- 01 11月, 2016 34 次提交
-
-
由 A. Unique TensorFlower 提交于
Split code to compute control flow structure of a graph out of graph_partition.cc and into its own module (control_flow.{cc,h}). Change: 137756517
-
由 A. Unique TensorFlower 提交于
Change: 137756007
-
由 A. Unique TensorFlower 提交于
std::unordered_map and std::unordered_set. Add default template argument of std::hash<Key> to gtl::FlatMap and gtl::FlatSet to better match std::unordered_{map,set} Improves performance on an RPC-intensive benchmark by ~0.4% Change: 137754417
-
由 A. Unique TensorFlower 提交于
and marking Nudge as a device function. Change: 137753029
-
由 A. Unique TensorFlower 提交于
Change: 137750783
-
由 A. Unique TensorFlower 提交于
Change: 137749798
-
由 A. Unique TensorFlower 提交于
Each worker thread in the thread pool writes to its designated entry in the std::vector. The entry must be of a size that the CPU can write to without stepping on the adjacent entries. Moreover, std::vector<bool> is implemented using bitwise operations. The assignment operator sets or clears a bit in a word, which requires both a read and a write and is therefore not atomic. Change: 137749423
-
由 Yuefeng Zhou 提交于
Change: 137749204
-
由 A. Unique TensorFlower 提交于
Change: 137747341
-
由 A. Unique TensorFlower 提交于
. fix typo in call . re-export the functions so that they be accessed from outside code . do not create a new graph for the dummy placement variable. The new graph doesn't have the device setter applied, so the variable was not actually placed on the PS. Change: 137745823
-
由 A. Unique TensorFlower 提交于
Change: 137744670
-
由 A. Unique TensorFlower 提交于
Change: 137743958
-
由 Dan Smilkov 提交于
Rename fetchMetadata to fetchSpriteAndMetadata, and metadataInfo to spriteAndMetadataInfo. The sprite information is independent of the metadata file, so the projector should fetch the image regardless of whether the embeddings have labels or not. Also preserve the background of the sprite image in the scatter plot. Change: 137743425
-
由 Eugene Brevdo 提交于
Change: 137741689
-
由 A. Unique TensorFlower 提交于
Change: 137740850
-
由 A. Unique TensorFlower 提交于
Change: 137739075
-
由 A. Unique TensorFlower 提交于
Change: 137737291
-
由 Zongheng Yang 提交于
V2 is turned on: so test for the .index files. Example failure: https://ci.tensorflow.org/job/tensorflow-pull-requests-cpu/2379/console Change: 137734944
-
由 Charles Nicholson 提交于
bug where alpha was incorrectly disabled. Change: 137734683
-
由 A. Unique TensorFlower 提交于
Change: 137733513
-
由 Yuefeng Zhou 提交于
Change: 137731250
-
由 A. Unique TensorFlower 提交于
Does not implement these resources yet as this change is big enough already. Change: 137731142
-
由 Illia Polosukhin 提交于
Fixing sparse_ops_test string comparison failure. Fixing embedding_lookup_unique to support >2d params as per description. Change: 137730923
-
由 A. Unique TensorFlower 提交于
Change: 137730576
-
由 Dan Mané 提交于
Since we now use node names instead of tags, and don't require argument uniqueness, some logic for checking tag uniqueness was removed. The constructed graphs will be identical, except that the summaries may get different node names. Change: 137727808
-
由 A. Unique TensorFlower 提交于
Turning off the parallelized IsDirectory() call on iOS platform, due to problems with more than a few threads. Change: 137727687
-
由 A. Unique TensorFlower 提交于
Change: 137727599
-
由 Andrew Selle 提交于
- tf.where now can optionally take 3 arguments to create a ternary select behavior. tf.select will be eventually deprecated - tf.setdiff1d added in preparation to deprecate tf.listdiff - tf.pad now accepts lower or upper case constants (consatnts are now case insensitive) - added tf.multiply, tf.subtract, and tf.negative to match NumPy names - tf.mul, tf.sub, tf.neg will be deprecated. Change: 137726562
-
由 A. Unique TensorFlower 提交于
Change: 137721344
-
由 Zongheng Yang 提交于
V2 delivers (1) up to 3x peak memory reduction during restore, (2) much lower restore latency. These have proven to be critical for production use cases. Change: 137715407
-
由 A. Unique TensorFlower 提交于
Change callers of tf.image.per_image_whitening() to use tf.image.per_image_standardization(). Once these changes are submitted, per_image_whitening() can be removed. Change: 137714408
-
由 A. Unique TensorFlower 提交于
Change: 137709541
-
由 A. Unique TensorFlower 提交于
Instead of squeeze, which fails on single dimension shape tensors, use reshape to [-1] for shape tensor flattening. Change: 137708871
-
由 Charles Nicholson 提交于
glEnable(GL_ALPHA_TEST)). Disable premultiplied alpha for now, we don't have any translucent texture reads yet. Replace dynamic branching + discard in sprite fragment shader's analytical circle rendering, do a simple lerp instead. Simplify picking fragment shader as well, return the opacity in the alpha channel instead of discarding. Change: 137707296
-
- 31 10月, 2016 6 次提交
-
-
由 A. Unique TensorFlower 提交于
Change: 137705107
-
由 A. Unique TensorFlower 提交于
Change: 137701824
-
由 A. Unique TensorFlower 提交于
Change: 137701036
-
由 A. Unique TensorFlower 提交于
Change: 137699654
-
由 A. Unique TensorFlower 提交于
Change: 137695685
-
由 A. Unique TensorFlower 提交于
Change: 137692532
-