From 1c1cc028a46bb27a2535ecc5adeccbd47307abc8 Mon Sep 17 00:00:00 2001
From: Hao Jin <hjjn.amzn@gmail.com>
Date: Fri, 9 Aug 2019 21:37:25 +0000
Subject: [PATCH] add coverage for npx.relu and npx.sigmoid

---
 tests/python/unittest/test_numpy_op.py | 74 ++++++++++++++++++++++++++
 1 file changed, 74 insertions(+)

diff --git a/tests/python/unittest/test_numpy_op.py b/tests/python/unittest/test_numpy_op.py
index 818e0948dc65..574dd3b79d4a 100644
--- a/tests/python/unittest/test_numpy_op.py
+++ b/tests/python/unittest/test_numpy_op.py
@@ -224,6 +224,80 @@ def hybrid_forward(self, F, a, *args, **kwargs):
             check_unary_func(func, ref_grad, shape, low, high)
 
 
+@with_seed()
+@use_np
+def test_npx_relu():
+    def np_relu(x):
+        return _np.maximum(x, 0.0)
+    def np_relu_grad(x):
+        return 1.0 * (x > 0.0)
+
+    class TestReLU(HybridBlock):
+        def __init__(self):
+            super(TestReLU, self).__init__()
+
+        def hybrid_forward(self, F, a):
+            return F.npx.relu(a)
+
+    shapes = [(), (2, 3, 4), (2, 0, 3), (1, 0, 0)]
+    for hybridize in [True, False]:
+        for shape in shapes:
+            test_relu = TestReLU()
+            if hybridize:
+                test_relu.hybridize()
+            x = rand_ndarray(shape).as_np_ndarray()
+            x.attach_grad()
+            np_out = np_relu(x.asnumpy())
+            with mx.autograd.record():
+                mx_out = test_relu(x)
+            assert mx_out.shape == np_out.shape
+            assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+            mx_out.backward()
+            np_backward = np_relu_grad(x.asnumpy())
+            assert_almost_equal(x.grad.asnumpy(), np_backward, rtol=1e-3, atol=1e-5)
+
+            mx_out = npx.relu(x)
+            np_out = np_relu(x.asnumpy())
+            assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+
+
+@with_seed()
+@use_np
+def test_npx_sigmoid():
+    def np_sigmoid(x):
+        return _np.divide(1.0, (1.0 + _np.exp(-x)))
+    def np_sigmoid_grad(ya):
+        return ya * (1 - ya)
+
+    class TestSigmoid(HybridBlock):
+        def __init__(self):
+            super(TestSigmoid, self).__init__()
+
+        def hybrid_forward(self, F, a):
+            return F.npx.sigmoid(a)
+
+    shapes = [(), (2, 3, 4), (2, 0, 3), (1, 0, 0)]
+    for hybridize in [True, False]:
+        for shape in shapes:
+            test_sigmoid = TestSigmoid()
+            if hybridize:
+                test_sigmoid.hybridize()
+            x = rand_ndarray(shape).as_np_ndarray()
+            x.attach_grad()
+            np_out = np_sigmoid(x.asnumpy())
+            with mx.autograd.record():
+                mx_out = test_sigmoid(x)
+            assert mx_out.shape == np_out.shape
+            assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+            mx_out.backward()
+            np_backward = np_sigmoid_grad(np_out)
+            assert_almost_equal(x.grad.asnumpy(), np_backward, rtol=1e-3, atol=1e-5)
+
+            mx_out = npx.sigmoid(x)
+            np_out = np_sigmoid(x.asnumpy())
+            assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
+
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()