experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Addressed comments.

Browse Source
This commit is contained in:
mdfaijul 2020-10-08 17:15:18 -07:00
parent f6490af3f1
commit c53f7a7b1b
2 changed files with 31 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/core/util/mkl_util.h
View File

@ -894,7 +894,7 @@ inline void SetDummyMklDnnShapeOutput(OpKernelContext* context,
} }
// If the input tensor has ref count as 1, it is forwarded to the desired // If the input tensor has ref count as 1, it is forwarded to the desired
// output port and the function reutrns true. In that case, it also allocates // output port and the function returns true. In that case, it also allocates
// the serialized MklDnnShape object. Otherwise, the function returns false. // the serialized MklDnnShape object. Otherwise, the function returns false.
inline bool ForwardMklTensorInToOutWithMklShape(OpKernelContext* context, inline bool ForwardMklTensorInToOutWithMklShape(OpKernelContext* context,
int idx_in, int idx_out, int idx_in, int idx_out,
@ -921,9 +921,8 @@ inline bool ForwardMklTensorInToOutWithMklShape(OpKernelContext* context,
if (is_forwarded || always_forward) { if (is_forwarded || always_forward) {
AllocateOutputSetMklShape(context, idx_out, mkl_shape); AllocateOutputSetMklShape(context, idx_out, mkl_shape);
return true; return true;
} else {
return false;
} }
return false;
} }
// Forward the MKL shape ONLY (used in elementwise and other ops where // Forward the MKL shape ONLY (used in elementwise and other ops where

50
tensorflow/python/kernel_tests/conv_ops_test.py
View File

@ -3270,9 +3270,7 @@ def GetInceptionBackFilterTest(input_size, filter_size, output_size, strides,
class FusedConv2DTest(test.TestCase): class FusedConv2DTest(test.TestCase):
def _CreateNumpyTensor(self, shape): def _CreateNumpyTensor(self, shape):
total_size = 1 total_size = np.prod(shape)
for s in shape:
total_size *= s
return np.arange(1, total_size + 1, dtype=np.float32).reshape(shape) return np.arange(1, total_size + 1, dtype=np.float32).reshape(shape)
def _CreateConv2D(self, input_values, filters, def _CreateConv2D(self, input_values, filters,
@ -3283,7 +3281,9 @@ class FusedConv2DTest(test.TestCase):
strides=strides, strides=strides,
padding=padding) padding=padding)
@test_util.deprecated_graph_mode_only # Tests tensor forwarding of a fused Conv2D+BiasAdd+Add op when the input to
# Add has refcount 1.
@test_util.run_in_graph_and_eager_modes(use_gpu=False)
def testAddWithRefCountOne(self): def testAddWithRefCountOne(self):
expected_output = [ expected_output = [
113377, 125570, 77305, 86738, 19433, 22226, 60681, 113377, 125570, 77305, 86738, 19433, 22226, 60681,
@ -3297,12 +3297,12 @@ class FusedConv2DTest(test.TestCase):
filter_in = self._CreateNumpyTensor(filter_in_sizes) filter_in = self._CreateNumpyTensor(filter_in_sizes)
bias_in = self._CreateNumpyTensor(bias_in_sizes) bias_in = self._CreateNumpyTensor(bias_in_sizes)
# To get different weights for filter # To get different weights for filter
ofs = 1 offset = 1
conv1 = self._CreateConv2D(x, filter_in) conv1 = self._CreateConv2D(x, filter_in)
conv2 = self._CreateConv2D(conv1, filter_in + ofs) conv2 = self._CreateConv2D(conv1, filter_in + offset)
conv = self._CreateConv2D(conv1, filter_in - ofs) conv = self._CreateConv2D(conv1, filter_in - offset)
bias_add = nn_ops.bias_add(conv, bias_in) bias_add = nn_ops.bias_add(conv, bias_in)
add = math_ops.add_n([bias_add, conv2]) add = math_ops.add_n([bias_add, conv2])
@ -3310,7 +3310,9 @@ class FusedConv2DTest(test.TestCase):
np.rint(expected_output), np.rint(expected_output),
self.evaluate(add).reshape(-1)) self.evaluate(add).reshape(-1))
@test_util.deprecated_graph_mode_only # Tests tensor forwarding of a fused Conv2D+BiasAdd+Add op when the input to
# Add has a total refcount of 2, and Add is its last consumer.
@test_util.run_in_graph_and_eager_modes(use_gpu=False)
def testAddWithRefCountTwoAndRunAddLast(self): def testAddWithRefCountTwoAndRunAddLast(self):
expected_output = [ expected_output = [
1.907175e+06, 2.253505e+06, 7.809210e+05, 9.537180e+05, 1.907175e+06, 2.253505e+06, 7.809210e+05, 9.537180e+05,
@ -3326,12 +3328,12 @@ class FusedConv2DTest(test.TestCase):
filter_in = self._CreateNumpyTensor(filter_in_sizes) filter_in = self._CreateNumpyTensor(filter_in_sizes)
bias_in = self._CreateNumpyTensor(bias_in_sizes) bias_in = self._CreateNumpyTensor(bias_in_sizes)
# To get different weights for filter # To get different weights for filter
ofs = 1 offset = 1
conv1 = self._CreateConv2D(x, filter_in) conv1 = self._CreateConv2D(x, filter_in)
conv2 = self._CreateConv2D(conv1, filter_in + ofs) conv2 = self._CreateConv2D(conv1, filter_in + offset)
conv = self._CreateConv2D(conv2, filter_in - ofs) conv = self._CreateConv2D(conv2, filter_in - offset)
bias_add = nn_ops.bias_add(conv, bias_in) bias_add = nn_ops.bias_add(conv, bias_in)
add = math_ops.add_n([bias_add, conv1]) add = math_ops.add_n([bias_add, conv1])
@ -3339,7 +3341,9 @@ class FusedConv2DTest(test.TestCase):
np.rint(expected_output), np.rint(expected_output),
self.evaluate(add).reshape(-1)) self.evaluate(add).reshape(-1))
@test_util.deprecated_graph_mode_only # Tests tensor forwarding of a fused Conv2D+BiasAdd+Add op when the input to
# Add has refcount 2 and Add (in the fused Conv2D op) is its first consumer.
@test_util.run_in_graph_and_eager_modes(use_gpu=False)
def testAddWithRefCountTwoAndRunAddFirst(self): def testAddWithRefCountTwoAndRunAddFirst(self):
expected_output = [ expected_output = [
176161, 194450, 120673, 134822, 30545, 34734, 96041, 176161, 194450, 120673, 134822, 30545, 34734, 96041,
@ -3353,12 +3357,12 @@ class FusedConv2DTest(test.TestCase):
filter_in = self._CreateNumpyTensor(filter_in_sizes) filter_in = self._CreateNumpyTensor(filter_in_sizes)
bias_in = self._CreateNumpyTensor(bias_in_sizes) bias_in = self._CreateNumpyTensor(bias_in_sizes)
# To get different weights for filter # To get different weights for filter
ofs = 1 offset = 1
conv1 = self._CreateConv2D(x, filter_in) conv1 = self._CreateConv2D(x, filter_in)
conv2 = self._CreateConv2D(conv1, filter_in + ofs) conv2 = self._CreateConv2D(conv1, filter_in + offset)
conv = self._CreateConv2D(conv1, filter_in - ofs) conv = self._CreateConv2D(conv1, filter_in - offset)
bias_add = nn_ops.bias_add(conv, bias_in) bias_add = nn_ops.bias_add(conv, bias_in)
add = math_ops.add_n([bias_add, conv2]) add = math_ops.add_n([bias_add, conv2])
@ -3369,7 +3373,9 @@ class FusedConv2DTest(test.TestCase):
np.rint(expected_output), np.rint(expected_output),
self.evaluate(output).reshape(-1)) self.evaluate(output).reshape(-1))
@test_util.deprecated_graph_mode_only # Tests tensor forwarding of a fused Conv2D+BiasAdd+Add op when the input to
# Add has refcount 2, and there is no dependency between its two consumers.
@test_util.run_in_graph_and_eager_modes(use_gpu=False)
def testAddWithRefCountTwoAndNoDependence(self): def testAddWithRefCountTwoAndNoDependence(self):
expected_output = [ expected_output = [
176161, 194450, 120673, 134822, 30545, 34734, 96041, 176161, 194450, 120673, 134822, 30545, 34734, 96041,
@ -3383,12 +3389,12 @@ class FusedConv2DTest(test.TestCase):
filter_in = self._CreateNumpyTensor(filter_in_sizes) filter_in = self._CreateNumpyTensor(filter_in_sizes)
bias_in = self._CreateNumpyTensor(bias_in_sizes) bias_in = self._CreateNumpyTensor(bias_in_sizes)
# To get different weights for filter # To get different weights for filter
ofs = 1 offset = 1
conv1 = self._CreateConv2D(x, filter_in) conv1 = self._CreateConv2D(x, filter_in)
conv2 = self._CreateConv2D(conv1, filter_in + ofs) conv2 = self._CreateConv2D(conv1, filter_in + offset)
conv = self._CreateConv2D(conv1, filter_in - ofs) conv = self._CreateConv2D(conv1, filter_in - offset)
bias_add = nn_ops.bias_add(conv, bias_in) bias_add = nn_ops.bias_add(conv, bias_in)
add = math_ops.add_n([bias_add, conv2]) add = math_ops.add_n([bias_add, conv2])
@ -3400,8 +3406,10 @@ class FusedConv2DTest(test.TestCase):
np.rint(expected_output), np.rint(expected_output),
self.evaluate(output).reshape(-1)) self.evaluate(output).reshape(-1))
# Tests tensor forwarding of a fused Conv2D+BiasAdd+Add op when the input to
@test_util.deprecated_graph_mode_only # Add is the same as the input to the fused Conv2D op and needs a tensor
# buffer.
@test_util.run_in_graph_and_eager_modes(use_gpu=False)
def testAddWithSameSrcAndAddTensorBuffer(self): def testAddWithSameSrcAndAddTensorBuffer(self):
expected_output = [ expected_output = [
57157, 63298, 39249, 44026, 9971, 11402, 31193, 36306, 57157, 63298, 39249, 44026, 9971, 11402, 31193, 36306,