- 15 8月, 2017 33 次提交
-
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165274865
-
由 A. Unique TensorFlower 提交于
tf.name_scope, tf.device, tf.control_dependencies EAGER friendly. PiperOrigin-RevId: 165269572
-
由 Yao Zhang 提交于
Avoid creating new nodes and updating the input names of output nodes, if a folded node has only 1 output; instead, simply replace the existing folded node with a const node. This is to speed up the constant folding. PiperOrigin-RevId: 165261642
-
由 Brennan Saeta 提交于
Most CPU-based input pipeline image operations operate in the tensor layout NHWC. While this is reasonable for CPUs, may devices operate most efficiently using alternate layouts. (e.g. NVidia GPUs opt for NCHW). This change allows us to perform the layout reshapes on the CPU ahead of feeding them into the device. This change allows us to adopt non-batch-major image layouts while supporting automatic sharding. This is useful both for image inputs, as well as RNNs which sometimes would like to be time-major ordered. PiperOrigin-RevId: 165257772
-
由 Francois Chollet 提交于
It is largely a tiny refactor. One addition to the public API: add method `count_params` to base layer class (which allows us to use the `model.summary()` method with models built with core layers). PiperOrigin-RevId: 165255776
-
由 A. Unique TensorFlower 提交于
[tf.contrib.data] Add warning in dataset.map if user only specifies output_buffer_size but not num_threads. PiperOrigin-RevId: 165255129
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165253738
-
由 Brennan Saeta 提交于
Previously, the split op did not have any placement hints associated with it, which would sometimes result in very suboptimal placement. This change adds placement hints to avoid unnecessary network transfers. PiperOrigin-RevId: 165250631
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165249531
-
由 Toby Boyd 提交于
PiperOrigin-RevId: 165247204
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165245157
-
由 Yangzihao Wang 提交于
PiperOrigin-RevId: 165243700
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165242767
-
由 Alexandre Passos 提交于
PiperOrigin-RevId: 165242437
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165241905
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165239782
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165239093
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165236032
-
由 A. Unique TensorFlower 提交于
165233164 by andrewharp: internal change -- 165227502 by saeta: Add dataset.shard() for parallel (distributed) training. When running across multiple shards (i.e. multiple hosts), in order to avoid reading the same data redundantly and potentially overfitting during early training steps, it is important to split a dataset into non-overlapping subsets. The `.shard()` operator makes this very easy to do. -- 165227077 by A. Unique TensorFlower: Maybe export after each round of evaluation in `continuous_train_and_eval`. -- 165226561 by A. Unique TensorFlower: Adding an option to also update the tensorflow package name to tf-nightly for the nightly releases. -- 165224405 by tanmingxing: Add while_loop op benchmark for further performance study. -- 165223476 by andrewharp: Android demo: correct SSD detector url and handle model archives stored in sub-directories during Gradle build. Resolves #12265 Also remove deprecated mobile multibox model from Gradle build to bring parity with Bazel build. -- PiperOrigin-RevId: 165233164
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165220185
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165217993
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165214339
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165211238
-
由 Derek Murray 提交于
This ensures that the various polymorphic datasets will support all of TensorFlow's primitive types. In particular, it adds support for tf.uint16 in `Dataset.from_tensor_slices()`, `Dataset.batch()`, and `Dataset.dense_to_sparse_batch()`; and quantized types in `Dataset.from_sparse_tensor_slices()` and `Dataset.padded_batch(). Note that we do not use TF_CALL_ALL_TYPES, because that macro does not include the quantized types. PiperOrigin-RevId: 165209405
-
由 Shanqing Cai 提交于
PiperOrigin-RevId: 165207803
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165207133
-
由 Alexandre Passos 提交于
Temporary hack to make most composite ops work. PiperOrigin-RevId: 165205218
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165200062
-
由 Alexandre Passos 提交于
PiperOrigin-RevId: 165199520
-
由 Alexandre Passos 提交于
PiperOrigin-RevId: 165195680
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165195190
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165195082
-
由 A. Unique TensorFlower 提交于
1. There is a bug in in deriving the formulas for batch norm gradient calculation. This commit corrects the formula and updates the implementation. 2. Learning from mistake, this commit uses gradient_checker to test batch_norm_grad in a more generic way -- this method calculates the "real" gradients by evaluating the graph twice, and then compare it with the gradient. RELNOTES: Update formula for fused batch norm in xla PiperOrigin-RevId: 165190486
-
- 14 8月, 2017 1 次提交
-
-
由 A. Unique TensorFlower 提交于
to reduce memory consumption. On the wire, this should be fully compatible unless values > uint32max were actually ever used, which is unlikely. PiperOrigin-RevId: 165171764
-
- 13 8月, 2017 2 次提交
-
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165078268
-
由 A. Unique TensorFlower 提交于
wrapped in the generated code. PiperOrigin-RevId: 165078237
-
- 12 8月, 2017 4 次提交
-
-
由 Shanqing Cai 提交于
PiperOrigin-RevId: 165053647
-
由 Benoit Steiner 提交于
PiperOrigin-RevId: 165049643
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 165049580
-
由 Ali Yahya 提交于
PiperOrigin-RevId: 165049398
-