BEGIN_PUBLIC

Allow a different output shape from the input in tf.contrib.image.transform (tensorflow#17011).
END_PUBLIC

RELNOTES: Allow a different output shape from the input in tf.contrib.image.transform.

Thanks qyu@ for making the original change and fixing a few other prior issues!

Automated rollback of commit 07fdb69

PiperOrigin-RevId: 208248183

tensorflow/contrib/image/kernels/image_ops.cc

@@ -71,6 +71,7 @@ class ImageProjectiveTransform : public OpKernel {
   void Compute(OpKernelContext* ctx) override {
     const Tensor& images_t = ctx->input(0);
     const Tensor& transform_t = ctx->input(1);
+    const Tensor& shape_t = ctx->input(2);
     OP_REQUIRES(ctx, images_t.shape().dims() == 4,
                 errors::InvalidArgument("Input images must have rank 4"));
     OP_REQUIRES(ctx,
@@ -81,11 +82,28 @@ class ImageProjectiveTransform : public OpKernel {
                      ProjectiveGenerator<Device, T>::kNumParameters),
                 errors::InvalidArgument(
                     "Input transform should be num_images x 8 or 1 x 8"));
-    auto images = images_t.tensor<T, 4>();
-    auto transform = transform_t.matrix<float>();
+    OP_REQUIRES(ctx, shape_t.dims() == 1,
+                errors::InvalidArgument("output shape must be 1-dimensional",
+                                        shape_t.shape().DebugString()));
+    OP_REQUIRES(ctx, shape_t.NumElements() == 2,
+                errors::InvalidArgument("output shape must have two elements",
+                                        shape_t.shape().DebugString()));
+    auto shape_vec = shape_t.vec<int32>();
+    int32 out_height = shape_vec(0);
+    int32 out_width = shape_vec(1);
+    OP_REQUIRES(ctx, out_height > 0 && out_width > 0,
+                errors::InvalidArgument("output dimensions must be positive"));
+
     Tensor* output_t;
-    OP_REQUIRES_OK(ctx, ctx->allocate_output(0, images_t.shape(), &output_t));
+    OP_REQUIRES_OK(ctx, ctx->allocate_output(
+                            0,
+                            TensorShape({images_t.dim_size(0), out_height,
+                                         out_width, images_t.dim_size(3)}),
+                            &output_t));
     auto output = output_t->tensor<T, 4>();
+    auto images = images_t.tensor<T, 4>();
+    auto transform = transform_t.matrix<float>();
+
     (FillProjectiveTransform<Device, T>(interpolation_))(
         ctx->eigen_device<Device>(), &output, images, transform);
   }
@@ -129,10 +147,11 @@ TF_CALL_double(DECLARE_FUNCTOR);
 
 }  // end namespace functor
 
-#define REGISTER(TYPE)                                        \
-  REGISTER_KERNEL_BUILDER(Name("ImageProjectiveTransform")    \
-                              .Device(DEVICE_GPU)             \
-                              .TypeConstraint<TYPE>("dtype"), \
+#define REGISTER(TYPE)                                       \
+  REGISTER_KERNEL_BUILDER(Name("ImageProjectiveTransform")   \
+                              .Device(DEVICE_GPU)            \
+                              .TypeConstraint<TYPE>("dtype") \
+                              .HostMemory("output_shape"),   \
                           ImageProjectiveTransform<GPUDevice, TYPE>)
 
 TF_CALL_uint8(REGISTER);

tensorflow/contrib/image/kernels/image_ops.h

@@ -167,7 +167,7 @@ struct FillProjectiveTransform {
   void operator()(const Device& device, OutputType* output,
                   const InputType& images,
                   const TransformsType& transform) const {
-    output->device(device) = images.generate(
+    output->device(device) = output->generate(
         ProjectiveGenerator<Device, T>(images, transform, interpolation_));
   }
 };

tensorflow/contrib/image/ops/image_ops.cc

@@ -19,23 +19,66 @@ limitations under the License.
 
 namespace tensorflow {
 
+using shape_inference::DimensionHandle;
 using shape_inference::InferenceContext;
 using shape_inference::ShapeHandle;
 
+namespace {
+
+// Sets output[0] to shape [batch_dim,height,width,channel_dim], where
+// height and width come from the size_tensor.
+Status SetOutputToSizedImage(InferenceContext* c, DimensionHandle batch_dim,
+                             int size_input_idx, DimensionHandle channel_dim) {
+  // Verify shape of size input.
+  ShapeHandle size;
+  TF_RETURN_IF_ERROR(c->WithRank(c->input(size_input_idx), 1, &size));
+  DimensionHandle unused;
+  TF_RETURN_IF_ERROR(c->WithValue(c->Dim(size, 0), 2, &unused));
+
+  // Get size values from the size tensor.
+  const Tensor* size_tensor = c->input_tensor(size_input_idx);
+  DimensionHandle width;
+  DimensionHandle height;
+  if (size_tensor == nullptr) {
+    width = c->UnknownDim();
+    height = c->UnknownDim();
+  } else {
+    // TODO(petewarden) - Remove once we have constant evaluation in C++ only.
+    if (size_tensor->dtype() != DT_INT32) {
+      return errors::InvalidArgument(
+          "Bad size input type for SetOutputToSizedImage: Expected DT_INT32 "
+          "but got ",
+          DataTypeString(size_tensor->dtype()), " for input #", size_input_idx,
+          " in ", c->DebugString());
+    }
+    auto vec = size_tensor->vec<int32>();
+    height = c->MakeDim(vec(0));
+    width = c->MakeDim(vec(1));
+  }
+  c->set_output(0, c->MakeShape({batch_dim, height, width, channel_dim}));
+  return Status::OK();
+}
+
+// TODO(qyu): Move this to core/framework/common_shape_fns.h
+Status ResizeShapeFn(InferenceContext* c) {
+  ShapeHandle input;
+  TF_RETURN_IF_ERROR(c->WithRank(c->input(0), 4, &input));
+  return SetOutputToSizedImage(c, c->Dim(input, 0), 2 /* size_input_idx */,
+                               c->Dim(input, 3));
+}
+
+}  // namespace
+
 // TODO(ringwalt): Add a "fill_mode" argument with "constant", "mirror", etc.
 // TODO(ringwalt): Add a "fill_constant" argument for constant mode (default 0).
-// TODO(ringwalt): Add an "output_shape" argument. This is sufficient to
-// implement "same" and "valid" modes in the Python function.
 REGISTER_OP("ImageProjectiveTransform")
     .Input("images: dtype")
     .Input("transforms: float32")
+    .Input("output_shape: int32")
     .Attr("dtype: {uint8, int32, int64, float16, float32, float64}")
     .Attr("interpolation: string")
     .Output("transformed_images: dtype")
-    .SetShapeFn([](InferenceContext* c) {
-      c->set_output(0, c->input(0));
-      return Status::OK();
-    })
+    .SetShapeFn(ResizeShapeFn)
     .Doc(R"doc(
 Applies the given transform to each of the images.
 
@@ -49,7 +92,7 @@ If one row of `transforms` is `[a0, a1, a2, b0, b1, b2, c0, c1]`, then it maps
 the *output* point `(x, y)` to a transformed *input* point
 `(x', y') = ((a0 x + a1 y + a2) / k, (b0 x + b1 y + b2) / k)`, where
 `k = c0 x + c1 y + 1`. If the transformed point lays outside of the input
-image, the output pixel is set to 0. The output is the same size as the input,
+image, the output pixel is set to 0.
 
 images: 4D `Tensor`, input image(s) in NHWC format.
 transforms: 2D `Tensor`, projective transform(s) to apply to the image(s).

tensorflow/contrib/image/python/kernel_tests/image_ops_test.py

@@ -27,6 +27,7 @@
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import gradient_checker
 from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import random_ops
 from tensorflow.python.platform import googletest
 
 _DTYPES = set(
@@ -194,6 +195,19 @@ def test_bilinear_uint8(self):
            [0.0, 149, 233, 149, 0.0],
            [0.0, 0.0, 87., 0.0, 0.0]])
 
+  def test_rotate_static_shape(self):
+    image = array_ops.diag([1., 2., 3.])
+    result = image_ops.rotate(
+        image, random_ops.random_uniform((), -1, 1), interpolation="BILINEAR")
+    self.assertEqual(image.get_shape(), result.get_shape())
+
+  def test_transform_static_output_shape(self):
+    image = constant_op.constant([[1., 2.], [3., 4.]])
+    result = image_ops.transform(
+        image, random_ops.random_uniform([8], -1, 1),
+        output_shape=constant_op.constant([3, 5]))
+    self.assertAllEqual([3, 5], result.get_shape())
+
   def _test_grad(self, shape_to_test):
     with self.test_session():
       test_image_shape = shape_to_test
@@ -213,10 +227,40 @@ def _test_grad(self, shape_to_test):
           x_init_value=test_image)
       self.assertLess(left_err, 1e-10)
 
+  def _test_grad_different_shape(self, input_shape, output_shape):
+    with self.test_session():
+      test_image_shape = input_shape
+      test_image = np.random.randn(*test_image_shape)
+      test_image_tensor = constant_op.constant(
+          test_image, shape=test_image_shape)
+      test_transform = image_ops.angles_to_projective_transforms(
+          np.pi / 2, 4, 4)
+
+      if len(output_shape) == 2:
+        resize_shape = output_shape
+      elif len(output_shape) == 3:
+        resize_shape = output_shape[0:2]
+      elif len(output_shape) == 4:
+        resize_shape = output_shape[1:3]
+      output = image_ops.transform(
+          images=test_image_tensor,
+          transforms=test_transform,
+          output_shape=resize_shape)
+      left_err = gradient_checker.compute_gradient_error(
+          test_image_tensor,
+          test_image_shape,
+          output,
+          output_shape,
+          x_init_value=test_image)
+      self.assertLess(left_err, 1e-10)
+
   def test_grad(self):
     self._test_grad([16, 16])
     self._test_grad([4, 12, 12])
     self._test_grad([3, 4, 12, 12])
+    self._test_grad_different_shape([16, 16], [8, 8])
+    self._test_grad_different_shape([4, 12, 3], [8, 24, 3])
+    self._test_grad_different_shape([3, 4, 12, 3], [3, 8, 24, 3])
 
 
 class BipartiteMatchTest(test_util.TensorFlowTestCase):

tensorflow/contrib/image/python/ops/image_ops.py

@@ -23,6 +23,7 @@
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import tensor_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import linalg_ops
@@ -40,6 +41,9 @@
 ops.RegisterShape("ImageProjectiveTransform")(common_shapes.call_cpp_shape_fn)
 
 
+# TODO(ringwalt): Support a "reshape" (name used by SciPy) or "expand" (name
+# used by PIL, maybe more readable) mode, which determines the correct
+# output_shape and translation for the transform.
 def rotate(images, angles, interpolation="NEAREST", name=None):
   """Rotate image(s) counterclockwise by the passed angle(s) in radians.
 
@@ -213,7 +217,11 @@ def translations_to_projective_transforms(translations, name=None):
         axis=1)
 
 
-def transform(images, transforms, interpolation="NEAREST", name=None):
+def transform(images,
+              transforms,
+              interpolation="NEAREST",
+              output_shape=None,
+              name=None):
   """Applies the given transform(s) to the image(s).
 
   Args:
@@ -230,6 +238,10 @@ def transform(images, transforms, interpolation="NEAREST", name=None):
        the transform mapping input points to output points. Note that gradients
        are not backpropagated into transformation parameters.
     interpolation: Interpolation mode. Supported values: "NEAREST", "BILINEAR".
+    output_shape: Output dimesion after the transform, [height, width].
+       If None, output is the same size as input image.
+
+    name: The name of the op.
 
   Returns:
     Image(s) with the same type and shape as `images`, with the given
@@ -238,6 +250,7 @@ def transform(images, transforms, interpolation="NEAREST", name=None):
 
   Raises:
     TypeError: If `image` is an invalid type.
+    ValueError: If output shape is not 1-D int32 Tensor.
   """
   with ops.name_scope(name, "transform"):
     image_or_images = ops.convert_to_tensor(images, name="images")
@@ -256,6 +269,17 @@ def transform(images, transforms, interpolation="NEAREST", name=None):
     else:
       raise TypeError("Images should have rank between 2 and 4.")
 
+    if output_shape is None:
+      output_shape = tensor_util.constant_value(
+          array_ops.shape(images)[1:3]) or array_ops.shape(images)[1:3]
+
+    output_shape = ops.convert_to_tensor(
+        output_shape, dtypes.int32, name="output_shape")
+
+    if not output_shape.get_shape().is_compatible_with([2]):
+      raise ValueError("output_shape must be a 1-D Tensor of 2 elements: "
+                       "new_height, new_width")
+
     if len(transform_or_transforms.get_shape()) == 1:
       transforms = transform_or_transforms[None]
     elif transform_or_transforms.get_shape().ndims is None:
@@ -265,8 +289,12 @@ def transform(images, transforms, interpolation="NEAREST", name=None):
       transforms = transform_or_transforms
     else:
       raise TypeError("Transforms should have rank 1 or 2.")
+
     output = gen_image_ops.image_projective_transform(
-        images, transforms, interpolation=interpolation.upper())
+        images,
+        output_shape=output_shape,
+        transforms=transforms,
+        interpolation=interpolation.upper())
     if len(image_or_images.get_shape()) == 2:
       return output[0, :, :, 0]
     elif len(image_or_images.get_shape()) == 3:
@@ -376,14 +404,6 @@ def _image_projective_transform_grad(op, grad):
 
   if image_or_images.dtype.base_dtype not in _IMAGE_DTYPES:
     raise TypeError("Invalid dtype %s." % image_or_images.dtype)
-  if len(image_or_images.get_shape()) == 2:
-    images = image_or_images[None, :, :, None]
-  elif len(image_or_images.get_shape()) == 3:
-    images = image_or_images[None, :, :, :]
-  elif len(image_or_images.get_shape()) == 4:
-    images = image_or_images
-  else:
-    raise TypeError("Images should have rank between 2 and 4")
   if len(transform_or_transforms.get_shape()) == 1:
     transforms = transform_or_transforms[None]
   elif len(transform_or_transforms.get_shape()) == 2:
@@ -396,13 +416,11 @@ def _image_projective_transform_grad(op, grad):
   inverse = linalg_ops.matrix_inverse(transforms)
   transforms = matrices_to_flat_transforms(inverse)
   output = gen_image_ops.image_projective_transform(
-      grad, transforms, interpolation=interpolation)
-  if len(image_or_images.get_shape()) == 2:
-    return [output[0, :, :, 0], None]
-  elif len(image_or_images.get_shape()) == 3:
-    return [output[0, :, :, :], None]
-  else:
-    return [output, None]
+      images=grad,
+      transforms=transforms,
+      output_shape=array_ops.shape(image_or_images)[1:3],
+      interpolation=interpolation)
+  return [output, None, None]
 
 
 def bipartite_match(distance_mat,