This is an automated email from the ASF dual-hosted git repository.
lausen 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 07bfe55 parametrize test_deformable_convolution (#19600)
07bfe55 is described below
commit 07bfe55f8298563cf1352d908fc94bda615ebe84
Author: Leonard Lausen <[email protected]>
AuthorDate: Mon Nov 30 13:47:40 2020 -0700
parametrize test_deformable_convolution (#19600)
Fixes #19563
---
tests/python/unittest/test_operator.py | 73 ++++++++++++++++++----------------
1 file changed, 38 insertions(+), 35 deletions(-)
diff --git a/tests/python/unittest/test_operator.py
b/tests/python/unittest/test_operator.py
index 1e3a944..8a18d5d 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -5671,41 +5671,44 @@ def test_psroipooling_with_type():
'psroipool_rois': 'null'},
arg_params=arg_params)
-def test_deformable_convolution():
- for num_batch in [1, 2]:
- for num_channel_data, num_deformable_group in itertools.product([4,
8], [1, 2]):
- for input_height, input_width in itertools.product([5, 6], [5, 6]):
- for dilate in [(1, 1), (2, 2)]:
- for grad_nodes in [['im_data'], ['offset_data'],
['weight']]:
- output_height = input_height
- output_width = input_width
- im_data = np.random.rand(num_batch, num_channel_data,
input_height, input_width)
- offset_data = \
- np.random.rand(num_batch, num_deformable_group * 3
* 3 * 2, output_height, output_width)\
- * 0.8 + 0.1
-
- weight = np.random.normal(0, 0.001, (num_channel_data,
num_channel_data, 3, 3))
- bias = np.zeros(num_channel_data)
-
- im_data_var =
mx.symbol.Variable(name="im_data").as_np_ndarray()
- offset_data_var =
mx.symbol.Variable(name="offset_data").as_np_ndarray()
- weight_var =
mx.symbol.Variable(name="weight").as_np_ndarray()
- bias_var =
mx.symbol.Variable(name="bias").as_np_ndarray()
- op = mx.sym.npx.deformable_convolution(name='test_op',
data=im_data_var,
-
offset=offset_data_var,
-
weight=weight_var, bias=bias_var,
-
num_filter=num_channel_data, pad=dilate,
- kernel=(3, 3),
stride=(1, 1), dilate=dilate,
-
num_deformable_group=num_deformable_group)
- if grad_nodes[0] == 'offset_data':
- # wider tolerance needed for coordinate
differential
- rtol, atol = 1.0, 1e-2
- else:
- rtol, atol = 0.05, 1e-3
- # By now we only have gpu implementation
- if default_context().device_type == 'gpu':
- check_numeric_gradient(op, [im_data, offset_data,
weight, bias], rtol=rtol, atol=atol,
- grad_nodes=grad_nodes,
ctx=mx.gpu(0), numeric_eps=1.0/64)
[email protected]('num_batch', [1, 2])
[email protected]('num_channel_data_deformable_group',
itertools.product([4, 8], [1, 2]))
[email protected]('input_height_width', itertools.product([5, 6], [5,
6]))
[email protected]('dilate', [(1, 1), (2, 2)])
[email protected]('grad_nodes', [['im_data'], ['offset_data'],
['weight']])
+def test_deformable_convolution(num_batch, num_channel_data_deformable_group,
input_height_width,
+ dilate, grad_nodes):
+ num_channel_data, num_deformable_group = num_channel_data_deformable_group
+ input_height, input_width = input_height_width
+ output_height = input_height
+ output_width = input_width
+ im_data = np.random.rand(num_batch, num_channel_data, input_height,
input_width)
+ offset_data = \
+ np.random.rand(num_batch, num_deformable_group * 3 * 3 * 2,
output_height, output_width)\
+ * 0.8 + 0.1
+
+ weight = np.random.normal(0, 0.001, (num_channel_data, num_channel_data,
3, 3))
+ bias = np.zeros(num_channel_data)
+
+ im_data_var = mx.symbol.Variable(name="im_data").as_np_ndarray()
+ offset_data_var = mx.symbol.Variable(name="offset_data").as_np_ndarray()
+ weight_var = mx.symbol.Variable(name="weight").as_np_ndarray()
+ bias_var = mx.symbol.Variable(name="bias").as_np_ndarray()
+ op = mx.sym.npx.deformable_convolution(name='test_op', data=im_data_var,
+ offset=offset_data_var,
+ weight=weight_var, bias=bias_var,
+ num_filter=num_channel_data,
pad=dilate,
+ kernel=(3, 3), stride=(1, 1),
dilate=dilate,
+
num_deformable_group=num_deformable_group)
+ if grad_nodes[0] == 'offset_data':
+ # wider tolerance needed for coordinate differential
+ rtol, atol = 1.0, 1e-2
+ else:
+ rtol, atol = 0.05, 1e-3
+ # By now we only have gpu implementation
+ if default_context().device_type == 'gpu':
+ check_numeric_gradient(op, [im_data, offset_data, weight, bias],
rtol=rtol, atol=atol,
+ grad_nodes=grad_nodes, ctx=mx.gpu(0),
numeric_eps=1.0/64)
def _validate_sample_location(input_rois, input_offset, spatial_scale,
pooled_w, pooled_h, sample_per_part, part_size, output_dim, num_classes,
trans_std, feat_h, feat_w):
