[incubator-mxnet] branch master updated: Pad Operator Type Support (#12035)

[ns...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

nswamy pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git


The following commit(s) were added to refs/heads/master by this push:
     new b03227d  Pad Operator Type Support (#12035)
b03227d is described below

commit b03227d9dbd97f3cdd4c0ba8496e05797847a631
Author: Sebastian Bodenstein <se...@gmail.com>
AuthorDate: Mon Aug 20 16:40:38 2018 +0200

    Pad Operator Type Support (#12035)
    
    * fix no data type inference for pad
    
    * add support for int types
    
    * add tests for all types
    
    * fix gpu type switch
    
    * remove integer support
    
    * fix python op test style issues
    
    * fix type bug in python tests
---
 src/operator/pad-inl.h                 | 11 +++++++++++
 tests/python/unittest/test_operator.py | 24 ++++++++++++++----------
 2 files changed, 25 insertions(+), 10 deletions(-)

diff --git a/src/operator/pad-inl.h b/src/operator/pad-inl.h
index 520cd12..0b43e2d 100644
--- a/src/operator/pad-inl.h
+++ b/src/operator/pad-inl.h
@@ -189,6 +189,17 @@ class PadProp : public OperatorProperty {
     return param_.__DICT__();
   }
 
+  bool InferType(std::vector<int> *in_type,
+                 std::vector<int> *out_type,
+                 std::vector<int> *aux_type) const override {
+    int dtype = (*in_type)[0];
+    type_assign(&dtype, (*out_type)[0]);
+
+    TYPE_ASSIGN_CHECK(*in_type, 0, dtype);
+    TYPE_ASSIGN_CHECK(*out_type, 0, dtype);
+    return dtype != -1;
+  }
+
   bool InferShape(std::vector<TShape> *in_shape, std::vector<TShape> *out_shape,
                   std::vector<TShape> *aux_shape) const override {
     using namespace mshadow;
diff --git a/tests/python/unittest/test_operator.py b/tests/python/unittest/test_operator.py
index 04617de..125666b 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -3008,16 +3008,16 @@ def test_roipooling():
                            numeric_eps=1e-4, rtol=1e-1, atol=1E-4)
 
 
-def check_pad_with_shape(shape, xpu, pad_width, mode):
+def check_pad_with_shape(shape, xpu, pad_width, mode, dtype="float64"):
     # bind with label
-    X = mx.symbol.Variable('X')
+    X = mx.symbol.Variable('X', dtype=dtype)
     Y = mx.symbol.Pad(data=X, mode=mode, pad_width=pad_width)
-    x = mx.random.uniform(-1, 1, shape, ctx=mx.cpu()).copyto(xpu)
+    x = mx.random.uniform(-1, 1, shape, ctx=mx.cpu(), dtype=dtype).copyto(xpu)
     # numpy result
     pad_grouped = list(zip(*[iter(list(pad_width))] * 2))
     np_out = np.pad(x.asnumpy(), pad_grouped, mode)
     # mxnet result
-    grad = mx.nd.empty(shape, ctx = xpu)
+    grad = mx.nd.empty(shape, ctx = xpu, dtype=dtype)
     exec1 = Y.bind(xpu, args = [x], args_grad = {'X': grad})
     exec1.forward(is_train=True)
     out = exec1.outputs[0].asnumpy()
@@ -3029,16 +3029,20 @@ def check_pad_with_shape(shape, xpu, pad_width, mode):
 
 @with_seed()
 def test_pad():
+    ctx = default_context()
     shape1 = (2, 3, 3, 5)
     pad1 = (0, 0, 0, 0, 1, 2, 3, 4)
     shape2 = (2, 3, 3, 5, 4)
     pad2 = (0, 0, 0, 0, 1, 2, 3, 4, 3, 1)
-    check_pad_with_shape(shape1, default_context(), pad1, 'constant')
-    check_pad_with_shape(shape1, default_context(), pad1, 'edge')
-    check_pad_with_shape(shape2, default_context(), pad2, 'constant')
-    check_pad_with_shape(shape2, default_context(), pad2, 'edge')
-    check_pad_with_shape(shape1, default_context(), pad1, 'reflect')
-    check_pad_with_shape(shape2, default_context(), pad2, 'reflect')
+    # note: this op doesn't support ints yet. Add tests when supported
+    dtypes = ["float16", "float32", "float64"]
+    for dtype in dtypes:
+        check_pad_with_shape(shape1, ctx, pad1, 'constant', dtype)
+        check_pad_with_shape(shape1, ctx, pad1, 'edge', dtype)
+        check_pad_with_shape(shape2, ctx, pad2, 'constant', dtype)
+        check_pad_with_shape(shape2, ctx, pad2, 'edge', dtype)
+        check_pad_with_shape(shape1, ctx, pad1, 'reflect', dtype)
+        check_pad_with_shape(shape2, ctx, pad2, 'reflect', dtype)
 
 
 def np_instance_norm(data, weight, bias, eps):