return
     self._testSeparableConv2DEqualInputOutputDepth("NCHW")
 
+  def _testSeparableConv2dExplicitPadding(self, data_format):
+    tensor_in_sizes = [1, 4, 4, 2]
+    depthwise_filter_in_sizes = [2, 2, 2, 3]
+    pointwise_filter_in_sizes = [1, 1, 6, 7]
+    padding = [[0, 0], [1, 2], [3, 4], [0, 0]]
+    with self.cached_session(use_gpu=True):
+      # Compute the 'expected' values by manually padding before calling
+      # separable_conv2d
+      t1 = self._InitValues(tensor_in_sizes)
+      t1 = array_ops.pad(t1, padding)
+      f1 = self._InitValues(depthwise_filter_in_sizes)
+      f1.set_shape(depthwise_filter_in_sizes)
+      f2 = self._InitValues(pointwise_filter_in_sizes)
+      conv = nn_impl.separable_conv2d(
+          t1,
+          f1,
+          f2,
+          strides=[1, 1, 1, 1],
+          padding="VALID",
+          data_format="NHWC")
+      expected = self.evaluate(conv)
+      expected = np.ravel(expected)
+    self._VerifyValues(
+        tensor_in_sizes=tensor_in_sizes,
+        depthwise_filter_in_sizes=depthwise_filter_in_sizes,
+        pointwise_filter_in_sizes=pointwise_filter_in_sizes,
+        stride=1,
+        padding=padding,
+        expected=expected,
+        data_format=data_format)
+
+  def testSeparableConv2dExplicitPadding(self):
+    self._testSeparableConv2dExplicitPadding("NHWC")
+
+  def testSeparableConv2dExplicitPaddingNCHW(self):
+    if not test.is_gpu_available():
+      return
+    self._testSeparableConv2dExplicitPadding("NCHW")
+
 
 class DeepConv2DTest(test.TestCase):
 

tensorflow/python/ops/nn_impl.py

@@ -974,8 +974,14 @@ def separable_conv2d(input,
       filter to mix channels after `depthwise_filter` has convolved spatially.
     strides: 1-D of size 4.  The strides for the depthwise convolution for
       each dimension of `input`.
-    padding: A string, either `'VALID'` or `'SAME'`.  The padding algorithm.
-      See the "returns" section of `tf.nn.convolution` for details.
+    padding: Controls how to pad the image before applying the depthwise
+      convolution. Can be the string `"SAME"` or `"VALID"` indicating the type
+      of padding algorithm to use, or a Python list indicating the explicit
+      paddings at the start and end of each dimension. When explicit padding is
+      used and data_format is `"NHWC"`, this should be in the form `[[0, 0],
+      [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]]`. When explicit
+      padding used and data_format is `"NCHW"`, this should be in the form
+      `[[0, 0], [0, 0], [pad_top, pad_bottom], [pad_left, pad_right]]`.
     rate: 1-D of size 2. The dilation rate in which we sample input values
       across the `height` and `width` dimensions in atrous convolution. If it is
       greater than 1, then all values of strides must be 1.
@@ -1076,8 +1082,14 @@ def separable_conv2d_v2(
       `depthwise_filter` has convolved spatially.
     strides: 1-D of size 4.  The strides for the depthwise convolution for each
       dimension of `input`.
-    padding: A string, either `'VALID'` or `'SAME'`.  The padding algorithm. See
-      the "returns" section of `tf.nn.convolution` for details.
+    padding: Controls how to pad the image before applying the depthwise
+      convolution. Can be the string `"SAME"` or `"VALID"` indicating the type
+      of padding algorithm to use, or a Python list indicating the explicit
+      paddings at the start and end of each dimension. When explicit padding is
+      used and data_format is `"NHWC"`, this should be in the form `[[0, 0],
+      [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]]`. When explicit
+      padding used and data_format is `"NCHW"`, this should be in the form
+      `[[0, 0], [0, 0], [pad_top, pad_bottom], [pad_left, pad_right]]`.
     data_format: The data format for input. Either "NHWC" (default) or "NCHW".
     dilations: 1-D of size 2. The dilation rate in which we sample input values
       across the `height` and `width` dimensions in atrous convolution. If it is