[incubator-mxnet] branch master updated: [BUGFIX] [Numpy] MXNet fp16 initialization bug #19118 (#19270)

[sx...@apache.org]

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

sxjscience 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 c2d56dc  [BUGFIX] [Numpy] MXNet fp16 initialization bug #19118 (#19270)
c2d56dc is described below

commit c2d56dcde0c4256c91635670d49d90448865ab25
Author: Anshu Trivedi <an...@gmail.com>
AuthorDate: Wed Oct 7 23:47:01 2020 +0530

    [BUGFIX] [Numpy] MXNet fp16 initialization bug #19118 (#19270)
    
    * fix: MXNet fp16 initialization bug #19118
    
    * add fix for a couple more initializers, and test
    
    Co-authored-by: Sheng Zha <zh...@amazon.com>
---
 python/mxnet/initializer.py         |  8 ++++----
 tests/python/unittest/test_smoke.py | 29 +++++++++++++++++++++++++++--
 2 files changed, 31 insertions(+), 6 deletions(-)

diff --git a/python/mxnet/initializer.py b/python/mxnet/initializer.py
index 28f8af1..60a1d3d 100644
--- a/python/mxnet/initializer.py
+++ b/python/mxnet/initializer.py
@@ -507,7 +507,7 @@ class Uniform(Initializer):
 
     def _init_weight(self, _, arr):
         uniform_fn = _mx_np.random.uniform if is_np_array() else random.uniform
-        uniform_fn(-self.scale, self.scale, arr.shape, out=arr)
+        uniform_fn(-self.scale, self.scale, arr.shape, dtype=arr.dtype, out=arr)
 
 @register
 class Normal(Initializer):
@@ -541,7 +541,7 @@ class Normal(Initializer):
 
     def _init_weight(self, _, arr):
         normal_fn = _mx_np.random.normal if is_np_array() else random.normal
-        normal_fn(0, self.sigma, arr.shape, out=arr)
+        normal_fn(0, self.sigma, arr.shape, dtype=arr.dtype, out=arr)
 
 @register
 class Orthogonal(Initializer):
@@ -641,10 +641,10 @@ class Xavier(Initializer):
         scale = np.sqrt(self.magnitude / factor)
         if self.rnd_type == "uniform":
             uniform_fn = _mx_np.random.uniform if is_np_array() else random.uniform
-            uniform_fn(-scale, scale, arr.shape, out=arr)
+            uniform_fn(-scale, scale, arr.shape, dtype=arr.dtype, out=arr)
         elif self.rnd_type == "gaussian":
             normal_fn = _mx_np.random.normal if is_np_array() else random.normal
-            normal_fn(0, scale, arr.shape, out=arr)
+            normal_fn(0, scale, arr.shape, dtype=arr.dtype, out=arr)
         else:
             raise ValueError("Unknown random type")
 
diff --git a/tests/python/unittest/test_smoke.py b/tests/python/unittest/test_smoke.py
index c14310c..04f8679 100644
--- a/tests/python/unittest/test_smoke.py
+++ b/tests/python/unittest/test_smoke.py
@@ -15,8 +15,9 @@
 # specific language governing permissions and limitations
 # under the License.
 
-from mxnet import np, npx, use_np, autograd
+from mxnet import np, npx, use_np, autograd, initializer, gluon
 from common import setup_module, teardown_module, with_environment
+import pytest
 
 @use_np
 @with_environment('MXNET_ENGINE_TYPE', 'NaiveEngine')
@@ -66,4 +67,28 @@ def test_18934_empty_leaky_relu():
     autograd.mark_variables([arr], [arr_grad])
     with autograd.record():
         res = npx.leaky_relu(arr)
-    res.backward()
\ No newline at end of file
+    res.backward()
+
+@use_np
+@pytest.mark.parametrize('initializer',[
+    'zeros', 'ones', initializer.Constant(3),
+    initializer.Uniform(),
+    initializer.Normal(),
+    initializer.Orthogonal(),
+    initializer.Orthogonal(rand_type='normal'),
+    initializer.Xavier(),
+    initializer.Xavier(rnd_type='gaussian'),
+    initializer.MSRAPrelu(),
+    initializer.MSRAPrelu(factor_type='in'),
+    initializer.MSRAPrelu(factor_type='out'),
+    initializer.LSTMBias(),
+])
+@pytest.mark.parametrize('dtype', [
+    'float32', 'float64'
+])
+def test_19118(initializer, dtype):
+    net = gluon.nn.Dense(16, in_units=16)
+    net.cast(dtype)
+    net.initialize(initializer)
+    net.hybridize()
+    net(np.zeros((16, 16), dtype=dtype))