- 22 8月, 2019 34 次提交
-
-
由 Igor Saprykin 提交于
PiperOrigin-RevId: 264693076
-
由 Igor Saprykin 提交于
PiperOrigin-RevId: 264693064
-
由 George Karpenkov 提交于
The trait specifies that the `MemRefOf` has to have a static shape. PiperOrigin-RevId: 264692758
-
由 A. Unique TensorFlower 提交于
Move linear algebra ops 'lu_solve', 'lu_inverse', and 'lu_reconstruct' from TensorFlow Probability to TensorFlow core. Slightly refactor linear_operator_util.py to avoid a circular dependence with linalg_impl.py. PiperOrigin-RevId: 264692711
-
由 Ayush Dubey 提交于
PiperOrigin-RevId: 264692343
-
由 Loren Maggiore 提交于
PiperOrigin-RevId: 264692056
-
由 Feng Liu 提交于
execution The imported GraphDef might have extra attributes which are not part of the op registration but added by other tools. The existence of these attributes might cause error when they are added to the eager op for constant folding. So we want to ignore them when the op is converted to an NodeDef during the constant folding. PiperOrigin-RevId: 264691406
-
由 Jacques Pienaar 提交于
InitMlir performs both initializations while also allowing passing some flags to both sides (e.g., tool {flags to tf} -- {flags to llvm}). PiperOrigin-RevId: 264690725
-
由 Alex Stark 提交于
PiperOrigin-RevId: 264690662
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 264688956
-
由 Eugene Zhulenev 提交于
Not all fusion patterns in Remapper supported for NHWC and NCHW. Assign data layout before looking for fusable patterns. PiperOrigin-RevId: 264688604
-
由 Gaurav Jain 提交于
PiperOrigin-RevId: 264688574
-
由 Guangda Lai 提交于
added. This can help to assign the TRTEngineOp to the desired GPU if that isn't the first one. Also enhance the tests to make it more robust. PiperOrigin-RevId: 264686781
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264686678
-
由 Rohan Jain 提交于
Changing the Dataset options that allowed whitelisted ops to be replicated to a generic allow_stateful_ops flag rather than explicitly passing in a list of stateful ops to be whitelisted. In this case, we log a warning listing out which stateful ops we encountered along the way. PiperOrigin-RevId: 264685786
-
由 Sundeep Gottipati 提交于
PiperOrigin-RevId: 264683921
-
由 Alex Stark 提交于
This amounts to disabling Ruy paths for this cpuid instruction results lack selected features. PiperOrigin-RevId: 264683681
-
由 Taylor Robie 提交于
Refactor the keras TensorLikeDataAdapter (numpy array, EagerTensor, etc) to use tf.shuffle rather than np.shuffle. This allows us to use more of tf.data's pipelining machinery which both improves multi-epoch performance and decreases memory consumption. PiperOrigin-RevId: 264681581
-
由 Brian Zhao 提交于
PiperOrigin-RevId: 264681321
-
由 River Riddle 提交于
PiperOrigin-RevId: 264672975
-
由 Benoit Jacob 提交于
Don't round the allocator's storage size to the next power of two. This is typically a huge buffer. We're going to reach a steady state where we have only a few such buffers and they won't get frequently reallocated, anyway. PiperOrigin-RevId: 264669851
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264669653
-
由 Yifei Feng 提交于
PiperOrigin-RevId: 264668323
-
由 Benoit Jacob 提交于
in size_util. Part of it was AllocateFast not checking if ptr_ is null before using it (null deref with offset, so didn't look like a null deref). Part of it was using round_up_pot with a large size_t value that got implicitly casted to int as round_up_pot took an int argument. This showed how it's safer to just templatize the helpers in size_util.h, make them accept either int32 or int64 (guarded in floor_log2, which is the only of these functions that cares). I just changed allocator to use only signed types (std::size_t --> std::ptrdiff_t) because I didn't want to deal with the extra complexity of dealing with both signed and unsigned in size_util. PiperOrigin-RevId: 264668215
-
由 River Riddle 提交于
The ModuleState is only used for printing aliases, which is only done when printing the top-level module. PiperOrigin-RevId: 264664138
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264659402
-
由 Rohan Jain 提交于
1. Fixing variable_scope to not pass in partition_info argument if its in V2 mode. This ensures compatibility with V2 initializers 2. Adding a tracking_name argument to add_variable which allows for passing in a different name for tracking purposes without affecting the variable name. PiperOrigin-RevId: 264658202
-
由 Nick Kreeger 提交于
This op implementation closely matches the existing TFLite Op - but expects more details about the type of execution strategy and memory layout specific to Micro. For example, output shape and scratch tensor allocations should be known during model conversion. The implemented unit tests expect these values to be identified before execution. PiperOrigin-RevId: 264645863
-
由 Gunhan Gulsoy 提交于
PiperOrigin-RevId: 264641416
-
由 River Riddle 提交于
This will allow iterating the values of a non-opaque ElementsAttr, with all of the types currently supported by DenseElementsAttr. This should help reduce the amount of specialization on DenseElementsAttr. PiperOrigin-RevId: 264637293
-
由 A. Unique TensorFlower 提交于
Updates TF Lite benchmark tool documentation to explicitly mention the ability to specify an NNAPI accelerator PiperOrigin-RevId: 264635461
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264635106
-
由 River Riddle 提交于
This will allow for adding more hooks for controlling parser behavior without bloating Dialect in the common case. This cl also adds iteration support to the DialectInterfaceCollection. PiperOrigin-RevId: 264627846
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264626141
-
- 21 8月, 2019 6 次提交
-
-
由 Lei Zhang 提交于
PiperOrigin-RevId: 264612014
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264609586
-
由 Alex Stark 提交于
PiperOrigin-RevId: 264605397
-
由 Lasse Espeholt 提交于
PiperOrigin-RevId: 264605227
-
由 Alex Stark 提交于
PiperOrigin-RevId: 264602338
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 264600694
-