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

Merge pull request #5388 from vrv/master

Browse Source 
Revert IDCT jpeg change and nvcc flag change
This commit is contained in:
Vijay Vasudevan 2016-11-03 15:48:41 -07:00 committed by GitHub
commit a507438361
4 changed files with 6 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tensorflow/core/lib/jpeg/jpeg_mem.cc
View File

@ -139,11 +139,8 @@ uint8* UncompressLow(const void* srcdata, FewerArgsForCompiler* argball) {
cinfo.do_fancy_upsampling = boolean(flags.fancy_upscaling); cinfo.do_fancy_upsampling = boolean(flags.fancy_upscaling);
cinfo.scale_num = 1; cinfo.scale_num = 1;
cinfo.scale_denom = ratio; cinfo.scale_denom = ratio;
// Activating this has a quality/speed trade-off implication. // Activating this has a quality/speed trade-off implication:
// However, most JPEGs are already compressed, and so the faster IDCT // cinfo.dct_method = JDCT_IFAST;
// should have no effect on training. The fast setting speeds up training on the
// GPU, e.g. by about 30% for smaller networks such as AlexNet.
cinfo.dct_method = JDCT_IFAST;
jpeg_start_decompress(&cinfo); jpeg_start_decompress(&cinfo);

2
tensorflow/core/lib/jpeg/jpeg_mem_unittest.cc
View File

@ -165,7 +165,7 @@ TEST(JpegMemTest, Jpeg2) {
// Compare the two images // Compare the two images
const int totalerr = ComputeSumAbsoluteDifference( const int totalerr = ComputeSumAbsoluteDifference(
imgdata1.get(), refdata1.get(), in_w, in_h, stride1, stride1); imgdata1.get(), refdata1.get(), in_w, in_h, stride1, stride1);
CHECK_LE(totalerr, 120000); CHECK_LE(totalerr, 85000);
} }
// check the second image too. Should be bitwise identical to the first. // check the second image too. Should be bitwise identical to the first.

4
tensorflow/python/ops/image_ops_test.py
View File

@ -1803,10 +1803,10 @@ class JpegTest(test_util.TensorFlowTestCase):
jpeg0, image0, image1, image2 = sess.run([jpeg0, image0, image1, image2]) jpeg0, image0, image1, image2 = sess.run([jpeg0, image0, image1, image2])
# The decoded-encoded image should be similar to the input # The decoded-encoded image should be similar to the input
self.assertLess(self.averageError(image0, image1), 0.7) self.assertLess(self.averageError(image0, image1), 0.6)
# We should be very close to a fixpoint # We should be very close to a fixpoint
self.assertLess(self.averageError(image1, image2), 0.6) self.assertLess(self.averageError(image1, image2), 0.02)
# Smooth ramps compress well (input size is 153600) # Smooth ramps compress well (input size is 153600)
self.assertGreaterEqual(len(jpeg0), 5000) self.assertGreaterEqual(len(jpeg0), 5000)

2
tensorflow/tensorflow.bzl
View File

@ -416,7 +416,7 @@ def _cuda_copts():
"@local_config_cuda//cuda:using_nvcc": ( "@local_config_cuda//cuda:using_nvcc": (
common_cuda_opts + common_cuda_opts +
[ [
"-nvcc_options=expt-relaxed-constexpr", "-nvcc_options=relaxed-constexpr",
"-nvcc_options=ftz=true", "-nvcc_options=ftz=true",
] ]
), ),