Fix Flaky Test Higher Order Grad (apache#17325)

* failing seed * failed seed arcsin * use random_uniform_arrays to generate test arrays * use random_uniform_arrays for arccos * fix: handle case when len(s) == 0 * change lower and upper bounds of randomly generated array * failure case seed With this seed, one of the elements of head_grad for the first gradient is zero and leads to failure in `arctanh` due to the way it is implemented. * change gradient implementation of _backward_arctanh * remove seed * fix higher order gradient implementation for multiple trig operators * relax tolerance for tanh * address comments * add dtype argument. * consistent function signature. * concise list comprehension. * fix implementation for log1p * update function signature to be compatible with older python * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI * retrigger CI
anirudh2290 · Feb 4, 2020 · 71a5c9e · 71a5c9e
1 parent de30aa5
commit 71a5c9e
Show file tree

Hide file tree

Showing 5 changed files with 57 additions and 25 deletions.
diff --git a/python/mxnet/test_utils.py b/python/mxnet/test_utils.py
@@ -102,6 +102,18 @@ def random_arrays(*shapes):
     return arrays
 
 
+def random_uniform_arrays(*shapes, **kwargs):
+    """Generate some random numpy arrays."""
+    low = kwargs.pop('low', 0.0)
+    high = kwargs.pop('high', 1.0)
+    dtype = kwargs.pop('dtype', default_dtype())
+    if len(kwargs) > 0:
+        raise TypeError('Got unexpected argument/s : ' + str(kwargs.keys()))
+    arrays = [np.random.uniform(low, high, size=s).astype(dtype)
+              for s in shapes]
+    return arrays
+
+
 def random_sample(population, k):
     """Return a k length list of the elements chosen from the population sequence."""
     assert 0 <= k <= len(population)

diff --git a/src/nnvm/node_op_util.h b/src/nnvm/node_op_util.h
@@ -80,6 +80,12 @@ class NodeOpGen {
                                                    dependent_node->attrs.name + "_negative",
                                                    {x}, nullptr, &dependent_node)};
     }
+
+    nnvm::NodeEntry ones_like(const nnvm::NodeEntry &x) {
+        return nnvm::NodeEntry{mxnet::op::MakeNode("ones_like",
+                                                   dependent_node->attrs.name + "_oneslike",
+                                                   {x}, nullptr, &dependent_node)};
+    }
 };
 
 }  // namespace util

diff --git a/src/operator/tensor/elemwise_unary_op_logexp.cc b/src/operator/tensor/elemwise_unary_op_logexp.cc
@@ -214,11 +214,13 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_log1p,
     auto dydx_mul_dldy = nnvm::NodeEntry{n};  // f'(x) * head_grads
     auto op = mxnet::util::NodeOpGen{n};
 
-    auto dydx = op.div(dydx_mul_dldy, dldy);
-
-    auto d2ydx2_mid = op.mul(dydx_mul_dldy, dydx_mul_dldy);
+    auto ones = op.ones_like(x);
+    auto dydx = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_log1p",
+                                                  n->attrs.name + "_backward_log1p",
+                                                  {ones, x}, nullptr, &n)};
+    auto d2ydx2_mid = op.mul(dydx, dydx);
     auto d2ydx2_neg_mid = op.negative(d2ydx2_mid);
-    auto d2ydx2 = op.div(d2ydx2_neg_mid, dldy);
+    auto d2ydx2 = op.mul(d2ydx2_neg_mid, dldy);
 
     std::vector<nnvm::NodeEntry> ret;
 

diff --git a/src/operator/tensor/elemwise_unary_op_trig.cc b/src/operator/tensor/elemwise_unary_op_trig.cc
@@ -203,7 +203,10 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arcsin,
       auto dydx_mul_grad_x = nnvm::NodeEntry{n};
       auto op = mxnet::util::NodeOpGen{n};
 
-      auto x_grad = op.div(dydx_mul_grad_x, dydx);
+      auto ones = op.ones_like(x);
+      auto x_grad = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_arcsin",
+                                                   n->attrs.name + "_backward_arcsin",
+                                                   {ones, x}, nullptr, &n)};
       auto x_grad_square = op.square(x_grad);
       auto x_grad_square_mul_x = op.mul(x_grad_square, x);
       auto x_grad_grad = op.mul(dydx_mul_grad_x, x_grad_square_mul_x);
@@ -246,7 +249,10 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arccos,
       auto dydx_mul_grad_x = nnvm::NodeEntry{n};
       auto op = mxnet::util::NodeOpGen{n};
 
-      auto x_grad = op.div(dydx_mul_grad_x, dydx);
+      auto ones = op.ones_like(x);
+      auto x_grad = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_arccos",
+                                                   n->attrs.name + "_backward_arccos",
+                                                   {ones, x}, nullptr, &n)};
       auto x_grad_square = op.square(x_grad);
       auto x_grad_square_mul_x = op.mul(x_grad_square, x);
       auto x_grad_grad = op.mul(dydx_mul_grad_x, x_grad_square_mul_x);
@@ -295,7 +301,10 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arctan,
       auto dldy_mul_dydx = nnvm::NodeEntry{n};
       auto op = mxnet::util::NodeOpGen{n};
 
-      auto x_grad = op.div(dldy_mul_dydx, dldy);
+      auto ones = op.ones_like(x);
+      auto x_grad = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_arctan",
+                                                   n->attrs.name + "_backward_arctan",
+                                                   {ones, x}, nullptr, &n)};
       auto x_grad_square = op.square(x_grad);
       auto x_grad_square_mul_x = op.mul(x_grad_square, x);
       auto x_grad_square_mul_2_x = op.mul(-2.0, x_grad_square_mul_x);
@@ -501,7 +510,10 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arcsinh,
       auto dydx_mul_grad_x = nnvm::NodeEntry{n};
       auto op = mxnet::util::NodeOpGen{n};
 
-      auto grad_x = op.div(dydx_mul_grad_x, dydx);
+      auto ones = op.ones_like(x);
+      auto grad_x = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_arcsinh",
+                                                   n->attrs.name + "_backward_arcsinh",
+                                                   {ones, x}, nullptr, &n)};
       auto grad_x_square = op.square(grad_x);
       auto grad_x_square_mul_x = op.mul(grad_x_square, x);
       auto grad_grad_x = op.mul(dydx_mul_grad_x, grad_x_square_mul_x);
@@ -539,7 +551,10 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arccosh,
       auto dydx_mul_grad_x = nnvm::NodeEntry{n};
       auto op = mxnet::util::NodeOpGen{n};
 
-      auto grad_x = op.div(dydx_mul_grad_x, dydx);
+      auto ones = op.ones_like(x);
+      auto grad_x = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_arccosh",
+                                                   n->attrs.name + "_backward_arccosh",
+                                                   {ones, x}, nullptr, &n)};
       auto grad_x_square = op.square(grad_x);
       auto grad_x_square_mul_x = op.mul(grad_x_square, x);
       auto grad_grad_x = op.mul(dydx_mul_grad_x, grad_x_square_mul_x);
@@ -583,7 +598,10 @@ MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arctanh,
       auto dldy_mul_dydx = nnvm::NodeEntry{n};
       auto op = mxnet::util::NodeOpGen{n};
 
-      auto x_grad = op.div(dldy_mul_dydx, dldy);
+      auto ones = op.ones_like(x);
+      auto x_grad = nnvm::NodeEntry{mxnet::op::MakeNode("_backward_arctanh",
+                                                   n->attrs.name + "_backward_arctanh",
+                                                   {ones, x}, nullptr, &n)};
       auto x_grad_square = op.square(x_grad);
       auto x_grad_square_mul_x = op.mul(x_grad_square, x);
       auto x_grad_square_mul_2_x = op.mul(2.0, x_grad_square_mul_x);

diff --git a/tests/python/unittest/test_higher_order_grad.py b/tests/python/unittest/test_higher_order_grad.py
@@ -27,7 +27,8 @@
 from common import with_seed
 import mxnet
 from mxnet import nd, autograd, gluon
-from mxnet.test_utils import assert_almost_equal, random_arrays, rand_shape_nd, same
+from mxnet.test_utils import (
+    assert_almost_equal, random_arrays, random_uniform_arrays, rand_shape_nd, same)
 
 
 @with_seed()
@@ -131,7 +132,7 @@ def grad_grad_op(x):
         array = random_arrays(shape)
         check_nth_order_unary(array, tanh, grad_op, 1, rtol=1e-6, atol=1e-6)
         check_second_order_unary(
-            array, tanh, grad_grad_op, rtol=1e-6, atol=1e-6)
+            array, tanh, grad_grad_op, rtol=1e-6, atol=1e-5)
 
 
 @with_seed()
@@ -144,12 +145,8 @@ def grad_grad_op(x):
 
     for dim in range(1, 5):
         shape = rand_shape_nd(dim)
-        array = random_arrays(shape)
-        # Hack: Decrease std_dev to make
-        # sure all elements
-        # are in range -1 to 1
-        # i.e. Domain of arcsin
-        array *= 0.2
+        # Domain of arcsin is [-1, 1]
+        array = random_uniform_arrays(shape, low=-0.99, high=0.99)[0]
         check_second_order_unary(array, arcsin, grad_grad_op)
 
 
@@ -163,12 +160,8 @@ def grad_grad_op(x):
 
     for dim in range(1, 5):
         shape = rand_shape_nd(dim)
-        array = random_arrays(shape)
-        # Hack: Decrease std_dev to make
-        # sure all elements
-        # are in range -1 to 1
-        # i.e. Domain of arccos
-        array *= 0.2
+        # Domain of arccos is [-1, 1]
+        array = random_uniform_arrays(shape, low=-0.99, high=0.99)[0]
         check_second_order_unary(array, arccos, grad_grad_op)
 
 
@@ -233,7 +226,8 @@ def grad_grad_op(x):
 
     for dim in range(1, 5):
         shape = rand_shape_nd(dim)
-        array = random_arrays(shape)
+        # Domain of arctanh is (-1, 1)
+        array = random_uniform_arrays(shape, low=-0.99, high=0.99)[0]
         check_second_order_unary(array, arctanh, grad_grad_op)