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

Merge pull request #13264 from caisq/branch_169770126

Browse Source 
Branch 169770126
This commit is contained in:
Shanqing Cai 2017-09-23 22:01:48 -04:00 committed by GitHub
commit d2d42ee8b3
199 changed files with 4847 additions and 2014 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/BUILD
View File

@ -356,7 +356,6 @@ filegroup(
"//tensorflow/contrib/data:all_files",
"//tensorflow/contrib/data/python/kernel_tests:all_files",
"//tensorflow/contrib/data/python/ops:all_files",
"//tensorflow/contrib/data/python/util:all_files",
"//tensorflow/contrib/decision_trees/proto:all_files",
"//tensorflow/contrib/distributions:all_files",
"//tensorflow/contrib/eager/python:all_files",
@ -475,6 +474,9 @@ filegroup(
"//tensorflow/java/src/main/java/org/tensorflow/examples:all_files",
"//tensorflow/java/src/main/native:all_files",
"//tensorflow/python:all_files",
"//tensorflow/python/data:all_files",
"//tensorflow/python/data/ops:all_files",
"//tensorflow/python/data/util:all_files",
"//tensorflow/python/debug:all_files",
"//tensorflow/python/eager:all_files",
"//tensorflow/python/estimator:all_files",

6
tensorflow/cc/framework/scope.h
View File

@ -107,13 +107,13 @@ class Scope {
static Scope NewRootScope();
/// Return a new scope. Ops created with this scope will have
/// <name>/<child_scope_name> as the prefix. The actual name will be unique
/// `name/child_scope_name` as the prefix. The actual name will be unique
/// in the current scope. All other properties are inherited from the current
/// scope. If child_scope_name is empty, the '/' is elided.
/// scope. If `child_scope_name` is empty, the `/` is elided.
Scope NewSubScope(const string& child_scope_name) const;
/// Return a new scope. All ops created within the returned scope will have
/// names of the form <name>/<op_name>[_<suffix].
/// names of the form `name/op_name[_suffix]`.
Scope WithOpName(const string& op_name) const;
/// Return a new scope. All ops created within the returned scope will have as

6
tensorflow/cc/gradients/nn_grad.cc
View File

@ -50,7 +50,6 @@ REGISTER_GRADIENT_OP("Softmax", SoftmaxGrad);
Status LogSoftmaxGrad(const Scope& scope, const Operation& op,
const std::vector<Output>& grad_inputs,
std::vector<Output>* grad_outputs) {
auto softmax = Exp(scope, op.output(0));
auto sum = Sum(scope, grad_inputs[0], {1}, Sum::KeepDims(true));
auto mul = Mul(scope, sum, softmax);
@ -130,8 +129,7 @@ Status Conv2DGrad(const Scope& scope, const Operation& op,
TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "data_format", &data_format));
TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "padding", &padding));
TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "strides", &strides));
TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "use_cudnn_on_gpu",
&use_cudnn_on_gpu));
TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "use_cudnn_on_gpu", &use_cudnn_on_gpu));
Conv2DBackpropInput::Attrs input_attrs;
input_attrs.DataFormat(data_format);
input_attrs.UseCudnnOnGpu(use_cudnn_on_gpu);
@ -198,8 +196,6 @@ Status MaxPoolGradV2Helper(const Scope& scope, const Operation& op,
}
REGISTER_GRADIENT_OP("MaxPoolV2", MaxPoolGradV2Helper);
} // anonymous namespace
} // namespace ops
} // namespace tensorflow

54
tensorflow/cc/gradients/nn_grad_test.cc
View File

@ -36,7 +36,7 @@ class NNGradTest : public ::testing::Test {
float max_error;
TF_ASSERT_OK((ComputeGradientError<float, float, float>(
scope_, {x}, {x_shape}, {y}, {y_shape}, &max_error)));
EXPECT_LT(max_error, 2.2e-4);
EXPECT_LT(max_error, 1e-3);
}
void RunTest(const Output& x, const Tensor& x_init_value, const Output& y,
@ -44,7 +44,7 @@ class NNGradTest : public ::testing::Test {
float max_error;
TF_ASSERT_OK((ComputeGradientError<float, float, float>(
scope_, x, x_init_value, y, y_shape, &max_error)));
EXPECT_LT(max_error, 2.2e-4);
EXPECT_LT(max_error, 1e-3);
}
void RunTest(const OutputList& xs, const std::vector<TensorShape>& x_shapes,
@ -53,7 +53,25 @@ class NNGradTest : public ::testing::Test {
float max_error;
TF_ASSERT_OK((ComputeGradientError<float, float, float>(
scope_, xs, x_shapes, ys, y_shapes, &max_error)));
EXPECT_LT(max_error, 2.2e-4);
EXPECT_LT(max_error, 1e-3);
}
// Sets tensor with random values, ensuring that the max value is largest by
// a reasonable amount.
// This is an issue for MaxPool and MaxPoolV2, in which perturbations by the
// numeric gradient computation in the gradient checker can change the max
// value if values are too close together.
template <typename T>
void SetRandomValuesWithBumpedMax(Tensor* tensor) {
auto tensor_flat = tensor->flat<T>();
tensor_flat.setRandom();
int32 max_index = 0;
for (size_t i = 1; i < tensor->NumElements(); i++) {
if (tensor_flat(i) > tensor_flat(max_index)) {
max_index = i;
}
}
tensor_flat(max_index) += 1e-2;
}
Scope scope_;
@ -148,22 +166,30 @@ TEST_F(NNGradTest, Conv2DGrad) {
}
TEST_F(NNGradTest, MaxPoolGradHelper) {
TensorShape shape({1, 2, 2, 1});
auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
TensorShape x_shape({1, 2, 2, 1});
TensorShape y_shape({1, 1, 1, 1});
auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));
// Setup window and strides so that we only do one MaxPool.
const std::vector<int> ksize{1, 2, 2, 1};
const std::vector<int> strides{1, 1, 1, 1};
auto y = MaxPool(scope_, x, ksize, strides, "SAME");
RunTest(x, shape, y, shape);
const std::vector<int> strides{1, 2, 2, 1};
auto y = MaxPool(scope_, x, ksize, strides, "VALID");
Tensor x_init_value = Tensor(DT_FLOAT, x_shape);
SetRandomValuesWithBumpedMax<float>(&x_init_value);
RunTest(x, x_init_value, y, y_shape);
}
TEST_F(NNGradTest, MaxPoolGradV2Helper) {
TensorShape shape({1, 2, 2, 1});
auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
TensorShape x_shape({1, 2, 2, 1});
TensorShape y_shape({1, 1, 1, 1});
auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));
// Setup window and strides so that we only do one MaxPool.
Tensor ksize = test::AsTensor<int>({1, 2, 2, 1}, {4});
Tensor strides = test::AsTensor<int>({1, 1, 1, 1}, {4});
auto y = MaxPoolV2(scope_, x, ksize, strides, "SAME");
RunTest(x, shape, y, shape);
Tensor strides = test::AsTensor<int>({1, 2, 2, 1}, {4});
auto y = MaxPoolV2(scope_, x, ksize, strides, "VALID");
Tensor x_init_value = Tensor(DT_FLOAT, x_shape);
SetRandomValuesWithBumpedMax<float>(&x_init_value);
RunTest(x, x_init_value, y, y_shape);
}
} // namespace
} // namespace tensorflow

14
tensorflow/compiler/jit/BUILD
View File

@ -45,7 +45,7 @@ cc_library(
visibility = ["//visibility:public"],
deps = [
":jit_compilation_passes",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla/kernels:xla_ops",
"//tensorflow/compiler/xla/service:cpu_plugin",
],
@ -57,7 +57,7 @@ cc_library(
visibility = ["//visibility:public"],
deps = if_cuda([
":jit_compilation_passes",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla/kernels:xla_ops",
"//tensorflow/compiler/xla/service:gpu_plugin",
]),
@ -71,7 +71,7 @@ cc_library(
deps = [
":jit_compilation_passes",
":xla_device",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla:xla_compiler",
"//tensorflow/compiler/tf2xla/kernels:xla_ops",
"//tensorflow/compiler/xla/service:cpu_plugin", # buildcleaner: keep
@ -88,7 +88,7 @@ cc_library(
deps = [
":jit_compilation_passes",
":xla_device",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla:xla_compiler",
"//tensorflow/compiler/tf2xla/kernels:xla_ops",
"//tensorflow/compiler/xla/service:gpu_plugin", # buildcleaner: keep
@ -103,7 +103,7 @@ cc_library(
srcs = ["xla_interpreter_device.cc"],
deps = [
":xla_device",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla:xla_compiler",
],
alwayslink = True,
@ -213,7 +213,7 @@ cc_library(
deps = [
":common",
":compilation_passes",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla:const_analysis",
"//tensorflow/compiler/tf2xla:xla_compiler",
"//tensorflow/core:core_cpu_internal",
@ -297,7 +297,7 @@ tf_cc_test(
"//tensorflow/cc:cc_ops_internal",
"//tensorflow/cc:function_ops",
"//tensorflow/cc:ops",
"//tensorflow/compiler/jit/kernels:xla_local_launch_op",
"//tensorflow/compiler/jit/kernels:xla_launch_op",
"//tensorflow/compiler/tf2xla:xla_compiler",
"//tensorflow/compiler/tf2xla/kernels:xla_ops",
"//tensorflow/core:core_cpu",

2
tensorflow/compiler/jit/create_xla_launch_op.cc
View File

@ -14,7 +14,7 @@ limitations under the License.
==============================================================================*/
#include "tensorflow/compiler/jit/defs.h"
#include "tensorflow/compiler/jit/kernels/xla_local_launch_op.h"
#include "tensorflow/compiler/jit/kernels/xla_launch_op.h"
#include "tensorflow/compiler/jit/mark_for_compilation_pass.h"
#include "tensorflow/compiler/tf2xla/const_analysis.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"

6
tensorflow/compiler/jit/kernels/BUILD
View File

@ -7,9 +7,9 @@ package(
)
cc_library(
name = "xla_local_launch_op",
srcs = ["xla_local_launch_op.cc"],
hdrs = ["xla_local_launch_op.h"],
name = "xla_launch_op",
srcs = ["xla_launch_op.cc"],
hdrs = ["xla_launch_op.h"],
deps = [
"//tensorflow/compiler/jit:common",
"//tensorflow/compiler/jit:xla_compilation_cache",

4
tensorflow/compiler/jit/kernels/xla_local_launch_op.cc → tensorflow/compiler/jit/kernels/xla_launch_op.cc
View File

@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/compiler/jit/kernels/xla_local_launch_op.h"
#include "tensorflow/compiler/jit/kernels/xla_launch_op.h"
#include "tensorflow/compiler/jit/defs.h"
#include "tensorflow/compiler/jit/xla_device.h"
@ -194,7 +194,7 @@ std::vector<OptionalTensor> SnapshotResourceVariables(OpKernelContext* ctx,
Var* variable = nullptr;
ResourceHandle handle = HandleFromInput(ctx, first_variable + i);
if (LookupResource(ctx, handle, &variable).ok()) {
mutex_lock lock(*variable->mu());
tf_shared_lock lock(*variable->mu());
snapshot[i].name = handle.name();
snapshot[i].present = true;
snapshot[i].value = *variable->tensor();

0
tensorflow/compiler/jit/kernels/xla_local_launch_op.h → tensorflow/compiler/jit/kernels/xla_launch_op.h
View File

2
tensorflow/compiler/jit/xla_cpu_device.cc
View File

@ -16,7 +16,7 @@ limitations under the License.
// Registers the XLA_CPU device, which is an XlaDevice instantiation that runs
// operators using XLA via the XLA "Host" (CPU) backend.
#include "tensorflow/compiler/jit/kernels/xla_local_launch_op.h"
#include "tensorflow/compiler/jit/kernels/xla_launch_op.h"
#include "tensorflow/compiler/jit/xla_device.h"
#include "tensorflow/compiler/jit/xla_device_ops.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"

2
tensorflow/compiler/jit/xla_gpu_device.cc
View File

@ -16,7 +16,7 @@ limitations under the License.
// Registers the XLA_GPU device, which is an XlaDevice instantiation that runs
// operators using XLA via the XLA "CUDA" (GPU) backend.
#include "tensorflow/compiler/jit/kernels/xla_local_launch_op.h"
#include "tensorflow/compiler/jit/kernels/xla_launch_op.h"
#include "tensorflow/compiler/jit/xla_device.h"
#include "tensorflow/compiler/jit/xla_device_ops.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"

2
tensorflow/compiler/jit/xla_interpreter_device.cc
View File

@ -15,7 +15,7 @@ limitations under the License.
// Registers the XLA_INTERPRETER device which exposes the XLA Interpreter.
#include "tensorflow/compiler/jit/kernels/xla_local_launch_op.h"
#include "tensorflow/compiler/jit/kernels/xla_launch_op.h"
#include "tensorflow/compiler/jit/xla_device.h"
#include "tensorflow/compiler/jit/xla_device_ops.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"

16
tensorflow/compiler/xla/array2d.h
View File

@ -165,6 +165,22 @@ class Array2D {
return tensorflow::str_util::Join(pieces, "");
}
bool operator==(const Array2D<T>& other) const {
if (n1() != other.n1() || n2() != other.n2()) {
return false;
}
for (int64 i0 = 0; i0 < n1(); ++i0) {
for (int64 i1 = 0; i1 < n2(); ++i1) {
if ((*this)(i0, i1) != other(i0, i1)) {
return false;
}
}
}
return true;
}
bool operator!=(const Array2D<T>& other) const { return !(*this == other); }
private:
int64 n1_;
int64 n2_;

41
tensorflow/compiler/xla/array2d_test.cc
View File

@ -139,5 +139,46 @@ TEST(Array2dTest, Stringification) {
EXPECT_EQ(expected, arr->ToString());
}
TEST(Array2dTest, Equals) {
Array2D<int> arr0 = {{1, 2}, {3, 4}, {5, 6}};
Array2D<int> arr1 = {{1, 2}, {3, 4}, {5, 6}};
EXPECT_TRUE(arr0 == arr1);
EXPECT_FALSE(arr0 != arr1);
EXPECT_TRUE(arr1 == arr0);
EXPECT_FALSE(arr1 != arr0);
Array2D<int> arr2 = {{1, 2}, {3, 4}, {5, 6}, {7, 8}};
EXPECT_TRUE(arr0 != arr2);
EXPECT_FALSE(arr0 == arr2);
EXPECT_TRUE(arr2 != arr0);
EXPECT_FALSE(arr2 == arr0);
Array2D<int> arr3 = {{1, 2, 3}, {4, 5, 6}};
EXPECT_TRUE(arr0 != arr3);
EXPECT_FALSE(arr0 == arr3);
EXPECT_TRUE(arr3 != arr0);
EXPECT_FALSE(arr3 == arr0);
Array2D<int> arr4 = {{1, 2}, {3, 4}};
EXPECT_TRUE(arr0 != arr4);
EXPECT_FALSE(arr0 == arr4);
EXPECT_TRUE(arr4 != arr0);
EXPECT_FALSE(arr4 == arr0);
Array2D<int> arr5 = {{1, 2}, {13, 4}, {5, 6}};
EXPECT_TRUE(arr0 != arr5);
EXPECT_FALSE(arr0 == arr5);
EXPECT_TRUE(arr5 != arr0);
EXPECT_FALSE(arr5 == arr0);
Array2D<bool> bool_arr0 = {{false}, {true}};
Array2D<bool> bool_arr1 = {{false}, {true}};
EXPECT_TRUE(bool_arr0 == bool_arr1);
EXPECT_FALSE(bool_arr0 != bool_arr1);
Array2D<bool> bool_arr2 = {{false}, {false}};
EXPECT_FALSE(bool_arr0 == bool_arr2);
EXPECT_TRUE(bool_arr0 != bool_arr2);
}
} // namespace
} // namespace xla

15
tensorflow/compiler/xla/service/BUILD
View File

@ -503,6 +503,7 @@ cc_library(
"//tensorflow/compiler/xla:xla_data_proto",
"//tensorflow/compiler/xla/legacy_flags:debug_options_flags",
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
"//tensorflow/core:stream_executor_no_cuda",
],
)
@ -1966,6 +1967,20 @@ cc_library(
alwayslink = 1,
)
tf_cc_test(
name = "hlo_graph_dumper_test",
srcs = ["hlo_graph_dumper_test.cc"],
deps = [
":hlo",
":hlo_graph_dumper",
"//tensorflow/compiler/xla:test",
"//tensorflow/compiler/xla:xla_proto",
"//tensorflow/compiler/xla/tests:test_utils",
"//tensorflow/compiler/xla/tests:xla_internal_test_main", # fixdeps: keep
"//tensorflow/core:lib",
],
)
cc_library(
name = "transpose_folding",
srcs = ["transpose_folding.cc"],

7
tensorflow/compiler/xla/service/compile_only_service.cc
View File

@ -28,7 +28,9 @@ limitations under the License.
#include "tensorflow/compiler/xla/types.h"
#include "tensorflow/compiler/xla/util.h"
#include "tensorflow/core/lib/gtl/cleanup.h"
#include "tensorflow/core/lib/io/path.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/host_info.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/stream_executor_no_cuda.h"
@ -83,7 +85,10 @@ CompileOnlyService::CompileAheadOfTime(
"computation_", versioned_handle.handle.handle(), "__",
session_module->entry().name(), "__version_",
versioned_handle.version);
TF_RETURN_IF_ERROR(Executable::DumpToDirectory(directory_path, filename,
const string& per_host_path = tensorflow::io::JoinPath(
directory_path, tensorflow::port::Hostname());
TF_RETURN_IF_ERROR(Executable::DumpToDirectory(per_host_path, filename,
*session_module));
}

2
tensorflow/compiler/xla/service/cpu/cpu_compiler.cc
View File

@ -811,7 +811,7 @@ CpuCompiler::CompileAheadOfTime(std::vector<std::unique_ptr<HloModule>> modules,
/*is_entry_computation=*/true,
&module_sequence.at(computation)));
entry_function->setName(llvm_ir::AsStringRef(entry_point_name));
CHECK(entry_function->getName() == llvm_ir::AsStringRef(entry_point_name));
ModuleHook pre_optimization_ir_dump_hook;
ModuleHook post_optimization_ir_dump_hook;

100
tensorflow/compiler/xla/service/hlo_graph_dumper.cc
View File

@ -340,11 +340,8 @@ class HloDotDumper {
string Header();
string Footer();
// Maps HloComputations we should dump to their parent instruction in the
// outer computation.
std::unordered_map<const HloComputation*, const HloInstruction*>
SubcomputationsToDump();
bool ShouldShowSubcomputation(const HloComputation* subcomp);
bool ShouldShowFusionSubcomputation(const HloInstruction* instr);
string DumpSubcomputation(const HloComputation* subcomp,
const HloInstruction* parent_instr);
string DumpComputation(const HloComputation* comp);
@ -401,11 +398,6 @@ class HloDotDumper {
string HloDotDumper::Dump() {
string body;
for (const auto& kv : SubcomputationsToDump()) {
const HloComputation* subcomp = kv.first;
const HloInstruction* parent = kv.second;
StrAppend(&body, DumpSubcomputation(subcomp, parent));
}
StrAppend(&body, DumpComputation(computation_));
StrAppend(&body, DumpRootTag());
@ -525,33 +517,36 @@ stylesheet="
string HloDotDumper::Footer() { return StrCat(Join(edges_, "\n"), "\n}"); }
std::unordered_map<const HloComputation*, const HloInstruction*>
HloDotDumper::SubcomputationsToDump() {
// Dump the subcomputations of each instruction that's shown and doesn't have
// its operands omitted. If an instruction has just one subcomputation and
// it's trivial, omit it: We'll display that subcomputation inlined into the
// instruction's node when we draw it.
std::unordered_map<const HloComputation*, const HloInstruction*> to_dump;
for (const auto& instr : computation_->instructions()) {
if (!filter_.Show(instr.get()) ||
filter_.SomeOrAllOperandsOmitted(instr.get())) {
continue;
}
if (instr->opcode() == HloOpcode::kFusion) {
to_dump[instr->fused_instructions_computation()] = instr.get();
}
bool HloDotDumper::ShouldShowFusionSubcomputation(const HloInstruction* instr) {
CHECK_EQ(instr->opcode(), HloOpcode::kFusion);
return ShouldShowSubcomputation(instr->fused_instructions_computation());
}
for (const HloComputation* comp : instr->called_computations()) {
if (!MatchTrivialComputation(comp)) {
to_dump[comp] = instr.get();
}
bool HloDotDumper::ShouldShowSubcomputation(const HloComputation* subcomp) {
if (subcomp->IsFusionComputation()) {
const HloInstruction* fusion = subcomp->FusionInstruction();
if (!filter_.Show(fusion) || filter_.SomeOrAllOperandsOmitted(fusion)) {
return false;
}
}
return to_dump;
// Don't show trivial subcomputations on non-fusion nodes -- these are inlined
// into the graph.
if (!subcomp->IsFusionComputation() && MatchTrivialComputation(subcomp)) {
return false;
}
// Show the subcomputation if we're showing any of its members.
return std::any_of(computation_->instructions().begin(),
computation_->instructions().end(),
[&](const std::unique_ptr<HloInstruction>& instr) {
return filter_.Show(instr.get());
});
}
string HloDotDumper::DumpSubcomputation(const HloComputation* subcomp,
const HloInstruction* parent_instr) {
VLOG(2) << "Dumping subcomputation " << subcomp->name();
const char* computation_fmt = R"(subgraph %s {
%s
label = <%s>;
@ -593,20 +588,10 @@ tooltip = " ";
string comp_body = DumpComputation(subcomp);
if (parent_instr->opcode() == HloOpcode::kFusion) {
// Dump any nested fusion nodes.
for (const auto& subcomp_instr : subcomp->instructions()) {
if (subcomp_instr->opcode() == HloOpcode::kFusion) {
StrAppend(
&comp_body,
DumpSubcomputation(subcomp_instr->fused_instructions_computation(),
subcomp_instr.get()));
}
}
} else {
// Add an edge from the subcomputation to its parent node. If subcomp
// belongs to a fusion node, it's drawn in place of the fusion instruction,
// so there's no need to link those.
// Add an edge from the subcomputation to its parent node. If subcomp
// belongs to a fusion node, it's drawn in place of the fusion instruction,
// so there's no need to link those.
if (parent_instr->opcode() != HloOpcode::kFusion) {
VLOG(2) << "Edge: from " << subcomp->root_instruction()->name() << " to "
<< parent_instr->name() << " as " << next_edge_id_;
edge_ids_.insert(
@ -631,6 +616,14 @@ string HloDotDumper::DumpComputation(const HloComputation* comp) {
if (!filter_.Show(instr.get())) {
continue;
}
// Dump subcomputations within instr.
for (const HloComputation* subcomp : instr->called_computations()) {
if (ShouldShowSubcomputation(subcomp)) {
StrAppend(&g, DumpSubcomputation(subcomp, instr.get()));
}
}
StrAppend(&g, DumpInstruction(instr.get()));
}
return g;
@ -638,6 +631,14 @@ string HloDotDumper::DumpComputation(const HloComputation* comp) {
string HloDotDumper::DumpRootTag() {
HloInstruction* from = computation_->root_instruction();
// Fusion nodes are expanded inline, so if root is an expanded fusion node,
// walk up the graph until we find a node that isn't.
while (from->opcode() == HloOpcode::kFusion &&
ShouldShowFusionSubcomputation(from)) {
from = from->fused_expression_root();
}
auto from_id = InstructionId(from);
if (!filter_.Show(from)) {
@ -678,7 +679,7 @@ string HloDotDumper::DumpInstruction(const HloInstruction* instr) {
// Omit the fusion node if its subcomputation is drawn, since the
// subcomputation will be drawn inline.
if (instr->opcode() == HloOpcode::kFusion &&
filter_.ShowFusionSubcomputation(instr)) {
ShouldShowFusionSubcomputation(instr)) {
return "";
}
@ -937,7 +938,7 @@ string HloDotDumper::GetInstructionNodeExtraInfo(const HloInstruction* instr) {
// Show the shape and layout of the instruction, unless it's an inlined fusion
// node -- there the shape and layout is present in the output node.
if (instr->opcode() != HloOpcode::kFusion ||
!filter_.ShowFusionSubcomputation(instr)) {
!ShouldShowFusionSubcomputation(instr)) {
string instr_shape = ShapeUtil::HumanString(instr->shape());
// Show layout of non-tuple shapes with more than one dimension.
@ -982,7 +983,7 @@ void HloDotDumper::AddInstructionIncomingEdges(const HloInstruction* instr) {
// fusion node and the node's subcomputation is shown, we draw our edge
// starting at the fusion node's root instead of at the fusion node itself.
if (from->opcode() == HloOpcode::kFusion &&
filter_.ShowFusionSubcomputation(from)) {
ShouldShowFusionSubcomputation(from)) {
from = from->fused_expression_root();
}
if (!filter_.Show(from) || from->opcode() == HloOpcode::kConstant) {
@ -1147,6 +1148,11 @@ NodeFilter MakeNodeFilter(const HloInstruction* root, int64 radius) {
}
}
// Traverse into instr's nested computations.
for (const HloComputation* computation : instr->called_computations()) {
worklist.push_back({computation->root_instruction(), depth + 1});
}
// Traverse into instr's users, unless:
//
// - there are a ton of them, in which case they're probably not

122
tensorflow/compiler/xla/service/hlo_graph_dumper_test.cc Normal file
View File

@ -0,0 +1,122 @@
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/compiler/xla/service/hlo_graph_dumper.h"
#include "tensorflow/compiler/xla/service/hlo_computation.h"
#include "tensorflow/compiler/xla/service/hlo_instruction.h"
#include "tensorflow/compiler/xla/service/hlo_module.h"
#include "tensorflow/compiler/xla/service/hlo_opcode.h"
#include "tensorflow/compiler/xla/test.h"
#include "tensorflow/compiler/xla/tests/test_utils.h"
#include "tensorflow/compiler/xla/xla.pb.h"
#include "tensorflow/core/lib/strings/strcat.h"
namespace xla {
namespace {
using ::tensorflow::strings::StrCat;
using ::testing::HasSubstr;
string TestName() {
return ::testing::UnitTest::GetInstance()->current_test_info()->name();
}
class DotRenderer : public hlo_graph_dumper::GraphRendererInterface {
public:
string RenderGraph(const string& graph, GraphKind graph_kind,
const DebugOptions& debug_options) override {
return graph;
}
private:
string last_graph_;
};
XLA_REGISTER_GRAPH_RENDERER(DotRenderer, std::numeric_limits<int>::max());
TEST(HloGraphDumperTest, NestedFusion) {
HloComputation::Builder b("b");
// Build param0 + param1 + param2 + param3 + param4.
auto shape = ShapeUtil::MakeShape(F32, {10, 100});
std::vector<HloInstruction*> params;
for (int i = 0; i <= 4; ++i) {
params.push_back(b.AddInstruction(
HloInstruction::CreateParameter(i, shape, StrCat("param", i))));
}
std::vector<HloInstruction*> sums;
sums.push_back(b.AddInstruction(HloInstruction::CreateBinary(
shape, HloOpcode::kAdd, params[0], params[1])));
for (int i = 0; i <= 2; ++i) {
sums.push_back(b.AddInstruction(HloInstruction::CreateBinary(
shape, HloOpcode::kAdd, sums[i], params[i + 2])));
}
HloModule m(TestName());
m.AddEntryComputation(b.Build());
HloComputation* root_computation = m.entry_computation();
// Fuse into fusion(param0 + param1 + param2 + param3 + param4).
auto* outer_fusion = root_computation->CreateFusionInstruction(
{sums[3], sums[2], sums[1], sums[0]}, HloInstruction::FusionKind::kLoop);
// Fusing invalidates the pointers in sums -- the instructions are cloned when
// they're moved to the new computation. Get the updated pointers to sums.
std::vector<HloInstruction*> fused_sums;
for (auto* instr : outer_fusion->fused_instructions_computation()
->MakeInstructionPostOrder()) {
if (instr->opcode() == HloOpcode::kAdd) {
fused_sums.push_back(instr);
}
}
// Fuse into fusion(fusion(param0 + param1 + param2) + param3 + param4).
auto* inner_fusion =
outer_fusion->fused_instructions_computation()->CreateFusionInstruction(
{fused_sums[1], fused_sums[0]}, HloInstruction::FusionKind::kLoop);
// Generate the graph; all nodes should be present.
string graph = hlo_graph_dumper::DumpGraph(*root_computation, /*label=*/"",
DebugOptions());
for (const HloComputation* computation :
{root_computation, //
inner_fusion->fused_instructions_computation(),
outer_fusion->fused_instructions_computation()}) {
for (const std::unique_ptr<HloInstruction>& instruction :
computation->instructions()) {
EXPECT_THAT(graph, HasSubstr(instruction->name()));
}
}
// Dump a neighborhood around one of the inner sum nodes. We don't really
// care that the outer nodes are omitted -- whether they are or not is based
// fiddly heuristics -- but we do care that the node we asked for is printed.
const HloInstruction* inner_sum = nullptr;
for (const std::unique_ptr<HloInstruction>& instruction :
inner_fusion->fused_instructions_computation()->instructions()) {
if (instruction->opcode() == HloOpcode::kAdd) {
inner_sum = instruction.get();
break;
}
}
ASSERT_NE(inner_sum, nullptr);
EXPECT_THAT(
hlo_graph_dumper::DumpNeighborhoodAround(*inner_sum, /*radius=*/1),
HasSubstr(inner_sum->name()));
}
} // anonymous namespace
} // namespace xla

5
tensorflow/compiler/xla/service/service.cc
View File

@ -930,9 +930,10 @@ tensorflow::Status Service::TransferToClient(const TransferToClientRequest* arg,
}
Literal literal;
auto status = LiteralFromAllocation(allocation, *literal_shape, &literal);
TF_RETURN_IF_ERROR(
LiteralFromAllocation(allocation, *literal_shape, &literal));
*result->mutable_literal() = literal.ToProto();
return status;
return tensorflow::Status::OK();
}
tensorflow::Status Service::TransferToServer(const TransferToServerRequest* arg,

77
tensorflow/compiler/xla/tests/reduce_window_test.cc
View File

@ -32,6 +32,8 @@ limitations under the License.
#include "tensorflow/compiler/xla/tests/literal_test_util.h"
#include "tensorflow/compiler/xla/tests/test_macros.h"
#include "tensorflow/compiler/xla/xla_data.pb.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/lib/gtl/array_slice.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/platform/types.h"
@ -320,6 +322,81 @@ TEST_F(ReduceWindowTest, R4UnitWindow) {
ErrorSpec(1e-3, 1e-3));
}
XLA_TEST_F(HloTestBase, R6AddMultipleStrides) {
auto b = HloComputation::Builder(TestName());
std::vector<int64> input_dims(6, 8);
auto shape = ShapeUtil::MakeShape(F32, input_dims);
std::unique_ptr<Literal> arg_literal = Literal::CreateFromShape(shape);
auto generator = [&](tensorflow::gtl::ArraySlice<int64> indexes) -> float {
return 1.0f;
};
TF_EXPECT_OK(arg_literal->Populate<float>(generator));
auto input =
b.AddInstruction(HloInstruction::CreateConstant(std::move(arg_literal)));
auto init_value = b.AddInstruction(
HloInstruction::CreateConstant(Literal::CreateR0<float>(0.f)));
HloComputation::Builder add_computation("add");
Shape scalar_shape = ShapeUtil::MakeShape(F32, {});
auto param_lhs = add_computation.AddInstruction(
HloInstruction::CreateParameter(0, scalar_shape, "lhs"));
auto param_rhs = add_computation.AddInstruction(
HloInstruction::CreateParameter(1, scalar_shape, "rhs"));
add_computation.AddInstruction(HloInstruction::CreateBinary(
scalar_shape, HloOpcode::kAdd, param_lhs, param_rhs));
auto module = CreateNewModule();
auto add_func = module->AddEmbeddedComputation(add_computation.Build());
WindowDimension trivial_dim;
trivial_dim.set_size(1);
trivial_dim.set_stride(1);
trivial_dim.set_padding_low(0);
trivial_dim.set_padding_high(0);
trivial_dim.set_window_dilation(1);
trivial_dim.set_base_dilation(1);
WindowDimension active_dim;
active_dim.set_size(3);
active_dim.set_stride(1);
active_dim.set_padding_low(0);
active_dim.set_padding_high(0);
active_dim.set_window_dilation(1);
active_dim.set_base_dilation(1);
Window window;
*window.add_dimensions() = active_dim;
*window.add_dimensions() = trivial_dim;
*window.add_dimensions() = active_dim;
*window.add_dimensions() = active_dim;
*window.add_dimensions() = trivial_dim;
*window.add_dimensions() = trivial_dim;
// Non-monotonic output layout with minor dims trivial.
std::vector<int64> output_layout = {1, 5, 3, 2, 0, 4};
std::vector<int64> output_dims = {6, 8, 6, 6, 8, 8};
Shape result_shape =
ShapeUtil::MakeShapeWithLayout(F32, output_dims, output_layout);
b.AddInstruction(HloInstruction::CreateReduceWindow(
result_shape, input, init_value, window, add_func));
std::unique_ptr<Literal> expected = Literal::CreateFromShape(result_shape);
auto out_generator =
[&](tensorflow::gtl::ArraySlice<int64> indexes) -> float {
return 27.0f;
};
TF_EXPECT_OK(expected->Populate<float>(out_generator));
module->AddEntryComputation(b.Build());
auto actual = ExecuteAndTransfer(std::move(module), {});
LiteralTestUtil::ExpectNear(*actual, *expected, ErrorSpec(1e-3, 1e-3));
}
XLA_TEST_F(HloTestBase, R6Add) {
auto b = HloComputation::Builder(TestName());

3
tensorflow/contrib/boosted_trees/BUILD
View File

@ -222,9 +222,6 @@ py_test(
size = "small",
srcs = ["python/kernel_tests/split_handler_ops_test.py"],
srcs_version = "PY2AND3",
tags = [
"nomac", # b/63258195
],
deps = [
":split_handler_ops_py",
"//tensorflow/contrib/boosted_trees/proto:learner_proto_py",

3
tensorflow/contrib/boosted_trees/estimator_batch/BUILD
View File

@ -90,9 +90,6 @@ py_test(
size = "small",
srcs = ["custom_export_strategy_test.py"],
srcs_version = "PY2AND3",
tags = [
"nomac", # b/63258195
],
deps = [
":custom_export_strategy",
"//tensorflow/contrib/decision_trees/proto:generic_tree_model_extensions_py",

6
tensorflow/contrib/boosted_trees/lib/BUILD
View File

@ -281,9 +281,6 @@ py_test(
name = "categorical_split_handler_test",
srcs = ["learner/batch/categorical_split_handler_test.py"],
srcs_version = "PY2AND3",
tags = [
"nomac", # b/63258195
],
deps = [
":categorical_split_handler",
"//tensorflow/contrib/boosted_trees/proto:learner_proto_py",
@ -309,9 +306,6 @@ py_test(
name = "ordinal_split_handler_test",
srcs = ["learner/batch/ordinal_split_handler_test.py"],
srcs_version = "PY2AND3",
tags = [
"nomac", # b/63258195
],
deps = [
":ordinal_split_handler",
"//tensorflow/contrib/boosted_trees/proto:learner_proto_py",

5
tensorflow/contrib/cmake/tf_python.cmake
View File

@ -204,6 +204,9 @@ add_python_module("tensorflow/examples/tutorials")
add_python_module("tensorflow/examples/tutorials/mnist")
add_python_module("tensorflow/python")
add_python_module("tensorflow/python/client")
add_python_module("tensorflow/python/data")
add_python_module("tensorflow/python/data/ops")
add_python_module("tensorflow/python/data/util")
add_python_module("tensorflow/python/debug")
add_python_module("tensorflow/python/debug/cli")
add_python_module("tensorflow/python/debug/examples")
@ -237,6 +240,7 @@ add_python_module("tensorflow/python/keras/datasets/cifar100")
add_python_module("tensorflow/python/keras/datasets/imdb")
add_python_module("tensorflow/python/keras/datasets/mnist")
add_python_module("tensorflow/python/keras/datasets/reuters")
add_python_module("tensorflow/python/keras/estimator")
add_python_module("tensorflow/python/keras/initializers")
add_python_module("tensorflow/python/keras/layers")
add_python_module("tensorflow/python/keras/losses")
@ -333,7 +337,6 @@ add_python_module("tensorflow/contrib/data")
add_python_module("tensorflow/contrib/data/python")
add_python_module("tensorflow/contrib/data/python/kernel_tests")
add_python_module("tensorflow/contrib/data/python/ops")
add_python_module("tensorflow/contrib/data/python/util")
add_python_module("tensorflow/contrib/decision_trees")
add_python_module("tensorflow/contrib/decision_trees/proto")
add_python_module("tensorflow/contrib/deprecated")

1
tensorflow/contrib/data/BUILD
View File

@ -12,6 +12,7 @@ py_library(
"//tensorflow/contrib/data/python/ops:dataset_ops",
"//tensorflow/contrib/data/python/ops:sloppy_ops",
"//tensorflow/python:util",
"//tensorflow/python/data/ops:dataset_ops",
],
)

15
tensorflow/contrib/data/__init__.py
View File

@ -21,10 +21,15 @@
@@TextLineDataset
@@batch_and_drop_remainder
@@read_batch_features
@@rejection_resample
@@dense_to_sparse_batch
@@enumerate_dataset
@@group_by_window
@@ignore_errors
@@read_batch_features
@@unbatch
@@rejection_resample
@@sloppy_interleave
"""
from __future__ import absolute_import
@ -34,15 +39,19 @@ from __future__ import print_function
# pylint: disable=unused-import
from tensorflow.contrib.data.python.ops.dataset_ops import batch_and_drop_remainder
from tensorflow.contrib.data.python.ops.dataset_ops import Dataset
from tensorflow.contrib.data.python.ops.dataset_ops import dense_to_sparse_batch
from tensorflow.contrib.data.python.ops.dataset_ops import enumerate_dataset
from tensorflow.contrib.data.python.ops.dataset_ops import FixedLengthRecordDataset
from tensorflow.contrib.data.python.ops.dataset_ops import group_by_window
from tensorflow.contrib.data.python.ops.dataset_ops import Iterator
from tensorflow.contrib.data.python.ops.dataset_ops import ignore_errors
from tensorflow.contrib.data.python.ops.dataset_ops import read_batch_features
from tensorflow.contrib.data.python.ops.dataset_ops import rejection_resample
from tensorflow.contrib.data.python.ops.dataset_ops import SqlDataset
from tensorflow.contrib.data.python.ops.dataset_ops import TextLineDataset
from tensorflow.contrib.data.python.ops.dataset_ops import TFRecordDataset
from tensorflow.contrib.data.python.ops.dataset_ops import unbatch
from tensorflow.contrib.data.python.ops.sloppy_ops import sloppy_interleave
from tensorflow.python.data.ops.dataset_ops import Iterator
# pylint: enable=unused-import
from tensorflow.python.util.all_util import remove_undocumented

14
tensorflow/contrib/data/python/kernel_tests/BUILD
View File

@ -89,6 +89,7 @@ py_test(
"//tensorflow/python:math_ops",
"//tensorflow/python:string_ops",
"//tensorflow/python:tensor_shape",
"//tensorflow/python/data/ops:dataset_ops",
"//third_party/py/numpy",
],
)
@ -104,13 +105,17 @@ py_test(
],
deps = [
"//tensorflow/contrib/data/python/ops:dataset_ops",
"//tensorflow/core:protos_all_py",
"//tensorflow/python:array_ops",
"//tensorflow/python:client_testlib",
"//tensorflow/python:dtypes",
"//tensorflow/python:errors",
"//tensorflow/python:framework_ops",
"//tensorflow/python:math_ops",
"//tensorflow/python:resource_variable_ops",
"//tensorflow/python:session",
"//tensorflow/python:sparse_tensor",
"//tensorflow/python:util",
"//tensorflow/python/data/util:nest",
"//third_party/py/numpy",
],
)
@ -195,6 +200,7 @@ py_test(
"//tensorflow/python:data_flow_ops",
"//tensorflow/python:dtypes",
"//tensorflow/python:errors",
"//tensorflow/python:functional_ops",
"//tensorflow/python:io_ops",
"//tensorflow/python:lookup_ops",
"//tensorflow/python:math_ops",
@ -217,9 +223,13 @@ py_test(
"//tensorflow/python:array_ops",
"//tensorflow/python:client_testlib",
"//tensorflow/python:constant_op",
"//tensorflow/python:dataset_ops_gen",
"//tensorflow/python:dtypes",
"//tensorflow/python:errors",
"//tensorflow/python:framework_ops",
"//tensorflow/python:platform",
"//tensorflow/python:tensor_shape",
"//tensorflow/python:variables",
],
)
@ -358,10 +368,10 @@ py_test(
srcs_version = "PY2AND3",
deps = [
"//tensorflow/contrib/data/python/ops:dataset_ops",
"//tensorflow/contrib/data/python/util:nest",
"//tensorflow/python:client_testlib",
"//tensorflow/python:errors",
"//tensorflow/python:tensor_shape",
"//tensorflow/python/data/util:nest",
"//third_party/py/numpy",
],
)

16
tensorflow/contrib/data/python/kernel_tests/batch_dataset_op_test.py
View File

@ -229,8 +229,9 @@ class BatchDatasetTest(test.TestCase):
def testDenseToSparseBatchDataset(self):
components = np.random.randint(12, size=(100,)).astype(np.int32)
iterator = (dataset_ops.Dataset.from_tensor_slices(components)
.map(lambda x: array_ops.fill([x], x)).dense_to_sparse_batch(
4, [12]).make_initializable_iterator())
.map(lambda x: array_ops.fill([x], x)).apply(
dataset_ops.dense_to_sparse_batch(4, [12]))
.make_initializable_iterator())
init_op = iterator.initializer
get_next = sparse_tensor.SparseTensor(*iterator.get_next())
@ -253,8 +254,9 @@ class BatchDatasetTest(test.TestCase):
def testDenseToSparseBatchDatasetShapeErrors(self):
input_tensor = array_ops.placeholder(dtypes.int32)
iterator = (dataset_ops.Dataset.from_tensors(input_tensor)
.dense_to_sparse_batch(4, [12]).make_initializable_iterator())
iterator = (dataset_ops.Dataset.from_tensors(input_tensor).apply(
dataset_ops.dense_to_sparse_batch(4, [12]))
.make_initializable_iterator())
init_op = iterator.initializer
get_next = sparse_tensor.SparseTensor(*iterator.get_next())
@ -277,7 +279,7 @@ class BatchDatasetTest(test.TestCase):
expected_types = (dtypes.int32,) * 3
data = data.batch(2)
self.assertEqual(expected_types, data.output_types)
data = data.unbatch()
data = data.apply(dataset_ops.unbatch())
self.assertEqual(expected_types, data.output_types)
iterator = data.make_one_shot_iterator()
@ -296,7 +298,7 @@ class BatchDatasetTest(test.TestCase):
expected_types = ((dtypes.int32,),) * 3
data = data.batch(2)
self.assertEqual(expected_types, data.output_types)
data = data.unbatch()
data = data.apply(dataset_ops.unbatch())
self.assertEqual(expected_types, data.output_types)
iterator = data.make_one_shot_iterator()
@ -317,7 +319,7 @@ class BatchDatasetTest(test.TestCase):
expected_types = ((dtypes.int32, dtypes.string),) * 3
data = data.batch(2)
self.assertAllEqual(expected_types, data.output_types)
data = data.unbatch()
data = data.apply(dataset_ops.unbatch())
self.assertAllEqual(expected_types, data.output_types)
iterator = data.make_one_shot_iterator()

2
tensorflow/contrib/data/python/kernel_tests/concatenate_dataset_op_test.py
View File

@ -20,7 +20,7 @@ from __future__ import print_function
import numpy as np
from tensorflow.contrib.data.python.ops import dataset_ops
from tensorflow.contrib.data.python.util import nest
from tensorflow.python.data.util import nest
from tensorflow.python.framework import errors
from tensorflow.python.framework import tensor_shape
from tensorflow.python.platform import test

2
tensorflow/contrib/data/python/kernel_tests/dataset_constructor_op_test.py
View File

@ -22,9 +22,9 @@ import threading
import numpy as np
from tensorflow.contrib.data.python.ops import dataset_ops
from tensorflow.contrib.data.python.util import nest
from tensorflow.core.protobuf import config_pb2
from tensorflow.python.client import session
from tensorflow.python.data.util import nest
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import errors
from tensorflow.python.framework import ops

18
tensorflow/contrib/data/python/kernel_tests/map_dataset_op_test.py
View File

@ -271,8 +271,8 @@ class MapDatasetTest(test.TestCase):
components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)
dataset = (dataset_ops.Dataset.from_tensor_slices(components)
.map(lambda x: array_ops.check_numerics(x, "message"))
.ignore_errors())
.map(lambda x: array_ops.check_numerics(x, "message")).apply(
dataset_ops.ignore_errors()))
iterator = dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
@ -287,10 +287,10 @@ class MapDatasetTest(test.TestCase):
def testParallelMapIgnoreError(self):
components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)
dataset = (dataset_ops.Dataset.from_tensor_slices(components)
.map(lambda x: array_ops.check_numerics(x, "message"),
num_threads=2, output_buffer_size=2)
.ignore_errors())
dataset = (dataset_ops.Dataset.from_tensor_slices(components).map(
lambda x: array_ops.check_numerics(x, "message"),
num_threads=2,
output_buffer_size=2).apply(dataset_ops.ignore_errors()))
iterator = dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
@ -311,9 +311,9 @@ class MapDatasetTest(test.TestCase):
for filename in filenames:
write_string_to_file(filename, filename)
dataset = (dataset_ops.Dataset.from_tensor_slices(filenames)
.map(io_ops.read_file, num_threads=2, output_buffer_size=2)
.ignore_errors())
dataset = (dataset_ops.Dataset.from_tensor_slices(filenames).map(
io_ops.read_file, num_threads=2, output_buffer_size=2).apply(
dataset_ops.ignore_errors()))
iterator = dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()

4
tensorflow/contrib/data/python/kernel_tests/range_dataset_op_test.py
View File

@ -169,8 +169,8 @@ class RangeDatasetTest(test.TestCase):
components = (["a", "b"], [1, 2], [37.0, 38])
start = constant_op.constant(20, dtype=dtypes.int64)
iterator = (dataset_ops.Dataset.from_tensor_slices(components).enumerate(
start=start).make_initializable_iterator())
iterator = (dataset_ops.Dataset.from_tensor_slices(components).apply(
dataset_ops.enumerate_dataset(start)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()

10
tensorflow/contrib/data/python/ops/BUILD
View File

@ -9,9 +9,7 @@ py_library(
srcs = ["dataset_ops.py"],
srcs_version = "PY2AND3",
deps = [
"//tensorflow/contrib/data/python/util:nest",
"//tensorflow/python:array_ops",
"//tensorflow/python:constant_op",
"//tensorflow/python:control_flow_ops",
"//tensorflow/python:dataset_ops_gen",
"//tensorflow/python:dtypes",
@ -22,12 +20,13 @@ py_library(
"//tensorflow/python:parsing_ops",
"//tensorflow/python:platform",
"//tensorflow/python:random_ops",
"//tensorflow/python:random_seed",
"//tensorflow/python:resource_variable_ops",
"//tensorflow/python:script_ops",
"//tensorflow/python:sparse_tensor",
"//tensorflow/python:tensor_shape",
"//tensorflow/python:tensor_util",
"//tensorflow/python/data/ops:dataset_ops",
"//tensorflow/python/data/util:nest",
"//third_party/py/numpy",
],
)
@ -37,13 +36,12 @@ py_library(
srcs = ["sloppy_ops.py"],
srcs_version = "PY2AND3",
deps = [
":dataset_ops",
"//tensorflow/contrib/data/python/util:nest",
"//tensorflow/python:dataset_ops_gen",
"//tensorflow/python:dtypes",
"//tensorflow/python:framework_ops",
"//tensorflow/python:function",
"//tensorflow/python:platform",
"//tensorflow/python/data/ops:dataset_ops",
"//tensorflow/python/data/util:nest",
],
)

2038
tensorflow/contrib/data/python/ops/dataset_ops.py
View File

File diff suppressed because it is too large Load Diff

4
tensorflow/contrib/data/python/ops/sloppy_ops.py
View File

@ -17,8 +17,8 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.contrib.data.python.ops import dataset_ops
from tensorflow.contrib.data.python.util import nest
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.data.util import nest
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import function
from tensorflow.python.framework import ops

3
tensorflow/contrib/eager/python/BUILD
View File

@ -41,11 +41,10 @@ py_library(
srcs_version = "PY2AND3",
visibility = ["//tensorflow:internal"],
deps = [
"//tensorflow/contrib/data/python/ops:dataset_ops",
"//tensorflow/contrib/data/python/util:nest",
"//tensorflow/python:dataset_ops_gen",
"//tensorflow/python:errors",
"//tensorflow/python:resource_variable_ops",
"//tensorflow/python/data/util:nest",
"//tensorflow/python/eager:context",
],
)

2
tensorflow/contrib/eager/python/datasets.py
View File

@ -20,7 +20,7 @@ from __future__ import print_function
import threading
from tensorflow.contrib.data.python.util import nest
from tensorflow.python.data.util import nest
from tensorflow.python.eager import context
from tensorflow.python.framework import errors
from tensorflow.python.ops import gen_dataset_ops

1
tensorflow/contrib/fused_conv/BUILD
View File

@ -63,6 +63,7 @@ tf_kernel_library(
"kernels/fused_conv_ops_gpu.h",
],
prefix = "fused_conv2d_bias_activation_op",
visibility = ["//visibility:public"],
deps = [
"//tensorflow/core:framework",
"//tensorflow/core:lib",

1
tensorflow/contrib/gan/BUILD
View File

@ -41,7 +41,6 @@ py_library(
"//tensorflow/contrib/training:training_py",
"//tensorflow/python:array_ops",
"//tensorflow/python:check_ops",
"//tensorflow/python:dtypes",
"//tensorflow/python:framework_ops",
"//tensorflow/python:init_ops",
"//tensorflow/python:training",

5
tensorflow/contrib/gan/python/train.py
View File

@ -26,7 +26,6 @@ from tensorflow.contrib.gan.python import losses as tfgan_losses
from tensorflow.contrib.gan.python import namedtuples
from tensorflow.contrib.slim.python.slim import learning as slim_learning
from tensorflow.contrib.training.python.training import training
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import check_ops
@ -549,7 +548,7 @@ def gan_train_ops(
generator_global_step = variable_scope.get_variable(
'dummy_global_step_generator',
shape=[],
dtype=dtypes.int64,
dtype=global_step.dtype.base_dtype,
initializer=init_ops.zeros_initializer(),
trainable=False,
collections=[ops.GraphKeys.GLOBAL_VARIABLES])
@ -570,7 +569,7 @@ def gan_train_ops(
discriminator_global_step = variable_scope.get_variable(
'dummy_global_step_discriminator',
shape=[],
dtype=dtypes.int64,
dtype=global_step.dtype.base_dtype,
initializer=init_ops.zeros_initializer(),
trainable=False,
collections=[ops.GraphKeys.GLOBAL_VARIABLES])

11
tensorflow/contrib/gan/python/train_test.py
View File

@ -542,11 +542,17 @@ class GANTrainOpsTest(test.TestCase):
def test_unused_update_ops_callable_acgan_provideupdates(self):
self._test_unused_update_ops(create_callable_acgan_model, True)
def _test_sync_replicas_helper(self, create_gan_model_fn):
def _test_sync_replicas_helper(
self, create_gan_model_fn, create_global_step=False):
model = create_gan_model_fn()
loss = train.gan_loss(model)
num_trainable_vars = len(variables_lib.get_trainable_variables())
if create_global_step:
gstep = variable_scope.get_variable(
'custom_gstep', dtype=dtypes.int32, initializer=0, trainable=False)
ops.add_to_collection(ops.GraphKeys.GLOBAL_STEP, gstep)
g_opt = get_sync_optimizer()
d_opt = get_sync_optimizer()
train_ops = train.gan_train_ops(
@ -610,6 +616,9 @@ class GANTrainOpsTest(test.TestCase):
def test_sync_replicas_callable_acgan(self):
self._test_sync_replicas_helper(create_callable_acgan_model)
def test_global_step_can_be_int32(self):
self._test_sync_replicas_helper(create_gan_model, create_global_step=True)
class GANTrainTest(test.TestCase):
"""Tests for `gan_train`."""

2
tensorflow/contrib/receptive_field/README.md
View File

@ -38,7 +38,7 @@ models are available to you. This can be done in three simple commands:
```sh
git clone https://github.com/tensorflow/models
cd models/slim
cd models/research/slim
sudo python setup.py install_lib
```

4
tensorflow/contrib/seq2seq/python/kernel_tests/beam_search_decoder_test.py
View File

@ -121,7 +121,7 @@ class TestBeamStep(test.TestCase):
log_probs=nn_ops.log_softmax(
array_ops.ones([self.batch_size, self.beam_width])),
lengths=constant_op.constant(
2, shape=[self.batch_size, self.beam_width], dtype=dtypes.int32),
2, shape=[self.batch_size, self.beam_width], dtype=dtypes.int64),
finished=array_ops.zeros(
[self.batch_size, self.beam_width], dtype=dtypes.bool))
@ -176,7 +176,7 @@ class TestBeamStep(test.TestCase):
log_probs=nn_ops.log_softmax(
array_ops.ones([self.batch_size, self.beam_width])),
lengths=ops.convert_to_tensor(
[[2, 1, 2], [2, 2, 1]], dtype=dtypes.int32),
[[2, 1, 2], [2, 2, 1]], dtype=dtypes.int64),
finished=ops.convert_to_tensor(
[[False, True, False], [False, False, True]], dtype=dtypes.bool))

16
tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py
View File

@ -22,6 +22,7 @@ import collections
from tensorflow.contrib.seq2seq.python.ops import beam_search_ops
from tensorflow.contrib.seq2seq.python.ops import decoder
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
@ -256,7 +257,7 @@ class BeamSearchDecoder(decoder.Decoder):
dtype=nest.flatten(self._initial_cell_state)[0].dtype),
finished=finished,
lengths=array_ops.zeros(
[self._batch_size, self._beam_width], dtype=dtypes.int32))
[self._batch_size, self._beam_width], dtype=dtypes.int64))
return (finished, start_inputs, initial_state)
@ -267,7 +268,7 @@ class BeamSearchDecoder(decoder.Decoder):
outputs: An instance of BeamSearchDecoderOutput.
final_state: An instance of BeamSearchDecoderState. Passed through to the
output.
sequence_lengths: An `int32` tensor shaped `[batch_size, beam_width]`.
sequence_lengths: An `int64` tensor shaped `[batch_size, beam_width]`.
The sequence lengths determined for each beam during decode.
Returns:
@ -491,9 +492,10 @@ def _beam_search_step(time, logits, next_cell_state, beam_state, batch_size,
indices=array_ops.tile(
array_ops.reshape(end_token, [1, 1]), [batch_size, beam_width]),
depth=vocab_size,
on_value=0,
off_value=1)
add_mask = (1 - math_ops.to_int32(previously_finished))
on_value=constant_op.constant(0, dtype=dtypes.int64),
off_value=constant_op.constant(1, dtype=dtypes.int64),
dtype=dtypes.int64)
add_mask = (1 - math_ops.to_int64(previously_finished))
lengths_to_add = array_ops.expand_dims(add_mask, 2) * lengths_to_add
new_prediction_lengths = (
lengths_to_add + array_ops.expand_dims(prediction_lengths, 2))
@ -547,9 +549,9 @@ def _beam_search_step(time, logits, next_cell_state, beam_state, batch_size,
# 1. Finished beams remain unchanged
# 2. Beams that are now finished (EOS predicted) remain unchanged
# 3. Beams that are not yet finished have their length increased by 1
lengths_to_add = math_ops.to_int32(
lengths_to_add = math_ops.to_int64(
math_ops.not_equal(next_word_ids, end_token))
lengths_to_add = (1 - math_ops.to_int32(next_finished)) * lengths_to_add
lengths_to_add = (1 - math_ops.to_int64(next_finished)) * lengths_to_add
next_prediction_len = _tensor_gather_helper(
gather_indices=next_beam_ids,
gather_from=beam_state.lengths,

4
tensorflow/contrib/signal/python/kernel_tests/mel_ops_test.py
View File

@ -33,7 +33,7 @@ def hertz_to_mel(frequencies_hertz):
"""Convert frequencies to mel scale using HTK formula.
Copied from
https://github.com/tensorflow/models/blob/master/audioset/mel_features.py.
https://github.com/tensorflow/models/blob/master/research/audioset/mel_features.py.
Args:
frequencies_hertz: Scalar or np.array of frequencies in hertz.
@ -54,7 +54,7 @@ def spectrogram_to_mel_matrix(num_mel_bins=20,
"""Return a matrix that can post-multiply spectrogram rows to make mel.
Copied from
https://github.com/tensorflow/models/blob/master/audioset/mel_features.py.
https://github.com/tensorflow/models/blob/master/research/audioset/mel_features.py.
Returns a np.array matrix A that can be used to post-multiply a matrix S of
spectrogram values (STFT magnitudes) arranged as frames x bins to generate a

1
tensorflow/contrib/tpu/BUILD
View File

@ -13,6 +13,7 @@ load("//tensorflow:tensorflow.bzl", "tf_py_test")
package(
default_visibility = [
"//cloud/vmm/testing/tests/tpu:__subpackages__",
"//learning/brain:__subpackages__",
"//tensorflow:__subpackages__",
],

3
tensorflow/contrib/verbs/BUILD
View File

@ -50,11 +50,8 @@ cc_library(
srcs = ["verbs_util.cc"],
hdrs = ["verbs_util.h"],
deps = [
"//tensorflow/core:core_cpu_internal",
"//tensorflow/core:framework",
"//tensorflow/core:gpu_runtime",
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
],
)

8
tensorflow/contrib/verbs/verbs_util.cc
View File

@ -15,9 +15,13 @@ limitations under the License.
#include "tensorflow/contrib/verbs/verbs_util.h"
#include "tensorflow/core/common_runtime/gpu/gpu_util.h"
#include "tensorflow/core/lib/core/notification.h"
#include <vector>
#include "tensorflow/core/lib/core/stringpiece.h"
#include "tensorflow/core/lib/strings/numbers.h"
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
namespace tensorflow {
// static

6
tensorflow/contrib/verbs/verbs_util.h
View File

@ -18,14 +18,10 @@ limitations under the License.
#include <string>
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/framework/types.h"
namespace tensorflow {
class TensorProto;
class VerbsUtil {
public:
static string AppendStepidToKey(const string& key, int64 step_id);

6
tensorflow/core/grappler/BUILD
View File

@ -55,12 +55,14 @@ tf_cuda_library(
name = "devices",
srcs = ["devices.cc"],
hdrs = ["devices.h"],
cuda_deps = [
"//tensorflow/core:gpu_init",
"//tensorflow/core:stream_executor",
],
visibility = ["//visibility:public"],
deps = [
"//tensorflow/core:gpu_init",
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
"//tensorflow/core:stream_executor",
],
)

10
tensorflow/core/kernels/BUILD
View File

@ -2300,7 +2300,12 @@ tf_kernel_library(
tf_kernel_library(
name = "self_adjoint_eig_v2_op",
prefix = "self_adjoint_eig_v2_op",
deps = LINALG_DEPS,
deps = LINALG_DEPS + if_cuda([
":cast_op",
":cwise_op",
":cuda_solvers",
":transpose_functor",
]),
)
tf_kernel_library(
@ -4011,7 +4016,6 @@ tf_kernel_library(
name = "word2vec_kernels",
prefix = "word2vec_kernels",
deps = [
"//tensorflow/core",
"//tensorflow/core:framework",
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
@ -4803,8 +4807,8 @@ tf_kernel_library(
":image_resizer_state",
":ops_util",
":pooling_ops",
"//tensorflow/core",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:core_cpu",
"//tensorflow/core:framework",
"//tensorflow/core:lib",
"//tensorflow/core:math_ops_op_lib",

284
tensorflow/core/kernels/cuda_solvers.cc
View File

@ -194,15 +194,14 @@ Status CudaSolver::CopyLapackInfoToHostAsync(
#define TF_CALL_LAPACK_TYPES_NO_COMPLEX(m) m(float, S) m(double, D)
// Macros to construct cusolverDn method names.
#define DN_SOLVER_FN(method, lapack_prefix) cusolverDn##lapack_prefix##method
#define DN_SOLVER_NAME(method, lapack_prefix) \
"cusolverDn" #lapack_prefix #method
#define DN_BUFSIZE_FN(method, lapack_prefix) \
cusolverDn##lapack_prefix##method##_bufferSize
#define DN_SOLVER_FN(method, type_prefix) cusolverDn##type_prefix##method
#define DN_SOLVER_NAME(method, type_prefix) "cusolverDn" #type_prefix #method
#define DN_BUFSIZE_FN(method, type_prefix) \
cusolverDn##type_prefix##method##_bufferSize
// Macros to construct cublas method names.
#define BLAS_SOLVER_FN(method, lapack_prefix) cublas##lapack_prefix##method
#define BLAS_SOLVER_NAME(method, lapack_prefix) "cublas" #lapack_prefix #method
#define BLAS_SOLVER_FN(method, type_prefix) cublas##type_prefix##method
#define BLAS_SOLVER_NAME(method, type_prefix) "cublas" #type_prefix #method
//=============================================================================
// Wrappers of cuSolverDN computational methods begin here.
@ -229,17 +228,16 @@ static inline Status GeamImpl(SolverFnT solver, cublasHandle_t cublas_handle,
return Status::OK();
}
#define GEAM_INSTANCE(Scalar, lapack_prefix) \
template <> \
Status CudaSolver::Geam<Scalar>( \
cublasOperation_t transa, cublasOperation_t transb, int m, int n, \
const Scalar* alpha, /* host or device pointer */ \
const Scalar* A, int lda, \
const Scalar* beta, /* host or device pointer */ \
const Scalar* B, int ldb, Scalar* C, int ldc) const { \
return GeamImpl(BLAS_SOLVER_FN(geam, lapack_prefix), cublas_handle_, \
transa, transb, m, n, alpha, A, lda, beta, B, ldb, C, \
ldc); \
#define GEAM_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::Geam<Scalar>( \
cublasOperation_t transa, cublasOperation_t transb, int m, int n, \
const Scalar* alpha, /* host or device pointer */ \
const Scalar* A, int lda, \
const Scalar* beta, /* host or device pointer */ \
const Scalar* B, int ldb, Scalar* C, int ldc) const { \
return GeamImpl(BLAS_SOLVER_FN(geam, type_prefix), cublas_handle_, transa, \
transb, m, n, alpha, A, lda, beta, B, ldb, C, ldc); \
}
TF_CALL_LAPACK_TYPES(GEAM_INSTANCE);
@ -263,12 +261,12 @@ static inline Status PotrfImpl(BufSizeFnT bufsize, SolverFnT solver,
return Status::OK();
}
#define POTRF_INSTANCE(Scalar, lapack_prefix) \
#define POTRF_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::Potrf<Scalar>(cublasFillMode_t uplo, int n, Scalar* A, \
int lda, int* dev_lapack_info) const { \
return PotrfImpl(DN_BUFSIZE_FN(potrf, lapack_prefix), \
DN_SOLVER_FN(potrf, lapack_prefix), context_, \
return PotrfImpl(DN_BUFSIZE_FN(potrf, type_prefix), \
DN_SOLVER_FN(potrf, type_prefix), context_, \
cusolver_dn_handle_, uplo, n, A, lda, dev_lapack_info); \
}
@ -293,13 +291,13 @@ static inline Status GetrfImpl(BufSizeFnT bufsize, SolverFnT solver,
return Status::OK();
}
#define GETRF_INSTANCE(Scalar, lapack_prefix) \
#define GETRF_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::Getrf<Scalar>(int m, int n, Scalar* A, int lda, \
int* dev_pivots, int* dev_lapack_info) \
const { \
return GetrfImpl(DN_BUFSIZE_FN(getrf, lapack_prefix), \
DN_SOLVER_FN(getrf, lapack_prefix), context_, \
return GetrfImpl(DN_BUFSIZE_FN(getrf, type_prefix), \
DN_SOLVER_FN(getrf, type_prefix), context_, \
cusolver_dn_handle_, m, n, A, lda, dev_pivots, \
dev_lapack_info); \
}
@ -319,53 +317,18 @@ static inline Status GetrsImpl(SolverFnT solver, OpKernelContext* context,
return Status::OK();
}
#define GETRS_INSTANCE(Scalar, lapack_prefix) \
#define GETRS_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::Getrs<Scalar>( \
cublasOperation_t trans, int n, int nrhs, const Scalar* A, int lda, \
const int* pivots, Scalar* B, int ldb, int* dev_lapack_info) const { \
return GetrsImpl(DN_SOLVER_FN(getrs, lapack_prefix), context_, \
return GetrsImpl(DN_SOLVER_FN(getrs, type_prefix), context_, \
cusolver_dn_handle_, trans, n, nrhs, A, lda, pivots, B, \
ldb, dev_lapack_info); \
}
TF_CALL_LAPACK_TYPES(GETRS_INSTANCE);
template <typename Scalar, typename BufSizeFnT, typename SolverFnT>
static inline Status GesvdImpl(BufSizeFnT bufsize, SolverFnT solver,
OpKernelContext* context,
cusolverDnHandle_t cusolver_dn_handle,
signed char jobu, signed char jobvt, int m,
int n, Scalar* A, int lda, Scalar* S, Scalar* U,
int ldu, Scalar* VT, int ldvt,
int* dev_lapack_info) {
/* Get amount of workspace memory required. */
int lwork;
TF_RETURN_IF_CUSOLVER_ERROR(bufsize(cusolver_dn_handle, m, n, &lwork));
/* Allocate device memory for workspace. */
ScratchSpace<Scalar> dev_workspace(context, lwork, /* on_host */ false);
/* Launch the solver kernel. */
TF_RETURN_IF_CUSOLVER_ERROR(solver(
cusolver_dn_handle, jobu, jobvt, m, n, CUDAComplex(A), lda, S,
CUDAComplex(U), ldu, CUDAComplex(VT), ldvt,
CUDAComplex(dev_workspace.mutable_data()), lwork, NULL, dev_lapack_info));
return Status::OK();
}
#define GESVD_INSTANCE(Scalar, lapack_prefix) \
template <> \
Status CudaSolver::Gesvd<Scalar>( \
signed char jobu, signed char jobvt, int m, int n, Scalar* dev_A, \
int lda, Scalar* dev_S, Scalar* dev_U, int ldu, Scalar* dev_VT, \
int ldvt, int* dev_lapack_info) const { \
return GesvdImpl(DN_BUFSIZE_FN(gesvd, lapack_prefix), \
DN_SOLVER_FN(gesvd, lapack_prefix), context_, \
cusolver_dn_handle_, jobu, jobvt, m, n, dev_A, lda, \
dev_S, dev_U, ldu, dev_VT, ldvt, dev_lapack_info); \
}
TF_CALL_LAPACK_TYPES_NO_COMPLEX(GESVD_INSTANCE);
template <typename Scalar, typename BufSizeFnT, typename SolverFnT>
static inline Status GeqrfImpl(BufSizeFnT bufsize, SolverFnT solver,
OpKernelContext* context,
@ -385,19 +348,19 @@ static inline Status GeqrfImpl(BufSizeFnT bufsize, SolverFnT solver,
return Status::OK();
}
#define GEQRF_INSTANCE(Scalar, lapack_prefix) \
#define GEQRF_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::Geqrf<Scalar>(int m, int n, Scalar* A, int lda, \
Scalar* tau, int* dev_lapack_info) const { \
return GeqrfImpl(DN_BUFSIZE_FN(geqrf, lapack_prefix), \
DN_SOLVER_FN(geqrf, lapack_prefix), context_, \
return GeqrfImpl(DN_BUFSIZE_FN(geqrf, type_prefix), \
DN_SOLVER_FN(geqrf, type_prefix), context_, \
cusolver_dn_handle_, m, n, A, lda, tau, dev_lapack_info); \
}
TF_CALL_LAPACK_TYPES(GEQRF_INSTANCE);
template <typename Scalar, typename BufSizeFnT, typename SolverFnT>
static inline Status OrmqrImpl(BufSizeFnT bufsize, SolverFnT solver,
static inline Status UnmqrImpl(BufSizeFnT bufsize, SolverFnT solver,
OpKernelContext* context,
cusolverDnHandle_t cusolver_dn_handle,
cublasSideMode_t side, cublasOperation_t trans,
@ -422,47 +385,25 @@ static inline Status OrmqrImpl(BufSizeFnT bufsize, SolverFnT solver,
// Unfortunately the LAPACK function name differs for the real and complex case
// (complex ones are prefixed with "UN" for "unitary"), so we instantiate each
// one separately.
template <>
Status CudaSolver::Ormqr(cublasSideMode_t side, cublasOperation_t trans, int m,
int n, int k, const float* dev_a, int lda,
const float* dev_tau, float* dev_c, int ldc,
int* dev_lapack_info) const {
return OrmqrImpl(DN_BUFSIZE_FN(ormqr, S), DN_SOLVER_FN(ormqr, S), context_,
cusolver_dn_handle_, side, trans, m, n, k, dev_a, lda,
dev_tau, dev_c, ldc, dev_lapack_info);
}
template <>
Status CudaSolver::Ormqr(cublasSideMode_t side, cublasOperation_t trans, int m,
int n, int k, const double* dev_a, int lda,
const double* dev_tau, double* dev_c, int ldc,
int* dev_lapack_info) const {
return OrmqrImpl(DN_BUFSIZE_FN(ormqr, D), DN_SOLVER_FN(ormqr, D), context_,
cusolver_dn_handle_, side, trans, m, n, k, dev_a, lda,
dev_tau, dev_c, ldc, dev_lapack_info);
}
template <>
Status CudaSolver::Ormqr(cublasSideMode_t side, cublasOperation_t trans, int m,
int n, int k, const std::complex<float>* dev_a,
int lda, const std::complex<float>* dev_tau,
std::complex<float>* dev_c, int ldc,
int* dev_lapack_info) const {
return OrmqrImpl(DN_BUFSIZE_FN(unmqr, C), DN_SOLVER_FN(unmqr, C), context_,
cusolver_dn_handle_, side, trans, m, n, k, dev_a, lda,
dev_tau, dev_c, ldc, dev_lapack_info);
}
template <>
Status CudaSolver::Ormqr(cublasSideMode_t side, cublasOperation_t trans, int m,
int n, int k, const std::complex<double>* dev_a,
int lda, const std::complex<double>* dev_tau,
std::complex<double>* dev_c, int ldc,
int* dev_lapack_info) const {
return OrmqrImpl(DN_BUFSIZE_FN(unmqr, Z), DN_SOLVER_FN(unmqr, Z), context_,
cusolver_dn_handle_, side, trans, m, n, k, dev_a, lda,
dev_tau, dev_c, ldc, dev_lapack_info);
}
#define UNMQR_INSTANCE(Scalar, function_prefix, type_prefix) \
template <> \
Status CudaSolver::Unmqr(cublasSideMode_t side, cublasOperation_t trans, \
int m, int n, int k, const Scalar* dev_a, int lda, \
const Scalar* dev_tau, Scalar* dev_c, int ldc, \
int* dev_lapack_info) const { \
return UnmqrImpl(DN_BUFSIZE_FN(function_prefix##mqr, type_prefix), \
DN_SOLVER_FN(function_prefix##mqr, type_prefix), \
context_, cusolver_dn_handle_, side, trans, m, n, k, \
dev_a, lda, dev_tau, dev_c, ldc, dev_lapack_info); \
}
UNMQR_INSTANCE(float, or, S);
UNMQR_INSTANCE(double, or, D);
UNMQR_INSTANCE(complex64, un, C);
UNMQR_INSTANCE(complex128, un, Z);
template <typename Scalar, typename BufSizeFnT, typename SolverFnT>
static inline Status OrgqrImpl(BufSizeFnT bufsize, SolverFnT solver,
static inline Status UngqrImpl(BufSizeFnT bufsize, SolverFnT solver,
OpKernelContext* context,
cusolverDnHandle_t cusolver_dn_handle, int m,
int n, int k, Scalar* dev_a, int lda,
@ -482,40 +423,97 @@ static inline Status OrgqrImpl(BufSizeFnT bufsize, SolverFnT solver,
return Status::OK();
}
// Unfortunately the LAPACK function name differs for the real and complex case
// (complex ones are prefixed with "UN" for "unitary"), so we instantiate each
// one separately.
template <>
Status CudaSolver::Orgqr(int m, int n, int k, float* dev_a, int lda,
const float* dev_tau, int* dev_lapack_info) const {
return OrgqrImpl(DN_BUFSIZE_FN(orgqr, S), DN_SOLVER_FN(orgqr, S), context_,
cusolver_dn_handle_, m, n, k, dev_a, lda, dev_tau,
dev_lapack_info);
#define UNGQR_INSTANCE(Scalar, function_prefix, type_prefix) \
template <> \
Status CudaSolver::Ungqr(int m, int n, int k, Scalar* dev_a, int lda, \
const Scalar* dev_tau, int* dev_lapack_info) \
const { \
return UngqrImpl(DN_BUFSIZE_FN(function_prefix##gqr, type_prefix), \
DN_SOLVER_FN(function_prefix##gqr, type_prefix), \
context_, cusolver_dn_handle_, m, n, k, dev_a, lda, \
dev_tau, dev_lapack_info); \
}
UNGQR_INSTANCE(float, or, S);
UNGQR_INSTANCE(double, or, D);
UNGQR_INSTANCE(complex64, un, C);
UNGQR_INSTANCE(complex128, un, Z);
template <typename Scalar, typename BufSizeFnT, typename SolverFnT>
static inline Status HeevdImpl(BufSizeFnT bufsize, SolverFnT solver,
OpKernelContext* context,
cusolverDnHandle_t cusolver_dn_handle,
cusolverEigMode_t jobz, cublasFillMode_t uplo,
int n, Scalar* dev_A, int lda,
typename Eigen::NumTraits<Scalar>::Real* dev_W,
int* dev_lapack_info) {
/* Get amount of workspace memory required. */
int lwork;
TF_RETURN_IF_CUSOLVER_ERROR(bufsize(cusolver_dn_handle, jobz, uplo, n,
CUDAComplex(dev_A), lda,
CUDAComplex(dev_W), &lwork));
/* Allocate device memory for workspace. */
ScratchSpace<Scalar> dev_workspace(context, lwork, /* on_host */ false);
/* Launch the solver kernel. */
TF_RETURN_IF_CUSOLVER_ERROR(
solver(cusolver_dn_handle, jobz, uplo, n, CUDAComplex(dev_A), lda,
CUDAComplex(dev_W), CUDAComplex(dev_workspace.mutable_data()),
lwork, dev_lapack_info));
return Status::OK();
}
template <>
Status CudaSolver::Orgqr(int m, int n, int k, double* dev_a, int lda,
const double* dev_tau, int* dev_lapack_info) const {
return OrgqrImpl(DN_BUFSIZE_FN(orgqr, D), DN_SOLVER_FN(orgqr, D), context_,
cusolver_dn_handle_, m, n, k, dev_a, lda, dev_tau,
dev_lapack_info);
}
template <>
Status CudaSolver::Orgqr(int m, int n, int k, std::complex<float>* dev_a,
int lda, const std::complex<float>* dev_tau,
int* dev_lapack_info) const {
return OrgqrImpl(DN_BUFSIZE_FN(ungqr, C), DN_SOLVER_FN(ungqr, C), context_,
cusolver_dn_handle_, m, n, k, dev_a, lda, dev_tau,
dev_lapack_info);
}
template <>
Status CudaSolver::Orgqr(int m, int n, int k, std::complex<double>* dev_a,
int lda, const std::complex<double>* dev_tau,
int* dev_lapack_info) const {
return OrgqrImpl(DN_BUFSIZE_FN(ungqr, Z), DN_SOLVER_FN(ungqr, Z), context_,
cusolver_dn_handle_, m, n, k, dev_a, lda, dev_tau,
dev_lapack_info);
#define HEEVD_INSTANCE(Scalar, function_prefix, type_prefix) \
template <> \
Status CudaSolver::Heevd(cusolverEigMode_t jobz, cublasFillMode_t uplo, \
int n, Scalar* dev_A, int lda, \
typename Eigen::NumTraits<Scalar>::Real* dev_W, \
int* dev_lapack_info) const { \
return HeevdImpl(DN_BUFSIZE_FN(function_prefix##evd, type_prefix), \
DN_SOLVER_FN(function_prefix##evd, type_prefix), \
context_, cusolver_dn_handle_, jobz, uplo, n, dev_A, lda, \
dev_W, dev_lapack_info); \
}
HEEVD_INSTANCE(float, sy, S);
HEEVD_INSTANCE(double, sy, D);
HEEVD_INSTANCE(complex64, he, C);
HEEVD_INSTANCE(complex128, he, Z);
template <typename Scalar, typename BufSizeFnT, typename SolverFnT>
static inline Status GesvdImpl(BufSizeFnT bufsize, SolverFnT solver,
OpKernelContext* context,
cusolverDnHandle_t cusolver_dn_handle,
signed char jobu, signed char jobvt, int m,
int n, Scalar* A, int lda, Scalar* S, Scalar* U,
int ldu, Scalar* VT, int ldvt,
int* dev_lapack_info) {
/* Get amount of workspace memory required. */
int lwork;
TF_RETURN_IF_CUSOLVER_ERROR(bufsize(cusolver_dn_handle, m, n, &lwork));
/* Allocate device memory for workspace. */
ScratchSpace<Scalar> dev_workspace(context, lwork, /* on_host */ false);
/* Launch the solver kernel. */
TF_RETURN_IF_CUSOLVER_ERROR(solver(
cusolver_dn_handle, jobu, jobvt, m, n, CUDAComplex(A), lda, S,
CUDAComplex(U), ldu, CUDAComplex(VT), ldvt,
CUDAComplex(dev_workspace.mutable_data()), lwork, NULL, dev_lapack_info));
return Status::OK();
}
#define GESVD_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::Gesvd<Scalar>( \
signed char jobu, signed char jobvt, int m, int n, Scalar* dev_A, \
int lda, Scalar* dev_S, Scalar* dev_U, int ldu, Scalar* dev_VT, \
int ldvt, int* dev_lapack_info) const { \
return GesvdImpl(DN_BUFSIZE_FN(gesvd, type_prefix), \
DN_SOLVER_FN(gesvd, type_prefix), context_, \
cusolver_dn_handle_, jobu, jobvt, m, n, dev_A, lda, \
dev_S, dev_U, ldu, dev_VT, ldvt, dev_lapack_info); \
}
TF_CALL_LAPACK_TYPES_NO_COMPLEX(GESVD_INSTANCE);
//=============================================================================
// Wrappers of cuBlas computational methods begin here.
//
@ -542,12 +540,12 @@ static inline Status GetrfBatchedImpl(
return Status::OK();
}
#define GETRF_BATCHED_INSTANCE(Scalar, lapack_prefix) \
#define GETRF_BATCHED_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::GetrfBatched( \
int n, const Scalar* host_a_dev_ptrs[], int lda, int* dev_pivots, \
DeviceLapackInfo* dev_lapack_info, int batch_size) const { \
return GetrfBatchedImpl(BLAS_SOLVER_FN(getrfBatched, lapack_prefix), \
return GetrfBatchedImpl(BLAS_SOLVER_FN(getrfBatched, type_prefix), \
context_, cublas_handle_, n, host_a_dev_ptrs, lda, \
dev_pivots, dev_lapack_info, batch_size); \
}
@ -580,14 +578,14 @@ static inline Status GetrsBatchedImpl(
return Status::OK();
}
#define GETRS_BATCHED_INSTANCE(Scalar, lapack_prefix) \
#define GETRS_BATCHED_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::GetrsBatched( \
cublasOperation_t trans, int n, int nrhs, \
const Scalar* host_a_dev_ptrs[], int lda, const int* dev_pivots, \
const Scalar* host_b_dev_ptrs[], int ldb, \
DeviceLapackInfo* dev_lapack_info, int batch_size) const { \
return GetrsBatchedImpl(BLAS_SOLVER_FN(getrsBatched, lapack_prefix), \
return GetrsBatchedImpl(BLAS_SOLVER_FN(getrsBatched, type_prefix), \
context_, cublas_handle_, trans, n, nrhs, \
host_a_dev_ptrs, lda, dev_pivots, host_b_dev_ptrs, \
ldb, dev_lapack_info, batch_size); \
@ -619,13 +617,13 @@ static inline Status GetriBatchedImpl(
return Status::OK();
}
#define GETRI_BATCHED_INSTANCE(Scalar, lapack_prefix) \
#define GETRI_BATCHED_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::GetriBatched( \
int n, const Scalar* host_a_dev_ptrs[], int lda, const int* dev_pivots, \
const Scalar* host_a_inv_dev_ptrs[], int ldainv, \
DeviceLapackInfo* dev_lapack_info, int batch_size) const { \
return GetriBatchedImpl(BLAS_SOLVER_FN(getriBatched, lapack_prefix), \
return GetriBatchedImpl(BLAS_SOLVER_FN(getriBatched, type_prefix), \
context_, cublas_handle_, n, host_a_dev_ptrs, lda, \
dev_pivots, host_a_inv_dev_ptrs, ldainv, \
dev_lapack_info, batch_size); \
@ -657,13 +655,13 @@ static inline Status MatInvBatchedImpl(
return Status::OK();
}
#define MATINV_BATCHED_INSTANCE(Scalar, lapack_prefix) \
#define MATINV_BATCHED_INSTANCE(Scalar, type_prefix) \
template <> \
Status CudaSolver::MatInvBatched( \
int n, const Scalar* host_a_dev_ptrs[], int lda, \
const Scalar* host_a_inv_dev_ptrs[], int ldainv, \
DeviceLapackInfo* dev_lapack_info, int batch_size) const { \
return MatInvBatchedImpl(BLAS_SOLVER_FN(matinvBatched, lapack_prefix), \
return MatInvBatchedImpl(BLAS_SOLVER_FN(matinvBatched, type_prefix), \
context_, cublas_handle_, n, host_a_dev_ptrs, \
lda, host_a_inv_dev_ptrs, ldainv, \
dev_lapack_info, batch_size); \

37
tensorflow/core/kernels/cuda_solvers.h
View File

@ -242,32 +242,39 @@ class CudaSolver {
Status Geqrf(int m, int n, Scalar* dev_A, int lda, Scalar* dev_tau,
int* dev_lapack_info) const TF_MUST_USE_RESULT;
// Overwrite matrix C by product of C and Householder matrix Q. The
// Householder matrix Q is represented by the output from Geqrf in dev_a and
// dev_tau.
// Overwrite matrix C by product of C and the unitary Householder matrix Q.
// The Householder matrix Q is represented by the output from Geqrf in dev_a
// and dev_tau.
// Notice: If Scalar is real, only trans=CUBLAS_OP_N or trans=CUBLAS_OP_T is
// supported. If Scalar is complex, trans=CUBLAS_OP_N or trans=CUBLAS_OP_C is
// supported.
// Returns Status::OK() if the kernel was launched successfully.
// See: http://docs.nvidia.com/cuda/cusolver/#cuds-lt-t-gt-ormqr
template <typename Scalar>
Status Ormqr(cublasSideMode_t side, cublasOperation_t trans, int m, int n,
Status Unmqr(cublasSideMode_t side, cublasOperation_t trans, int m, int n,
int k, const Scalar* dev_a, int lda, const Scalar* dev_tau,
Scalar* dev_c, int ldc,
int* dev_lapack_info) const TF_MUST_USE_RESULT;
// Overwrites QR factorization produced by Geqrf by Householder matrix Q.
// On input, the Householder matrix Q is represented by the output from Geqrf
// in dev_a and dev_tau. On output, dev_a is overwritten with the first n
// columns of Q.
// Requires m >= n >= 0.
// Overwrites QR factorization produced by Geqrf by the unitary Householder
// matrix Q. On input, the Householder matrix Q is represented by the output
// from Geqrf in dev_a and dev_tau. On output, dev_a is overwritten with the
// first n columns of Q. Requires m >= n >= 0.
// Returns Status::OK() if the kernel was launched successfully.
// See: http://docs.nvidia.com/cuda/cusolver/#cuds-lt-t-gt-orgqr
template <typename Scalar>
Status Orgqr(int m, int n, int k, Scalar* dev_a, int lda,
Status Ungqr(int m, int n, int k, Scalar* dev_a, int lda,
const Scalar* dev_tau,
int* dev_lapack_info) const TF_MUST_USE_RESULT;
// Hermitian (Symmetric) Eigen decomposition.
// See: http://docs.nvidia.com/cuda/cusolver/#cuds-lt-t-gt-syevd
template <typename Scalar>
Status Heevd(cusolverEigMode_t jobz, cublasFillMode_t uplo, int n,
Scalar* dev_A, int lda,
typename Eigen::NumTraits<Scalar>::Real* dev_W,
int* dev_lapack_info) const TF_MUST_USE_RESULT;
// Singular value decomposition.
// Returns Status::OK() if the kernel was launched successfully.
// TODO(rmlarsen, volunteers): Add support for complex types.
@ -277,16 +284,6 @@ class CudaSolver {
int lda, Scalar* dev_S, Scalar* dev_U, int ldu, Scalar* dev_VT,
int ldvt, int* dev_lapack_info) const TF_MUST_USE_RESULT;
/*
TODO(rmlarsen, volunteers): Implement the kernels below.
// Symmetric/Hermitian Eigen decomposition.
// See: http://docs.nvidia.com/cuda/cusolver/#cuds-lt-t-gt-syevd
template <typename Scalar>
Status Syevd(cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, Scalar*
dev_A, int lda, Scalar* dev_W, int* dev_lapack_info) const TF_MUST_USE_RESULT;
*/
private:
OpKernelContext* context_; // not owned.
cudaStream_t cuda_stream_;

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

@ -159,6 +159,8 @@ namespace functor {
DECLARE_GPU_SPEC_INDEX(T, int64);
TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC);
TF_CALL_int32(DECLARE_GPU_SPEC);
TF_CALL_int64(DECLARE_GPU_SPEC);
#undef DECLARE_GPU_SPEC_INDEX
#undef DECLARE_GPU_SPEC
@ -180,6 +182,8 @@ TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC);
REGISTER_ONE_HOT_GPU_INDEX(type, int64);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_ONE_HOT_GPU);
TF_CALL_int32(REGISTER_ONE_HOT_GPU);
TF_CALL_int64(REGISTER_ONE_HOT_GPU);
#undef REGISTER_ONE_HOT_GPU_INDEX
#undef REGISTER_ONE_HOT_GPU

2
tensorflow/core/kernels/one_hot_op_gpu.cu.cc
View File

@ -37,6 +37,8 @@ typedef Eigen::GpuDevice GPUDevice;
DEFINE_GPU_SPEC_INDEX(T, int64)
TF_CALL_GPU_NUMBER_TYPES(DEFINE_GPU_SPEC);
TF_CALL_int32(DEFINE_GPU_SPEC);
TF_CALL_int64(DEFINE_GPU_SPEC);
#undef DEFINE_GPU_SPEC_INDEX
#undef DEFINE_GPU_SPEC

8
tensorflow/core/kernels/qr_op_impl.h
View File

@ -248,12 +248,12 @@ class QrOpGpu : public AsyncOpKernel {
auto q_reshaped = q->flat_inner_dims<Scalar, 3>();
eye(device, q_reshaped);
for (int batch = 0; batch < batch_size; ++batch) {
// Notice: It appears that Ormqr does not write a zero into *info upon
// Notice: It appears that Unmqr does not write a zero into *info upon
// success (probably a bug), so we simply re-use the info array already
// zeroed by Geqrf above.
OP_REQUIRES_OK_ASYNC(
context,
solver.Ormqr(CUBLAS_SIDE_LEFT, CublasAdjointOp<Scalar>(), m, m,
solver.Unmqr(CUBLAS_SIDE_LEFT, CublasAdjointOp<Scalar>(), m, m,
min_size, &input_transposed_reshaped(batch, 0, 0), m,
&tau_matrix(batch, 0), &q_reshaped(batch, 0, 0), m,
dev_info.back().mutable_data() + batch),
@ -266,12 +266,12 @@ class QrOpGpu : public AsyncOpKernel {
}
} else {
// Generate m x n matrix Q. In this case we can use the more efficient
// algorithm in Orgqr to generate Q in place.
// algorithm in Ungqr to generate Q in place.
dev_info.emplace_back(context, batch_size, "orgqr");
for (int batch = 0; batch < batch_size; ++batch) {
OP_REQUIRES_OK_ASYNC(
context,
solver.Orgqr(
solver.Ungqr(
m, n, min_size, &input_transposed_reshaped(batch, 0, 0), m,
&tau_matrix(batch, 0), dev_info.back().mutable_data() + batch),
done);

194
tensorflow/core/kernels/self_adjoint_eig_v2_op_gpu.cc Normal file
View File

@ -0,0 +1,194 @@
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
// See docs in ../ops/linalg_ops.cc.
#if GOOGLE_CUDA
#include <numeric>
#include <type_traits>
#define EIGEN_USE_GPU
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/kernel_def_builder.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/kernels/cast_op.h"
#include "tensorflow/core/kernels/cuda_solvers.h"
#include "tensorflow/core/kernels/cwise_ops.h"
#include "tensorflow/core/kernels/transpose_functor.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/types.h"
namespace tensorflow {
typedef Eigen::GpuDevice GPUDevice;
template <class Scalar>
class SelfAdjointEigV2OpGpu : public AsyncOpKernel {
public:
explicit SelfAdjointEigV2OpGpu(OpKernelConstruction* context)
: AsyncOpKernel(context) {
OP_REQUIRES_OK(context, context->GetAttr("compute_v", &compute_v_));
}
void ComputeAsync(OpKernelContext* context, DoneCallback done) final {
const Tensor& input = context->input(0);
const int ndims = input.dims();
OP_REQUIRES_ASYNC(
context, ndims >= 2,
errors::InvalidArgument("Input must have rank >= 2, got ", ndims),
done);
const int64 n = input.dim_size(ndims - 1);
OP_REQUIRES_ASYNC(
context, input.dim_size(ndims - 2) == n,
errors::InvalidArgument("Input matrices must be squares, got",
input.dim_size(ndims - 2), " != ", n),
done);
const int64 batch_size =
input.template flat_inner_dims<Scalar, 3>().dimension(0);
// Allocate outputs.
Tensor* eigenvalues;
TensorShape eigenvalues_shape = input.shape();
eigenvalues_shape.RemoveLastDims(1);
OP_REQUIRES_OK_ASYNC(
context, context->allocate_output(0, eigenvalues_shape, &eigenvalues),
done);
Tensor* eigenvectors;
TensorShape eigenvectors_shape =
compute_v_ ? input.shape() : TensorShape({});
OP_REQUIRES_OK_ASYNC(
context, context->allocate_output(1, eigenvectors_shape, &eigenvectors),
done);
if (input.NumElements() == 0) {
done();
return;
}
// Allocate workspace.
Tensor eigenvalues_real;
using RealScalar = typename Eigen::NumTraits<Scalar>::Real;
if (std::is_same<Scalar, RealScalar>::value) {
eigenvalues_real = *eigenvalues;
} else {
OP_REQUIRES_OK_ASYNC(
context,
context->allocate_temp(DataTypeToEnum<RealScalar>::value,
eigenvalues_shape, &eigenvalues_real),
done);
}
Tensor input_copy;
OP_REQUIRES_OK_ASYNC(
context,
context->forward_input_or_allocate_temp(
{0}, DataTypeToEnum<Scalar>::value, input.shape(), &input_copy),
done);
// For real symmetric matrices, row-major and column-major are the same. For
// complex Hermitian, row-major and column-major differ by a conjugation,
// which is still cheaper than a transpose.
const GPUDevice& device = context->eigen_device<GPUDevice>();
if (!input.SharesBufferWith(input_copy)) {
if (Eigen::NumTraits<Scalar>::IsComplex) {
functor::UnaryFunctor<GPUDevice, functor::conj<Scalar>> conj;
conj(device, input_copy.flat<Scalar>() /*out*/,
input.flat<Scalar>() /*in*/);
} else {
device.memcpy(input_copy.flat<Scalar>().data(),
input.flat<Scalar>().data(),
input.NumElements() * sizeof(Scalar));
}
} else if (Eigen::NumTraits<Scalar>::IsComplex) {
functor::UnaryFunctor<GPUDevice, functor::conj<Scalar>> conj;
conj(device, const_cast<Tensor*>(&input)->flat<Scalar>() /*out*/,
input.flat<Scalar>() /*in*/);
}
// Compute eigen decomposition in-place in input_copy.
CudaSolver solver(context);
std::vector<DeviceLapackInfo> dev_info;
dev_info.emplace_back(context, batch_size, "heevd");
auto input_copy_reshaped = input_copy.flat_inner_dims<Scalar, 3>();
auto eigenvalues_real_reshaped =
eigenvalues_real.flat_inner_dims<RealScalar, 2>();
for (int batch = 0; batch < batch_size; ++batch) {
OP_REQUIRES_OK_ASYNC(context,
solver.Heevd(compute_v_ ? CUSOLVER_EIG_MODE_VECTOR
: CUSOLVER_EIG_MODE_NOVECTOR,
CUBLAS_FILL_MODE_UPPER, n,
&input_copy_reshaped(batch, 0, 0), n,
&eigenvalues_real_reshaped(batch, 0),
dev_info.back().mutable_data() + batch),
done);
}
if (!std::is_same<Scalar, RealScalar>::value) {
functor::CastFunctor<GPUDevice, Scalar, RealScalar> cast;
cast(device, eigenvalues->flat<Scalar>(),
const_cast<const Tensor*>(&eigenvalues_real)->flat<RealScalar>());
}
if (compute_v_) {
// Transpose eigenvectors now stored in input_copy in column-major form to
// output in row-major form.
std::vector<int> perm(ndims);
std::iota(perm.begin(), perm.end(), 0);
std::swap(perm[ndims - 2], perm[ndims - 1]);
OP_REQUIRES_OK_ASYNC(
context, DoTranspose(device, input_copy, perm, eigenvectors), done);
}
// Asynchronously check return status from cuSolver kernels.
TensorReference input_copy_ref(input_copy);
TensorReference eigenvalues_real_ref(eigenvalues_real);
auto info_checker = [context, dev_info, input_copy_ref,
eigenvalues_real_ref,
done](const Status& status,
const std::vector<HostLapackInfo>& host_infos) {
input_copy_ref.Unref();
eigenvalues_real_ref.Unref();
OP_REQUIRES_OK_ASYNC(context, status, done);
done();
};
OP_REQUIRES_OK_ASYNC(
context,
solver.CopyLapackInfoToHostAsync(dev_info, std::move(info_checker)),
done);
}
private:
bool compute_v_;
TF_DISALLOW_COPY_AND_ASSIGN(SelfAdjointEigV2OpGpu);
};
#define REGISTER(Scalar) \
REGISTER_KERNEL_BUILDER( \
Name("SelfAdjointEigV2").Device(DEVICE_GPU).TypeConstraint<Scalar>("T"), \
(SelfAdjointEigV2OpGpu<Scalar>))
REGISTER(float);
REGISTER(double);
REGISTER(complex64);
REGISTER(complex128);
#undef REGISTER
} // namespace tensorflow
#endif // GOOGLE_CUDA

2
tensorflow/core/lib/gtl/flatmap.h
View File

@ -146,8 +146,8 @@ class FlatMap {
friend class FlatMap;
Bucket* b_;
Bucket* end_;
char space_ alignas(value_type)[sizeof(value_type)];
uint32 i_;
char space_[sizeof(value_type)];
pointer val() { return reinterpret_cast<pointer>(space_); }
void FillValue() { new (space_) value_type(b_->key(i_), b_->val(i_)); }

3
tensorflow/core/platform/file_system.cc
View File

@ -56,6 +56,9 @@ void ForEach(int first, int last, const std::function<void(int)>& f) {
FileSystem::~FileSystem() {}
string FileSystem::TranslateName(const string& name) const {
// If the name is empty, CleanPath returns "." which is incorrect and
// we should return the empty path instead.
if (name.empty()) return name;
return io::CleanPath(name);
}

1
tensorflow/core/util/ctc/BUILD
View File

@ -62,7 +62,6 @@ cc_library(
],
deps = [
":ctc_loss_util_lib",
"//tensorflow/core:gpu_runtime",
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
"//third_party/eigen3",

2
tensorflow/docs_src/get_started/summaries_and_tensorboard.md
View File

@ -198,7 +198,7 @@ You're now all set to visualize this data using TensorBoard.
## Launching TensorBoard
To run TensorBoard, use the following command (alternatively `python -m
tensorflow.tensorboard`)
tensorboard.main`)
```bash
tensorboard --logdir=path/to/log-directory

14
tensorflow/python/BUILD
View File

@ -50,7 +50,6 @@ py_library(
"//tensorflow/tools/quantization:__pkg__", # TODO(b/34059704): remove when fixed
],
deps = [
":tf_optimizer",
":array_ops",
":bitwise_ops",
":check_ops",
@ -63,15 +62,20 @@ py_library(
":framework_for_generated_wrappers",
":functional_ops",
":gradient_checker",
":graph_util",
":histogram_ops",
":image_ops",
":initializers_ns",
":io_ops",
":layers",
":lib",
":linalg_ns",
":math_ops",
":metrics",
":nn",
":ops",
":platform",
":pywrap_tensorflow",
":script_ops",
":session_ops",
":sets",
@ -81,24 +85,24 @@ py_library(
":state_ops",
":string_ops",
":summary",
":metrics",
":layers",
":tensor_array_ops",
":training",
":ops",
":saver_test_utils",
":subscribe",
":test_ops", # TODO: Break testing code out into separate rule.
":tf_optimizer",
":util",
":weights_broadcast_ops",
"//third_party/py/numpy",
"//tensorflow/core:protos_all_py",
"//tensorflow/python/data",
"//tensorflow/python/estimator:estimator_py",
"//tensorflow/python/feature_column:feature_column_py",
"//tensorflow/python/keras",
"//tensorflow/python/ops/losses",
"//tensorflow/python/ops/distributions",
"//tensorflow/python/profiler",
"//tensorflow/python/saved_model",
"//tensorflow/python/keras",
] + if_not_windows([
"//tensorflow/contrib:contrib_py",
]),

27
tensorflow/python/data/BUILD Normal file
View File

@ -0,0 +1,27 @@
package(default_visibility = ["//tensorflow:internal"])
licenses(["notice"]) # Apache 2.0
exports_files(["LICENSE"])
py_library(
name = "data",
srcs = ["__init__.py"],
srcs_version = "PY2AND3",
deps = [
"//tensorflow/python:util",
"//tensorflow/python/data/ops:dataset_ops",
],
)
filegroup(
name = "all_files",
srcs = glob(
["**/*"],
exclude = [
"**/METADATA",
"**/OWNERS",
],
),
visibility = ["//tensorflow:__subpackages__"],
)

37
tensorflow/python/data/__init__.py Normal file
View File

@ -0,0 +1,37 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""`tf.data.Dataset` API for input pipelines.
@@Dataset
@@Iterator
@@TFRecordDataset
@@FixedLengthRecordDataset
@@TextLineDataset
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
# pylint: disable=unused-import
from tensorflow.python.data.ops.dataset_ops import Dataset
from tensorflow.python.data.ops.dataset_ops import FixedLengthRecordDataset
from tensorflow.python.data.ops.dataset_ops import Iterator
from tensorflow.python.data.ops.dataset_ops import TextLineDataset
from tensorflow.python.data.ops.dataset_ops import TFRecordDataset
# pylint: enable=unused-import
from tensorflow.python.util.all_util import remove_undocumented
remove_undocumented(__name__)

38
tensorflow/python/data/ops/BUILD Normal file
View File

@ -0,0 +1,38 @@
package(default_visibility = ["//tensorflow:internal"])
licenses(["notice"]) # Apache 2.0
exports_files(["LICENSE"])
py_library(
name = "dataset_ops",
srcs = ["dataset_ops.py"],
srcs_version = "PY2AND3",
deps = [
"//tensorflow/python:constant_op",
"//tensorflow/python:dataset_ops_gen",
"//tensorflow/python:dtypes",
"//tensorflow/python:framework_ops",
"//tensorflow/python:function",
"//tensorflow/python:math_ops",
"//tensorflow/python:random_seed",
"//tensorflow/python:script_ops",
"//tensorflow/python:sparse_tensor",
"//tensorflow/python:tensor_shape",
"//tensorflow/python:tensor_util",
"//tensorflow/python/data/util:nest",
"//third_party/py/numpy",
],
)
filegroup(
name = "all_files",
srcs = glob(
["**/*"],
exclude = [
"**/METADATA",
"**/OWNERS",
],
),
visibility = ["//tensorflow:__subpackages__"],
)

2032
tensorflow/python/data/ops/dataset_ops.py Normal file
View File

File diff suppressed because it is too large Load Diff

0
tensorflow/contrib/data/python/util/BUILD → tensorflow/python/data/util/BUILD
View File

1
tensorflow/contrib/data/python/util/nest.py → tensorflow/python/data/util/nest.py
View File

@ -13,6 +13,7 @@
# limitations under the License.
# ==============================================================================
# TODO(shivaniagrawal): Merge with core nest
"""## Functions for working with arbitrarily nested sequences of elements.
NOTE(mrry): This fork of the `tensorflow.python.util.nest` module

2
tensorflow/contrib/data/python/util/nest_test.py → tensorflow/python/data/util/nest_test.py
View File

@ -22,7 +22,7 @@ import collections
import numpy as np
from tensorflow.contrib.data.python.util import nest
from tensorflow.python.data.util import nest
from tensorflow.python.framework import constant_op
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops

49
tensorflow/python/estimator/training.py
View File

@ -203,6 +203,22 @@ class EvalSpec(
throttle_secs=throttle_secs)
class _StopAtSecsHook(session_run_hook.SessionRunHook):
"""Stops given secs after begin is called."""
def __init__(self, stop_after_secs):
self._stop_after_secs = stop_after_secs
self._start_time = None
def begin(self):
self._start_time = time.time()
def after_run(self, run_context, run_values):
del run_values
if time.time() - self._start_time >= self._stop_after_secs:
run_context.request_stop()
class UnimplementedError(Exception):
pass
@ -254,7 +270,38 @@ class _TrainingExecutor(object):
def run_local(self):
"""Runs training and evaluation locally (non-distributed)."""
raise UnimplementedError('Method run_local has not been implemented.')
def _should_stop_local_train(global_step):
if self._train_spec.max_steps is None:
return False
if global_step >= self._train_spec.max_steps:
return True
return False
if self._eval_spec.throttle_secs <= 0:
raise ValueError('eval_spec.throttle_secs should be positive, given: {}.'
'It is used do determine how long each training '
'iteration should go when train and evaluate '
'locally.'.format(
self._eval_spec.throttle_secs))
stop_hook = _StopAtSecsHook(self._eval_spec.throttle_secs)
train_hooks = list(self._train_spec.hooks) + [stop_hook]
logging.info('Start train and evaluate loop. The evaluate will happen '
'after {} secs (eval_spec.throttle_secs) or training is '
'finished.'.format(self._eval_spec.throttle_secs))
while True:
self._estimator.train(
input_fn=self._train_spec.input_fn,
max_steps=self._train_spec.max_steps,
hooks=train_hooks)
metrics = self._estimator.evaluate(
input_fn=self._eval_spec.input_fn,
steps=self._eval_spec.steps,
hooks=self._eval_spec.hooks,
name=self._eval_spec.name)
if _should_stop_local_train(metrics[ops.GraphKeys.GLOBAL_STEP]):
break
def _start_std_server(self, config):
"""Creates, starts, and returns a server_lib.Server."""

99
tensorflow/python/estimator/training_test.py
View File

@ -27,8 +27,10 @@ from tensorflow.python.estimator import estimator as estimator_lib
from tensorflow.python.estimator import run_config as run_config_lib
from tensorflow.python.estimator import training
from tensorflow.python.framework import ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.platform import test
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.training import monitored_session
from tensorflow.python.training import saver
from tensorflow.python.training import server_lib
from tensorflow.python.training import session_run_hook
@ -614,5 +616,102 @@ class TrainingExecutorRunPsTest(test.TestCase):
mock_eval_spec).run_ps()
class StopAtSecsHookTest(test.TestCase):
"""Tests StopAtSecsHook."""
@test.mock.patch.object(time, 'time')
def test_stops_after_time(self, mock_time):
mock_time.return_value = 1484695987.209386
hook = training._StopAtSecsHook(1000)
with ops.Graph().as_default():
no_op = control_flow_ops.no_op()
# some time passed before training starts
mock_time.return_value += 250
with monitored_session.MonitoredSession(hooks=[hook]) as sess:
self.assertFalse(sess.should_stop())
sess.run(no_op)
self.assertFalse(sess.should_stop())
mock_time.return_value += 500
sess.run(no_op)
self.assertFalse(sess.should_stop())
mock_time.return_value += 400
sess.run(no_op)
self.assertFalse(sess.should_stop())
mock_time.return_value += 200
sess.run(no_op)
self.assertTrue(sess.should_stop())
class TrainingExecutorRunLocalTest(test.TestCase):
"""Tests run_local of _TrainingExecutor."""
def test_send_stop_at_secs_to_train(self):
mock_est = test.mock.Mock(spec=estimator_lib.Estimator)
train_spec = training.TrainSpec(
input_fn=lambda: 1, max_steps=2, hooks=[_FakeHook()])
eval_spec = training.EvalSpec(
input_fn=lambda: 1, hooks=[_FakeHook()], throttle_secs=100)
mock_est.evaluate.return_value = {_GLOBAL_STEP_KEY: train_spec.max_steps}
executor = training._TrainingExecutor(mock_est, train_spec, eval_spec)
executor.run_local()
stop_hook = mock_est.train.call_args[1]['hooks'][-1]
self.assertIsInstance(stop_hook, training._StopAtSecsHook)
self.assertEqual(eval_spec.throttle_secs, stop_hook._stop_after_secs)
def test_runs_in_a_loop_until_max_steps(self):
mock_est = test.mock.Mock(spec=estimator_lib.Estimator)
train_spec = training.TrainSpec(
input_fn=lambda: 1, max_steps=300, hooks=[_FakeHook()])
eval_spec = training.EvalSpec(
input_fn=lambda: 1, hooks=[_FakeHook()], throttle_secs=100)
# should be called 3 times.
mock_est.evaluate.side_effect = [{
_GLOBAL_STEP_KEY: train_spec.max_steps - 100
}, {
_GLOBAL_STEP_KEY: train_spec.max_steps - 50
}, {
_GLOBAL_STEP_KEY: train_spec.max_steps
}]
executor = training._TrainingExecutor(mock_est, train_spec, eval_spec)
executor.run_local()
self.assertEqual(3, mock_est.train.call_count)
self.assertEqual(3, mock_est.evaluate.call_count)
def test_train_and_evaluate_args(self):
mock_est = test.mock.Mock(spec=estimator_lib.Estimator)
train_spec = training.TrainSpec(
input_fn=lambda: 1, max_steps=300, hooks=[_FakeHook()])
eval_spec = training.EvalSpec(
input_fn=lambda: 1, steps=2, hooks=[_FakeHook()], name='local_eval')
mock_est.evaluate.return_value = {_GLOBAL_STEP_KEY: train_spec.max_steps}
executor = training._TrainingExecutor(mock_est, train_spec, eval_spec)
executor.run_local()
mock_est.evaluate.assert_called_with(
name=eval_spec.name,
input_fn=eval_spec.input_fn,
steps=eval_spec.steps,
hooks=eval_spec.hooks)
train_args = mock_est.train.call_args[1]
self.assertEqual(list(train_spec.hooks), list(train_args['hooks'][:-1]))
self.assertEqual(train_spec.input_fn, train_args['input_fn'])
self.assertEqual(train_spec.max_steps, train_args['max_steps'])
def test_errors_out_if_throttle_secs_is_zero(self):
mock_est = test.mock.Mock(spec=estimator_lib.Estimator)
train_spec = training.TrainSpec(input_fn=lambda: 1)
eval_spec = training.EvalSpec(input_fn=lambda: 1, throttle_secs=0)
executor = training._TrainingExecutor(mock_est, train_spec, eval_spec)
with self.assertRaisesRegexp(ValueError, 'throttle_secs'):
executor.run_local()
if __name__ == '__main__':
test.main()

22
tensorflow/python/keras/BUILD
View File

@ -35,6 +35,7 @@ py_library(
"_impl/keras/engine/__init__.py",
"_impl/keras/engine/topology.py",
"_impl/keras/engine/training.py",
"_impl/keras/estimator.py",
"_impl/keras/initializers.py",
"_impl/keras/layers/__init__.py",
"_impl/keras/layers/advanced_activations.py",
@ -88,6 +89,7 @@ py_library(
"datasets/imdb/__init__.py",
"datasets/mnist/__init__.py",
"datasets/reuters/__init__.py",
"estimator/__init__.py",
"initializers/__init__.py",
"layers/__init__.py",
"losses/__init__.py",
@ -125,9 +127,11 @@ py_library(
"//tensorflow/python:layers_base",
"//tensorflow/python:logging_ops",
"//tensorflow/python:math_ops",
"//tensorflow/python:metrics",
"//tensorflow/python:nn",
"//tensorflow/python:platform",
"//tensorflow/python:random_ops",
"//tensorflow/python:session",
"//tensorflow/python:sparse_ops",
"//tensorflow/python:sparse_tensor",
"//tensorflow/python:state_ops",
@ -139,6 +143,8 @@ py_library(
"//tensorflow/python:util",
"//tensorflow/python:variable_scope",
"//tensorflow/python:variables",
"//tensorflow/python/estimator",
"//tensorflow/python/estimator:model_fn",
"@six_archive//:six",
],
)
@ -656,6 +662,22 @@ py_test(
],
)
py_test(
name = "estimator_test",
size = "medium",
srcs = ["_impl/keras/estimator_test.py"],
srcs_version = "PY2AND3",
deps = [
":keras",
"//tensorflow/python:client_testlib",
"//tensorflow/python:dtypes",
"//tensorflow/python:framework_ops",
"//tensorflow/python:platform",
"//tensorflow/python/estimator:numpy_io",
"//third_party/py/numpy",
],
)
py_test(
name = "backend_test",
size = "small",

1
tensorflow/python/keras/__init__.py
View File

@ -29,6 +29,7 @@ from tensorflow.python.keras import backend
from tensorflow.python.keras import callbacks
from tensorflow.python.keras import constraints
from tensorflow.python.keras import datasets
from tensorflow.python.keras import estimator
from tensorflow.python.keras import initializers
from tensorflow.python.keras import layers
from tensorflow.python.keras import losses

1
tensorflow/python/keras/_impl/keras/__init__.py
View File

@ -25,6 +25,7 @@ from tensorflow.python.keras._impl.keras import callbacks
from tensorflow.python.keras._impl.keras import constraints
from tensorflow.python.keras._impl.keras import datasets
from tensorflow.python.keras._impl.keras import engine
from tensorflow.python.keras._impl.keras import estimator
from tensorflow.python.keras._impl.keras import initializers
from tensorflow.python.keras._impl.keras import layers
from tensorflow.python.keras._impl.keras import losses

2
tensorflow/python/keras/_impl/keras/applications/mobilenet.py
View File

@ -57,7 +57,7 @@ the 100 % MobileNet on various input sizes:
The weights for all 16 models are obtained and translated
from Tensorflow checkpoints found at
https://github.com/tensorflow/models/blob/master/slim/nets/mobilenet_v1.md
https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md
# Reference
- [MobileNets: Efficient Convolutional Neural Networks for

37
tensorflow/python/keras/_impl/keras/backend.py
View File

@ -373,22 +373,7 @@ def get_session():
session = _SESSION
if not _MANUAL_VAR_INIT:
with session.graph.as_default():
variables = variables_module.global_variables()
candidate_vars = []
for v in variables:
if not getattr(v, '_keras_initialized', False):
candidate_vars.append(v)
# This step is expensive, so we only run it on variables not already
# marked as initialized.
is_initialized = session.run(
[variables_module.is_variable_initialized(v) for v in candidate_vars])
uninitialized_vars = []
for flag, v in zip(is_initialized, candidate_vars):
if not flag:
uninitialized_vars.append(v)
v._keras_initialized = True
if uninitialized_vars:
session.run(variables_module.variables_initializer(uninitialized_vars))
_initialize_variables(session)
return session
@ -556,6 +541,26 @@ def variable(value, dtype=None, name=None, constraint=None):
return v
def _initialize_variables(session):
"""Utility to initialize uninitialized variables on the fly."""
variables = variables_module.global_variables()
candidate_vars = []
for v in variables:
if not getattr(v, '_keras_initialized', False):
candidate_vars.append(v)
# This step is expensive, so we only run it on variables not already
# marked as initialized.
is_initialized = session.run(
[variables_module.is_variable_initialized(v) for v in candidate_vars])
uninitialized_vars = []
for flag, v in zip(is_initialized, candidate_vars):
if not flag:
uninitialized_vars.append(v)
v._keras_initialized = True
if uninitialized_vars:
session.run(variables_module.variables_initializer(uninitialized_vars))
def constant(value, dtype=None, shape=None, name=None):
"""Creates a constant tensor.

281
tensorflow/python/keras/_impl/keras/estimator.py Normal file
View File

@ -0,0 +1,281 @@
# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
# pylint: disable=protected-access
"""Home of estimator related functions.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.client import session
from tensorflow.python.estimator import estimator as estimator_lib
from tensorflow.python.estimator import model_fn as model_fn_lib
from tensorflow.python.framework import ops
from tensorflow.python.framework import random_seed
from tensorflow.python.framework import sparse_tensor as sparse_tensor_lib
from tensorflow.python.keras._impl.keras import backend as K
from tensorflow.python.keras._impl.keras import models
from tensorflow.python.keras._impl.keras.utils.generic_utils import CustomObjectScope
from tensorflow.python.ops import metrics as metrics_module
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.training import saver as saver_lib
from tensorflow.python.training import training_util
def _create_ordered_io(keras_model, estimator_io_dict, is_input=True):
"""Create a list of tensors from IO dictionary based on Keras IO order.
Args:
keras_model: an instance of compiled keras model.
estimator_io_dict: features or labels dictionary from model_fn.
is_input: True if dictionary is for inputs.
Returns:
a list of tensors based on Keras IO order.
Raises:
ValueError: if dictionary keys cannot be found in Keras model input_names
or output_names.
"""
if is_input:
keras_io_names = keras_model.input_names
else:
keras_io_names = keras_model.output_names
for key in estimator_io_dict:
if key not in keras_io_names:
raise ValueError(
'Cannot find %s with name "%s" in Keras Model. It needs to match '
'one of the following: %s' % ('input' if is_input else 'output', key,
', '.join(keras_io_names)))
tensors = []
for io_name in keras_io_names:
tensors.append(estimator_io_dict[io_name])
return tensors
def _clone_and_build_model(mode,
keras_model,
custom_objects,
features=None,
labels=None):
"""Clone and build the given keras_model.
Args:
mode: training mode.
keras_model: an instance of compiled keras model.
custom_objects: Dictionary for custom objects.
features:
labels:
Returns:
The newly built model.
"""
# Set to True during training, False for inference.
K.set_learning_phase(mode == model_fn_lib.ModeKeys.TRAIN)
# Clone keras model.
input_tensors = None if features is None else _create_ordered_io(
keras_model, features)
if custom_objects:
with CustomObjectScope(custom_objects):
model = models.clone_model(keras_model, input_tensors=input_tensors)
else:
model = models.clone_model(keras_model, input_tensors=input_tensors)
# Compile/Build model
if mode is model_fn_lib.ModeKeys.PREDICT and not model.built:
model.build()
else:
optimizer_config = keras_model.optimizer.get_config()
optimizer = keras_model.optimizer.__class__.from_config(optimizer_config)
optimizer.iterations = training_util.get_or_create_global_step()
# Get list of outputs.
if labels is None:
target_tensors = None
elif isinstance(labels, dict):
target_tensors = _create_ordered_io(keras_model, labels, is_input=False)
else:
target_tensors = [
sparse_tensor_lib.convert_to_tensor_or_sparse_tensor(labels)
]
model.compile(
optimizer,
keras_model.loss,
metrics=keras_model.metrics,
loss_weights=keras_model.loss_weights,
sample_weight_mode=keras_model.sample_weight_mode,
weighted_metrics=keras_model.weighted_metrics,
target_tensors=target_tensors)
if isinstance(model, models.Sequential):
model = model.model
return model
def _create_keras_model_fn(keras_model, custom_objects=None):
"""Creates model_fn for keras Estimator.
Args:
keras_model: an instance of compiled keras model.
custom_objects: Dictionary for custom objects.
Returns:
The model_fn for a keras Estimator.
"""
def model_fn(features, labels, mode):
"""model_fn for keras Estimator."""
model = _clone_and_build_model(mode, keras_model, custom_objects, features,
labels)
# Get inputs to EstimatorSpec
predictions = dict(zip(model.output_names, model.outputs))
loss = None
train_op = None
eval_metric_ops = None
# Set loss and metric only during train and evaluate.
if mode is not model_fn_lib.ModeKeys.PREDICT:
model._make_train_function() # pylint: disable=protected-access
loss = model.total_loss
if model.metrics:
eval_metric_ops = {}
# When each metric maps to an output
if isinstance(model.metrics, dict):
for i, output_name in enumerate(model.metrics.keys()):
metric_name = model.metrics[output_name]
if callable(metric_name):
metric_name = metric_name.__name__
# When some outputs use the same metric
if list(model.metrics.values()).count(metric_name) > 1:
metric_name += '_' + output_name
eval_metric_ops[metric_name] = metrics_module.mean(
model.metrics_tensors[i - len(model.metrics)])
else:
for i, metric_name in enumerate(model.metrics):
if callable(metric_name):
metric_name = metric_name.__name__
eval_metric_ops[metric_name] = metrics_module.mean(
model.metrics_tensors[i])
# Set train_op only during train.
if mode is model_fn_lib.ModeKeys.TRAIN:
train_op = model.train_function.updates_op
return model_fn_lib.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
return model_fn
def _save_first_checkpoint(keras_model, estimator, custom_objects,
keras_weights):
"""Save first checkpoint for the keras Estimator.
Args:
keras_model: an instance of compiled keras model.
estimator: keras estimator.
custom_objects: Dictionary for custom objects.
keras_weights: A flat list of Numpy arrays for weights of given keras_model.
Returns:
The model_fn for a keras Estimator.
"""
with ops.Graph().as_default() as g, g.device(estimator._device_fn):
random_seed.set_random_seed(estimator.config.tf_random_seed)
training_util.create_global_step()
model = _clone_and_build_model(model_fn_lib.ModeKeys.TRAIN, keras_model,
custom_objects)
if isinstance(model, models.Sequential):
model = model.model
# Load weights and save to checkpoint if there is no checkpoint
latest_path = saver_lib.latest_checkpoint(estimator.model_dir)
if not latest_path:
with session.Session() as sess:
model.set_weights(keras_weights)
# Make update ops and initialize all variables.
if not model.train_function:
# pylint: disable=protected-access
model._make_train_function()
K._initialize_variables(sess)
# pylint: enable=protected-access
saver = saver_lib.Saver()
saver.save(sess, estimator.model_dir + '/')
def model_to_estimator(keras_model=None,
keras_model_path=None,
custom_objects=None,
model_dir=None,
config=None):
"""Constructs an `Estimator` instance from given keras model.
Args:
keras_model: Keras model in memory.
keras_model_path: Directory to a keras model on disk.
custom_objects: Dictionary for custom objects.
model_dir: Directory to save Estimator model parameters, graph and etc.
config: Configuration object.
Returns:
An Estimator from given keras model.
Raises:
ValueError: if neither keras_model nor keras_model_path was given.
ValueError: if both keras_model and keras_model_path was given.
ValueError: if the keras_model_path is a GCS URI.
ValueError: if keras_model has not been compiled.
"""
if (not keras_model) and (not keras_model_path):
raise ValueError(
'Either keras_model or keras_model_path needs to be provided.')
if keras_model and keras_model_path:
raise ValueError(
'Please specity either keras_model or keras_model_path but not both.')
if not keras_model:
if keras_model_path.startswith(
'gs://') or 'storage.googleapis.com' in keras_model_path:
raise ValueError(
'%s is not a local path. Please copy the model locally first.' %
keras_model_path)
logging.info('Loading models from %s', keras_model_path)
keras_model = models.load_model(keras_model_path)
else:
logging.info('Using the Keras model from memory.')
keras_model = keras_model
if not hasattr(keras_model, 'optimizer'):
raise ValueError(
'Given keras model has not been compiled yet. Please compile first '
'before creating the estimator.')
keras_weights = keras_model.get_weights()
keras_model_fn = _create_keras_model_fn(keras_model, custom_objects)
est = estimator_lib.Estimator(
keras_model_fn, model_dir=model_dir, config=config)
# TODO(yifeif): move checkpoint initialization to scaffold.init_fn
_save_first_checkpoint(keras_model, est, custom_objects, keras_weights)
return est

392
tensorflow/python/keras/_impl/keras/estimator_test.py Normal file
View File

@ -0,0 +1,392 @@
# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for training routines."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from math import log10
import os
import tempfile
import numpy as np
from tensorflow.python.estimator.inputs import numpy_io
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.keras._impl import keras
from tensorflow.python.keras._impl.keras import testing_utils
from tensorflow.python.platform import gfile
from tensorflow.python.platform import test
try:
import h5py # pylint:disable=g-import-not-at-top
except ImportError:
h5py = None
def simple_sequential_model():
model = keras.models.Sequential()
model.add(
keras.layers.Conv2D(
32, kernel_size=(3, 3), activation='relu', input_shape=(14, 14, 3)))
model.add(keras.layers.MaxPooling2D(pool_size=(2, 2)))
model.add(keras.layers.Dropout(0.25))
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(16, activation='relu'))
model.add(keras.layers.Dropout(0.25))
model.add(keras.layers.Dense(3, activation='softmax'))
return model
def simple_functional_model():
a = keras.layers.Input(shape=(14, 14, 3))
b = keras.layers.Conv2D(32, kernel_size=(3, 3), activation='relu')(a)
b = keras.layers.MaxPooling2D(pool_size=(2, 2))(b)
b = keras.layers.Dropout(0.25)(b)
b = keras.layers.Flatten()(b)
b = keras.layers.Dense(16, activation='relu')(b)
b = keras.layers.Dropout(0.25)(b)
b = keras.layers.Dense(3, activation='softmax')(b)
model = keras.models.Model(inputs=[a], outputs=[b])
return model
def get_resource_for_simple_model(is_sequential, is_evaluate):
model = simple_sequential_model(
) if is_sequential else simple_functional_model()
if is_sequential:
model.build()
input_name = model.input_names[0]
np.random.seed(1337)
(x_train, y_train), (x_test, y_test) = testing_utils.get_test_data(
train_samples=200,
test_samples=100,
input_shape=(14, 14, 3),
num_classes=3)
y_train = keras.utils.to_categorical(y_train)
y_test = keras.utils.to_categorical(y_test)
train_input_fn = numpy_io.numpy_input_fn(
x={input_name: np.array(x_train, dtype=np.float32)},
y=np.array(y_train, dtype=np.float32),
shuffle=False,
num_epochs=None,
batch_size=16)
evaluate_input_fn = numpy_io.numpy_input_fn(
x={input_name: np.array(x_test, dtype=np.float32)},
y=np.array(y_test, dtype=np.float32),
num_epochs=1,
shuffle=False)
predict_input_fn = numpy_io.numpy_input_fn(
x={input_name: np.array(x_test, dtype=np.float32)},
num_epochs=1,
shuffle=False)
inference_input_fn = evaluate_input_fn if is_evaluate else predict_input_fn
return model, (x_train, y_train), (x_test,
y_test), train_input_fn, inference_input_fn
def multi_inputs_multi_outputs_model():
# test multi-input layer
a = keras.layers.Input(shape=(32,), name='input_a')
b = keras.layers.Input(shape=(32,), name='input_b')
dense = keras.layers.Dense(16, name='dense_1')
a_2 = dense(a)
b_2 = dense(b)
merged = keras.layers.concatenate([a_2, b_2], name='merge')
c = keras.layers.Dense(3, activation='softmax', name='dense_2')(merged)
d = keras.layers.Dense(2, activation='softmax', name='dense_3')(merged)
model = keras.models.Model(inputs=[a, b], outputs=[c, d])
model.compile(
loss='categorical_crossentropy',
optimizer='rmsprop',
metrics={'dense_2': 'accuracy',
'dense_3': 'accuracy'})
return model
class TestKerasEstimator(test.TestCase):
def setUp(self):
self._base_dir = os.path.join(self.get_temp_dir(), 'keras_estimator_test')
gfile.MakeDirs(self._base_dir)
def tearDown(self):
gfile.DeleteRecursively(self._base_dir)
def test_train(self):
for is_sequential in [True, False]:
keras_model, (_, _), (
_, _), train_input_fn, eval_input_fn = get_resource_for_simple_model(
is_sequential=is_sequential, is_evaluate=True)
keras_model.compile(
loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['accuracy', 'mse', keras.metrics.categorical_accuracy])
with self.test_session():
est_keras = keras.estimator.model_to_estimator(
keras_model=keras_model,
model_dir=tempfile.mkdtemp(dir=self._base_dir))
est_keras.train(input_fn=train_input_fn, steps=200 * 10 / 16)
eval_results = est_keras.evaluate(input_fn=eval_input_fn)
self.assertGreater(eval_results['accuracy'], 0.9)
self.assertGreater(eval_results['categorical_accuracy'], 0.9)
self.assertLess(eval_results['mse'], 0.1)
def test_evaluate(self):
keras_model, (x_train, y_train), (
x_test, y_test), _, eval_input_fn = get_resource_for_simple_model(
is_sequential=False, is_evaluate=True)
with self.test_session():
metrics = [
'binary_accuracy', 'binary_crossentropy', 'categorical_accuracy',
'categorical_crossentropy', 'cosine_proximity', 'hinge',
'kullback_leibler_divergence', 'mean_absolute_error',
'mean_absolute_percentage_error', 'mean_squared_error',
'mean_squared_logarithmic_error', 'poisson', 'squared_hinge',
'top_k_categorical_accuracy'
]
keras_model.compile(
loss='categorical_crossentropy', optimizer='adam', metrics=metrics)
keras_model.fit(x_train, y_train, epochs=1)
keras_eval = keras_model.evaluate(x_test, y_test, batch_size=32)
with self.test_session():
keras_est = keras.estimator.model_to_estimator(
keras_model=keras_model,
model_dir=tempfile.mkdtemp(dir=self._base_dir))
est_eval = keras_est.evaluate(input_fn=eval_input_fn)
metrics = ['loss'] + metrics
# Check loss and all metrics match between keras and estimator.
def shift(val):
return val / 10**int(log10(abs(val)))
for i, metric_name in enumerate(metrics):
self.assertAlmostEqual(
shift(est_eval[metric_name]),
shift(keras_eval[i]),
places=4,
msg='%s mismatch, keras model: %s, estimator: %s' %
(metric_name, est_eval[metric_name], keras_eval[i]))
def test_predict(self):
# Check that predict on a pretrained model yield the same result.
keras_model, (x_train, y_train), (
x_test, _), _, pred_input_fn = get_resource_for_simple_model(
is_sequential=True, is_evaluate=False)
with self.test_session():
keras_model.compile(
loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
keras_model.fit(x_train, y_train, epochs=1)
keras_pred = [np.argmax(y) for y in keras_model.predict(x_test)]
with self.test_session():
keras_est = keras.estimator.model_to_estimator(
keras_model=keras_model,
model_dir=tempfile.mkdtemp(dir=self._base_dir))
est_pred = [
np.argmax(y[keras_model.output_names[0]])
for y in keras_est.predict(input_fn=pred_input_fn)
]
self.assertAllEqual(est_pred, keras_pred)
def test_multi_inputs_multi_outputs(self):
np.random.seed(1337)
(a_train, c_train), (a_test, c_test) = testing_utils.get_test_data(
train_samples=200, test_samples=100, input_shape=(32,), num_classes=3)
(b_train, d_train), (b_test, d_test) = testing_utils.get_test_data(
train_samples=200, test_samples=100, input_shape=(32,), num_classes=2)
c_train = keras.utils.to_categorical(c_train)
c_test = keras.utils.to_categorical(c_test)
d_train = keras.utils.to_categorical(d_train)
d_test = keras.utils.to_categorical(d_test)
def train_input_fn():
input_dict = {
'input_a':
ops.convert_to_tensor(
np.array(a_train, dtype=np.float32), dtype=dtypes.float32),
'input_b':
ops.convert_to_tensor(
np.array(b_train, dtype=np.float32), dtype=dtypes.float32)
}
output_dict = {
'dense_2':
ops.convert_to_tensor(
np.array(c_train, dtype=np.float32), dtype=dtypes.float32),
'dense_3':
ops.convert_to_tensor(
np.array(d_train, dtype=np.float32), dtype=dtypes.float32)
}
return input_dict, output_dict
def evaluate_input_fn():
input_dict = {
'input_a':
ops.convert_to_tensor(
np.array(a_test, dtype=np.float32), dtype=dtypes.float32),
'input_b':
ops.convert_to_tensor(
np.array(b_test, dtype=np.float32), dtype=dtypes.float32)
}
output_dict = {
'dense_2':
ops.convert_to_tensor(
np.array(c_test, dtype=np.float32), dtype=dtypes.float32),
'dense_3':
ops.convert_to_tensor(
np.array(d_test, dtype=np.float32), dtype=dtypes.float32)
}
return input_dict, output_dict
with self.test_session():
model = multi_inputs_multi_outputs_model()
est_keras = keras.estimator.model_to_estimator(
keras_model=model, model_dir=tempfile.mkdtemp(dir=self._base_dir))
est_keras.train(input_fn=train_input_fn, steps=200 * 10 / 16)
eval_results = est_keras.evaluate(input_fn=evaluate_input_fn, steps=1)
self.assertGreater(eval_results['accuracy_dense_2'], 0.5)
self.assertGreater(eval_results['accuracy_dense_3'], 0.5)
def test_init_from_file(self):
if h5py is None:
return # Skip test if models cannot be saved.
keras_model, (x_train, y_train), (
x_test, _), _, pred_input_fn = get_resource_for_simple_model(
is_sequential=False, is_evaluate=False)
with self.test_session():
keras_model.compile(
loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])
keras_model.fit(x_train, y_train, epochs=1)
keras_pred = [np.argmax(y) for y in keras_model.predict(x_test)]
fname = os.path.join(self._base_dir, 'keras_model.h5')
keras.models.save_model(keras_model, fname)
with self.test_session():
keras_est = keras.estimator.model_to_estimator(
keras_model_path=fname,
model_dir=tempfile.mkdtemp(dir=self._base_dir))
est_pred = [
np.argmax(y[keras_model.output_names[0]])
for y in keras_est.predict(input_fn=pred_input_fn)
]
self.assertAllEqual(est_pred, keras_pred)
def test_keras_model_init_error(self):
with self.assertRaisesRegexp(ValueError, 'Either'):
keras.estimator.model_to_estimator()
with self.test_session():
keras_model = simple_sequential_model()
with self.assertRaisesRegexp(ValueError, 'not both'):
keras.estimator.model_to_estimator(
keras_model=keras_model,
keras_model_path=tempfile.mkdtemp(dir=self._base_dir))
with self.test_session():
keras_model = simple_sequential_model()
with self.assertRaisesRegexp(ValueError, 'compiled'):
keras.estimator.model_to_estimator(keras_model=keras_model)
with self.test_session():
keras_model = simple_sequential_model()
with self.assertRaisesRegexp(ValueError, 'not a local path'):
keras.estimator.model_to_estimator(
keras_model_path='gs://bucket/object')
def test_invalid_ionames_error(self):
np.random.seed(1337)
(x_train, y_train), (_, _) = testing_utils.get_test_data(
train_samples=200, test_samples=100, input_shape=(10,), num_classes=2)
y_train = keras.utils.to_categorical(y_train)
def invald_input_name_input_fn():
input_dict = {
'invalid_input_name':
ops.convert_to_tensor(
np.array(x_train, dtype=np.float32), dtype=dtypes.float32),
}
output = ops.convert_to_tensor(
np.array(y_train, dtype=np.float32), dtype=dtypes.float32)
return input_dict, output
def invald_output_name_input_fn():
input_dict = {
'input_1':
ops.convert_to_tensor(
np.array(x_train, dtype=np.float32), dtype=dtypes.float32),
}
output_dict = {
'invalid_output_name':
ops.convert_to_tensor(
np.array(y_train, dtype=np.float32), dtype=dtypes.float32),
}
return input_dict, output_dict
model = simple_functional_model()
model.compile(
loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])
est_keras = keras.estimator.model_to_estimator(
keras_model=model, model_dir=tempfile.mkdtemp(dir=self._base_dir))
with self.test_session():
with self.assertRaises(ValueError):
est_keras.train(input_fn=invald_input_name_input_fn, steps=100)
with self.assertRaises(ValueError):
est_keras.train(input_fn=invald_output_name_input_fn, steps=100)
def test_custom_objects(self):
keras_model, (_, _), (
_, _), train_input_fn, eval_input_fn = get_resource_for_simple_model(
is_sequential=True, is_evaluate=True)
class CustomOp(keras.optimizers.RMSprop):
pass
def custom_loss(y_true, y_pred):
return keras.losses.categorical_crossentropy(y_true, y_pred)
keras_model.compile(
loss=custom_loss, optimizer=CustomOp(), metrics=['accuracy'])
with self.test_session():
est_keras = keras.estimator.model_to_estimator(
keras_model=keras_model,
model_dir=tempfile.mkdtemp(dir=self._base_dir))
est_keras.train(input_fn=train_input_fn, steps=200 * 10 / 16)
eval_results = est_keras.evaluate(input_fn=eval_input_fn)
self.assertGreater(eval_results['accuracy'], 0.9)
if __name__ == '__main__':
test.main()

3
tensorflow/python/keras/_impl/keras/models.py
View File

@ -417,6 +417,9 @@ class Sequential(Model):
name = prefix + str(K.get_uid(prefix))
self.name = name
# Used by Layer base class.
self._dtype = None
# The following properties are not actually used by Keras;
# they exist for compatibility with TF's variable scoping mechanism.
self._updates = []

25
tensorflow/python/keras/estimator/__init__.py Normal file
View File

@ -0,0 +1,25 @@
# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Keras estimator API."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.keras._impl.keras.estimator import model_to_estimator
del absolute_import
del division
del print_function

54
tensorflow/python/kernel_tests/self_adjoint_eig_op_test.py
View File

@ -29,6 +29,13 @@ from tensorflow.python.ops import math_ops
from tensorflow.python.platform import test
def _AddTest(test_class, op_name, testcase_name, fn):
test_name = "_".join(["test", op_name, testcase_name])
if hasattr(test_class, test_name):
raise RuntimeError("Test %s defined more than once" % test_name)
setattr(test_class, test_name, fn)
class SelfAdjointEigTest(test.TestCase):
def testWrongDimensions(self):
@ -50,28 +57,30 @@ def SortEigenDecomposition(e, v):
return np.take(e, perm, -1), np.take(v, perm, -1)
def NormalizeEigenvectorsPhase(v):
"""Normalizes the phase of the Eigenvectors stored in the columns of `v`.
def EquilibrateEigenVectorPhases(x, y):
"""Equilibrate the phase of the Eigenvectors in the columns of `x` and `y`.
(complex) Eigenvectors are only unique up to an arbitrary phase.
We normalize the vectors such that the first component has phase 0.
Eigenvectors are only unique up to an arbitrary phase. This function rotates x
such that it matches y. Precondition: The coluns of x and y differ by a
multiplicative complex phase factor only.
Args:
v: `np.ndarray` with Eigenvectors as returned from `np.linalg.eigh`.
x: `np.ndarray` with Eigenvectors
y: `np.ndarray` with Eigenvectors
Returns:
`np.ndarray` normalized Eigenvectors.
`np.ndarray` containing an equilibrated version of x.
"""
reference = v / np.linalg.norm(v[..., 0:1, :], axis=-1, keepdims=True)
return v * reference.conj()
phases = np.sum(np.conj(x) * y, -2, keepdims=True)
phases /= np.abs(phases)
return phases * x
def _GetSelfAdjointEigTest(dtype_, shape_, compute_v_):
def CompareEigenVectors(self, x, y, tol):
x = NormalizeEigenvectorsPhase(x)
y = NormalizeEigenvectorsPhase(y)
self.assertAllClose(x, y, atol=tol, rtol=tol)
x = EquilibrateEigenVectorPhases(x, y)
self.assertAllClose(x, y, atol=tol)
def CompareEigenDecompositions(self, x_e, x_v, y_e, y_v, tol):
num_batches = int(np.prod(x_e.shape[:-1]))
@ -103,7 +112,7 @@ def _GetSelfAdjointEigTest(dtype_, shape_, compute_v_):
else:
atol = 1e-12
np_e, np_v = np.linalg.eigh(a)
with self.test_session():
with self.test_session(use_gpu=True):
if compute_v_:
tf_e, tf_v = linalg_ops.self_adjoint_eig(constant_op.constant(a))
@ -152,7 +161,7 @@ def _GetSelfAdjointEigGradTest(dtype_, shape_, compute_v_):
tol = 1e-2
else:
tol = 1e-7
with self.test_session():
with self.test_session(use_gpu=True):
tf_a = constant_op.constant(a)
if compute_v_:
tf_e, tf_v = linalg_ops.self_adjoint_eig(tf_a)
@ -185,17 +194,16 @@ def _GetSelfAdjointEigGradTest(dtype_, shape_, compute_v_):
return Test
if __name__ == '__main__':
for compute_v in [True, False]:
for dtype in (
dtypes_lib.float32, dtypes_lib.float64,
dtypes_lib.complex64, dtypes_lib.complex128):
if __name__ == "__main__":
for compute_v in True, False:
for dtype in (dtypes_lib.float32, dtypes_lib.float64, dtypes_lib.complex64,
dtypes_lib.complex128):
for size in 1, 2, 5, 10:
for batch_dims in [(), (3,)] + [(3, 2)] * (max(size, size) < 10):
shape = batch_dims + (size, size)
name = '%s_%s_%s' % (dtype, '_'.join(map(str, shape)), compute_v)
setattr(SelfAdjointEigTest, 'testSelfAdjointEig_' + name,
_GetSelfAdjointEigTest(dtype, shape, compute_v))
setattr(SelfAdjointEigGradTest, 'testSelfAdjointEigGrad_' + name,
_GetSelfAdjointEigGradTest(dtype, shape, compute_v))
name = "%s_%s_%s" % (dtype, "_".join(map(str, shape)), compute_v)
_AddTest(SelfAdjointEigTest, "SelfAdjointEig", name,
_GetSelfAdjointEigTest(dtype, shape, compute_v))
_AddTest(SelfAdjointEigGradTest, "SelfAdjointEigGrad", name,
_GetSelfAdjointEigGradTest(dtype, shape, compute_v))
test.main()

24
tensorflow/python/layers/base.py
View File

@ -57,7 +57,8 @@ class Layer(object):
Properties:
trainable: Whether the layer should be trained (boolean).
name: The name of the layer (string).
dtype: Default dtype of the layer (dtypes.float32).
dtype: Default dtype of the layer (default of None means use the
type of the first input).
trainable_variables: List of trainable variables.
non_trainable_variables: List of non-trainable variables.
variables: List of all variables of this layer, trainable and non-trainable.
@ -68,7 +69,7 @@ class Layer(object):
"""
def __init__(self, trainable=True, name=None,
dtype=dtypes.float32, **kwargs):
dtype=None, **kwargs):
# We use a kwargs dict here because these kwargs only exist
# for compatibility reasons.
# The list of kwargs is subject to changes in the future.
@ -97,7 +98,7 @@ class Layer(object):
self._graph = ops.get_default_graph()
self._per_input_losses = {}
self._per_input_updates = {}
self.dtype = dtypes.as_dtype(dtype).name
self._dtype = None if dtype is None else dtypes.as_dtype(dtype).name
self.input_spec = None
self._compute_previous_mask = ('mask' in estimator_util.fn_args(self.call)
or hasattr(self, 'compute_mask'))
@ -131,6 +132,10 @@ class Layer(object):
batch_size = kwargs.get('batch_size')
self.batch_input_shape = (batch_size,) + tuple(kwargs['input_shape'])
@property
def dtype(self):
return self._dtype
@property
def scope_name(self):
if not self._scope:
@ -389,7 +394,7 @@ class Layer(object):
Arguments:
name: variable name.
shape: variable shape.
dtype: The type of the variable. Defaults to `self.dtype`.
dtype: The type of the variable. Defaults to `self.dtype` or `float32`.
initializer: initializer instance (callable).
regularizer: regularizer instance (callable).
trainable: whether the variable should be part of the layer's
@ -414,7 +419,7 @@ class Layer(object):
raise RuntimeError('Variable regularization not supported in Eager '
'mode.')
if dtype is None:
dtype = self.dtype
dtype = self.dtype or dtypes.float32
self._set_scope(None)
vs_reuse = ((self.built or self._reuse)
@ -526,6 +531,11 @@ class Layer(object):
# Check input assumptions set before layer building, e.g. input rank.
self._assert_input_compatibility(inputs)
input_list = nest.flatten(inputs)
if input_list and self._dtype is None:
try:
self._dtype = input_list[0].dtype.name
except AttributeError:
pass
input_shapes = [x.get_shape() for x in input_list]
if len(input_shapes) == 1:
self.build(input_shapes[0])
@ -1406,8 +1416,8 @@ class Network(Layer):
self.trainable = True
# A Network does not create weights of its own, thus it is already built.
self.built = True
# A Network does not create weights of its own, thus dtype is not settable.
self.dtype = None
# A Network does not create weights of its own, thus has no dtype.
self._dtype = None
# The following are implemented as property functions:
# self.trainable_weights
# self.non_trainable_weights

5
tensorflow/python/lib/io/file_io_test.py
View File

@ -35,6 +35,11 @@ class FileIoTest(test.TestCase):
def tearDown(self):
file_io.delete_recursively(self._base_dir)
def testEmptyFilename(self):
f = file_io.FileIO("", mode="r")
with self.assertRaises(errors.NotFoundError):
_ = f.read()
def testFileDoesntExist(self):
file_path = os.path.join(self._base_dir, "temp_file")
self.assertFalse(file_io.file_exists(file_path))

7
tensorflow/tools/api/golden/tensorflow.keras.estimator.pbtxt Normal file
View File

@ -0,0 +1,7 @@
path: "tensorflow.keras.estimator"
tf_module {
member_method {
name: "model_to_estimator"
argspec: "args=[\'keras_model\', \'keras_model_path\', \'custom_objects\', \'model_dir\', \'config\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\'], "
}
}

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-activation.pbtxt
View File

@ -4,6 +4,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-activity-regularization.pbtxt
View File

@ -4,6 +4,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-add.pbtxt
View File

@ -5,6 +5,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-alpha-dropout.pbtxt
View File

@ -4,6 +4,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling1-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling2-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling3-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-average.pbtxt
View File

@ -5,6 +5,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool1-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool2-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool3-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-batch-normalization.pbtxt
View File

@ -5,6 +5,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-bidirectional.pbtxt
View File

@ -13,6 +13,10 @@ tf_class {
name: "constraints"
mtype: "<type \'property\'>"
}
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-concatenate.pbtxt
View File

@ -5,6 +5,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

4
tensorflow/tools/api/golden/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt
View File

@ -6,6 +6,10 @@ tf_class {
is_instance: "<class \'tensorflow.python.keras._impl.keras.engine.topology.Layer\'>"
is_instance: "<class \'tensorflow.python.layers.base.Layer\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
mtype: "<type \'property\'>"
}
member {
name: "graph"
mtype: "<type \'property\'>"

Some files were not shown because too many files have changed in this diff Show More