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

Enable half for resource scatter

Browse Source 
Call each type

Enable sparse adam test in eager mode

Use self.evaluate()

Update tests for adam

Enable half for scatter

Use assertAllCloseAccordingToType to pass float16 tests

imake linter happy

Run pylint

Relax tolerance for half
This commit is contained in:
WindQAQ 2020-07-09 23:05:09 -07:00
parent 7e279d6b0f
commit 5e28f0c5c8
5 changed files with 9 additions and 10 deletions

4
tensorflow/core/kernels/resource_variable_ops.cc
View File

@ -998,8 +998,8 @@ REGISTER_SCATTER_KERNEL(Variant, CPU, "ResourceScatterUpdate",
#define REGISTER_SCATTER_UPDATE_GPU(type) REGISTER_SCATTER_UPDATE(type, GPU);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_ARITHMETIC_GPU);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_MINMAX_GPU);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_SCATTER_ARITHMETIC_GPU);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_SCATTER_MINMAX_GPU);
REGISTER_KERNEL_BUILDER(Name("ResourceScatterUpdate")
.Device(DEVICE_GPU)

2
tensorflow/core/kernels/scatter_functor.cc
View File

@ -55,7 +55,7 @@ namespace functor {
DECLARE_GPU_SPECS_INDEX(T, int32); \
DECLARE_GPU_SPECS_INDEX(T, int64);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(DECLARE_GPU_SPECS);
TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPECS);
#undef DECLARE_GPU_SPECS
#undef DECLARE_GPU_SPECS_INDEX

4
tensorflow/core/kernels/scatter_functor_gpu.cu.cc
View File

@ -40,13 +40,11 @@ typedef Eigen::GpuDevice GPUDevice;
DEFINE_GPU_SPECS_INDEX(T, int32); \
DEFINE_GPU_SPECS_INDEX(T, int64);
DEFINE_GPU_SPECS(Eigen::half);
DEFINE_GPU_SPECS(float);
DEFINE_GPU_SPECS(double);
DEFINE_GPU_SPECS_OP(bool, int32, scatter_op::UpdateOp::ASSIGN);
DEFINE_GPU_SPECS_OP(bool, int64, scatter_op::UpdateOp::ASSIGN);
// TODO(b/27222123): The following fails to compile due to lack of support for
// fp16.
// TF_CALL_GPU_NUMBER_TYPES(DEFINE_GPU_SPECS);
#undef DEFINE_GPU_SPECS
#undef DEFINE_GPU_SPECS_INDEX

4
tensorflow/core/kernels/scatter_op.cc
View File

@ -286,9 +286,9 @@ TF_CALL_ALL_TYPES(REGISTER_SCATTER_UPDATE_CPU);
#define REGISTER_SCATTER_UPDATE_GPU(type) REGISTER_SCATTER_UPDATE(type, GPU);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_ARITHMETIC_GPU);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_SCATTER_ARITHMETIC_GPU);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_SCATTER_MINMAX_GPU);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_UPDATE_GPU);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_SCATTER_UPDATE_GPU);
#endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM

5
tensorflow/python/kernel_tests/scatter_ops_test.py
View File

@ -182,13 +182,14 @@ class ScatterTest(test.TestCase):
ref = variables.Variable(old)
self.evaluate(ref.initializer)
self.evaluate(tf_scatter(ref, indices, updates))
self.assertAllClose(self.evaluate(ref), new)
self.assertAllCloseAccordingToType(
self.evaluate(ref), new, half_rtol=5e-3, half_atol=5e-3)
def _VariableRankTests(self,
tf_scatter,
repeat_indices=False,
updates_are_scalar=False):
vtypes = [np.float32, np.float64]
vtypes = [np.float16, np.float32, np.float64]
if tf_scatter != state_ops.scatter_div:
vtypes.append(np.int32)