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

Enhance ShuffleDataset tests and add tests for ShuffleAndRepeatDataset

Browse Source
This commit is contained in:
Fei Hu 2019-04-22 19:43:24 -07:00
parent 6702143f14
commit d469a23e1e
1 changed files with 470 additions and 220 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

690
tensorflow/core/kernels/data/shuffle_dataset_op_test.cc
View File

@ -16,30 +16,41 @@ namespace tensorflow {
namespace data { namespace data {
namespace { namespace {
constexpr char kNodeName[] = "shuffle_dataset"; constexpr char kShuffleNodeName[] = "shuffle_dataset";
constexpr char kOpName[] = "ShuffleDataset"; constexpr char kShuffleOpName[] = "ShuffleDataset";
constexpr char kShuffleAndRepeatNodeName[] = "shuffle_and_repeat_dataset";
constexpr char kShuffleAndRepeatOpName[] = "ShuffleAndRepeatDataset";
class ShuffleDatasetOpTest : public DatasetOpsTestBase { class ShuffleDatasetOpTest : public DatasetOpsTestBase {
protected: protected:
// Creates a new `ShuffleDataset` op kernel // Creates a new `ShuffleDataset`/`ShuffleAndRepeatDataset` op kernel
Status CreateShuffleDatasetOpKernel( Status CreateDatasetOpKernel(
bool reshuffle_each_iteration, const DataTypeVector& output_types, int64 count, bool reshuffle_each_iteration,
const DataTypeVector& output_types,
const std::vector<PartialTensorShape>& output_shapes, const std::vector<PartialTensorShape>& output_shapes,
std::unique_ptr<OpKernel>* shuffle_dataset_kernel) { std::unique_ptr<OpKernel>* shuffle_dataset_kernel) {
NodeDef node_def = test::function::NDef( NodeDef node_def;
kNodeName, kOpName, {"input_dataset", "buffer_size", "seed", "seed2"}, if (count == 1) {
{{"reshuffle_each_iteration", reshuffle_each_iteration}, node_def = test::function::NDef(
{"output_types", output_types}, kShuffleNodeName, kShuffleOpName,
{"output_shapes", output_shapes}}); {"input_dataset", "buffer_size", "seed", "seed2"},
{{"reshuffle_each_iteration", reshuffle_each_iteration},
{"output_types", output_types},
{"output_shapes", output_shapes}});
} else {
node_def = test::function::NDef(
kShuffleAndRepeatNodeName, kShuffleAndRepeatOpName,
{"input_dataset", "buffer_size", "seed", "seed2", "count"},
{{"output_types", output_types}, {"output_shapes", output_shapes}});
}
TF_RETURN_IF_ERROR(CreateOpKernel(node_def, shuffle_dataset_kernel)); TF_RETURN_IF_ERROR(CreateOpKernel(node_def, shuffle_dataset_kernel));
return Status::OK(); return Status::OK();
} }
// Creates a new `ShuffleDataset` op kernel context. // Creates a new `ShuffleDataset`/`ShuffleAndRepeatDataset` op kernel context.
Status CreateShuffleDatasetContext( Status CreateDatasetContext(OpKernel* const op_kernel,
OpKernel* const op_kernel, gtl::InlinedVector<TensorValue, 4>* const inputs,
gtl::InlinedVector<TensorValue, 4>* const inputs, std::unique_ptr<OpKernelContext>* context) {
std::unique_ptr<OpKernelContext>* context) {
TF_RETURN_IF_ERROR(CheckOpKernelInput(*op_kernel, *inputs)); TF_RETURN_IF_ERROR(CheckOpKernelInput(*op_kernel, *inputs));
TF_RETURN_IF_ERROR(CreateOpKernelContext(op_kernel, inputs, context)); TF_RETURN_IF_ERROR(CreateOpKernelContext(op_kernel, inputs, context));
return Status::OK(); return Status::OK();
@ -57,8 +68,10 @@ struct TestCase {
Tensor buffer_size; Tensor buffer_size;
Tensor seed; Tensor seed;
Tensor seed2; Tensor seed2;
Tensor count;
bool reshuffle_each_iteration; bool reshuffle_each_iteration;
std::vector<Tensor> expected_outputs; std::vector<Tensor> expected_shuffle_outputs;
std::vector<Tensor> expected_reshuffle_outputs;
DataTypeVector expected_output_dtypes; DataTypeVector expected_output_dtypes;
std::vector<PartialTensorShape> expected_output_shapes; std::vector<PartialTensorShape> expected_output_shapes;
int64 expected_cardinality; int64 expected_cardinality;
@ -76,7 +89,7 @@ std::vector<Tensor> ConvertToTensorVec(std::vector<T> values) {
return tensors; return tensors;
} }
// Test case 1: normal case with reshuffle_each_iteration = false // Test case 1: test shuffle_dataset with reshuffle_each_iteration = false.
TestCase TestCase1() { TestCase TestCase1() {
return { return {
/*range_data_param*/ {0, 10, 1}, /*range_data_param*/ {0, 10, 1},
@ -84,16 +97,19 @@ TestCase TestCase1() {
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {3}), DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {3}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}), /*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}), /*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ false, /*reshuffle_each_iteration*/ false,
/*expected_outputs*/ /*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}), ConvertToTensorVec<int64>({2, 3, 0, 5, 6, 4, 7, 8, 9, 1}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({2, 3, 0, 5, 6, 4, 7, 8, 9, 1}),
/*expected_output_dtypes*/ {DT_INT64}, /*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})}, /*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 10, /*expected_cardinality*/ 10,
/*breakpoints*/ {0, 1, 9}}; /*breakpoints*/ {0, 1, 9}};
} }
// Test case 2: normal case with reshuffle_each_iteration = true // Test case 2: test shuffle_dataset with reshuffle_each_iteration = true.
TestCase TestCase2() { TestCase TestCase2() {
return { return {
/*range_data_param*/ {0, 10, 1}, /*range_data_param*/ {0, 10, 1},
@ -101,26 +117,73 @@ TestCase TestCase2() {
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}), DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}), /*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}), /*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ true, /*reshuffle_each_iteration*/ true,
/*expected_outputs*/ /*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}), ConvertToTensorVec<int64>({2, 6, 1, 3, 9, 5, 0, 8, 7, 4}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({1, 6, 0, 5, 2, 7, 4, 3, 9, 8}),
/*expected_output_dtypes*/ {DT_INT64}, /*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})}, /*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 10, /*expected_cardinality*/ 10,
/*breakpoints*/ {0, 1, 9}}; /*breakpoints*/ {0, 1, 9}};
} }
// Test case 3: special case with buffer_size = 1 & // Test case 3: similar with the test case 2 but a smaller buffer size than
// reshuffle_each_iteration = true // the input dataset.
TestCase TestCase3() { TestCase TestCase3() {
return {
/*range_data_param*/ {0, 10, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ true,
/*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({0, 2, 1, 3, 5, 6, 4, 7, 8, 9}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({1, 0, 2, 3, 4, 5, 6, 7, 9, 8}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 10,
/*breakpoints*/ {0, 1, 9}};
}
// Test case 4: similar with the test case 2 but has different seeds.
TestCase TestCase4() {
return {
/*range_data_param*/ {0, 10, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ true,
/*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({3, 0, 8, 1, 5, 4, 7, 2, 6, 9}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({4, 6, 9, 0, 1, 8, 2, 7, 3, 5}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 10,
/*breakpoints*/ {0, 1, 9}};
}
// Test case 5: test shuffle_dataset with buffer_size = 1 &
// reshuffle_each_iteration = true.
TestCase TestCase5() {
return { return {
/*range_data_param*/ {0, 10, 1}, /*range_data_param*/ {0, 10, 1},
/*buffer_size*/ /*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}), DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}), /*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}), /*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ true, /*reshuffle_each_iteration*/ true,
/*expected_outputs*/ /*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}), ConvertToTensorVec<int64>({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}),
/*expected_output_dtypes*/ {DT_INT64}, /*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})}, /*expected_output_shapes*/ {PartialTensorShape({})},
@ -128,21 +191,125 @@ TestCase TestCase3() {
/*breakpoints*/ {0, 1, 9}}; /*breakpoints*/ {0, 1, 9}};
} }
TestCase InvalidBufferSizeTestCase() { // Test case 6: test shuffle_dataset with an empty input dataset.
TestCase TestCase6() {
return {
/*range_data_param*/ {0, 0, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ true,
/*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 0,
/*breakpoints*/ {0, 1, 9}};
}
// Test case 7: test shuffle_and_repeat_dataset with buffer_size = 10 &
// count = 2.
TestCase TestCase7() {
return {
/*range_data_param*/ {0, 10, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*reshuffle_each_iteration*/ false,
/*expected_shuffle_outputs*/
ConvertToTensorVec<int64>(
{9, 0, 8, 6, 1, 3, 7, 2, 4, 5, 4, 3, 0, 5, 8, 2, 6, 9, 7, 1}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>(
{9, 0, 8, 6, 1, 3, 7, 2, 4, 5, 4, 3, 0, 5, 8, 2, 6, 9, 7, 1}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 20,
/*breakpoints*/ {0, 5, 22}};
}
// Test case 8: test shuffle_and_repeat_dataset with buffer_size = 10 &
// count = -1
TestCase TestCase8() {
return {
/*range_data_param*/ {0, 3, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {-1}),
/*reshuffle_each_iteration*/ false,
/*expected_shuffle_outputs*/
ConvertToTensorVec<int64>(
{2, 0, 1, 2, 0, 1, 1, 2, 0, 1, 0, 2, 2, 0, 1, 1, 0, 2, 2, 1, 0}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>(
{2, 0, 1, 2, 0, 1, 1, 2, 0, 1, 0, 2, 2, 0, 1, 1, 0, 2, 2, 1, 0}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ kInfiniteCardinality,
/*breakpoints*/ {0, 5, 20}};
}
TestCase InvalidBufferSizeTestCaseForShuffleDataset() {
return { return {
/*range_data_param*/ {0, 10, 1}, /*range_data_param*/ {0, 10, 1},
/*buffer_size*/ /*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {-1}), DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {-1}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}), /*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}), /*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*reshuffle_each_iteration*/ true, /*reshuffle_each_iteration*/ true,
/*expected_outputs*/ ConvertToTensorVec<int64>({}), /*expected_shuffle_outputs*/ ConvertToTensorVec<int64>({}),
/*expected_reshuffle_outputs*/ ConvertToTensorVec<int64>({}),
/*expected_output_dtypes*/ {DT_INT64}, /*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})}, /*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 10, /*expected_cardinality*/ 0,
/*breakpoints*/ {0, 1, 9}}; /*breakpoints*/ {0, 1, 9}};
} }
TestCase InvalidBufferSizeTestCaseForShuffleAndRepeatDataset() {
return {
/*range_data_param*/ {0, 10, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {-1}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*reshuffle_each_iteration*/ true,
/*expected_shuffle_outputs*/ ConvertToTensorVec<int64>({}),
/*expected_reshuffle_outputs*/ ConvertToTensorVec<int64>({}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 0,
/*breakpoints*/ {0, 1, 9}};
}
TestCase InvalidCountTestCaseForShuffleAndRepeatDataset() {
return {
/*range_data_param*/ {0, 3, 1},
/*buffer_size*/
DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {10}),
/*seed*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {1}),
/*seed2*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {2}),
/*count*/ DatasetOpsTestBase::CreateTensor<int64>(TensorShape({}), {0}),
/*reshuffle_each_iteration*/ false,
/*expected_shuffle_outputs*/
ConvertToTensorVec<int64>({}),
/*expected_reshuffle_outputs*/
ConvertToTensorVec<int64>({}),
/*expected_output_dtypes*/ {DT_INT64},
/*expected_output_shapes*/ {PartialTensorShape({})},
/*expected_cardinality*/ 0,
/*breakpoints*/ {0, 5, 20}};
}
class ParameterizedShuffleDatasetOpTest class ParameterizedShuffleDatasetOpTest
: public ShuffleDatasetOpTest, : public ShuffleDatasetOpTest,
public ::testing::WithParamInterface<TestCase> {}; public ::testing::WithParamInterface<TestCase> {};
@ -153,10 +320,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, GetNext) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -170,49 +340,78 @@ TEST_P(ParameterizedShuffleDatasetOpTest, GetNext) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
std::unique_ptr<IteratorContext> iterator_ctx; std::unique_ptr<IteratorContext> iterator_ctx;
TF_ASSERT_OK( TF_ASSERT_OK(CreateIteratorContext(dataset_context.get(), &iterator_ctx));
CreateIteratorContext(shuffle_dataset_context.get(), &iterator_ctx));
std::unique_ptr<IteratorBase> iterator; std::unique_ptr<IteratorBase> iterator;
TF_ASSERT_OK( TF_ASSERT_OK(
shuffle_dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator)); dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator));
bool end_of_sequence = false; bool end_of_sequence = false;
std::vector<Tensor> out_tensors; std::vector<Tensor> shuffled_out_tensors;
while (!end_of_sequence) { while (!end_of_sequence) {
std::vector<Tensor> next; std::vector<Tensor> next;
TF_EXPECT_OK( TF_EXPECT_OK(
iterator->GetNext(iterator_ctx.get(), &next, &end_of_sequence)); iterator->GetNext(iterator_ctx.get(), &next, &end_of_sequence));
out_tensors.insert(out_tensors.end(), next.begin(), next.end()); shuffled_out_tensors.insert(shuffled_out_tensors.end(), next.begin(),
next.end());
// For the forever-repeat case, we test only a finite number of steps of
// the infinite sequence.
if (count_value == -1 && shuffled_out_tensors.size() ==
test_case.expected_shuffle_outputs.size()) {
break;
}
} }
// When `buffer_size = 1`, the output sequence of `ShuffleDataset` will be in // Reshuffle the dataset.
// order, so we need to consider the element sequence when evaluating the end_of_sequence = false;
// result for this case. TF_ASSERT_OK(
bool expect_items_equal = test_case.buffer_size.flat<int64>()(0) > 1; dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator));
TF_EXPECT_OK(ExpectEqual(out_tensors, test_case.expected_outputs, std::vector<Tensor> reshuffled_out_tensors;
/*expect_items_equal*/ expect_items_equal)); while (!end_of_sequence) {
std::vector<Tensor> next;
TF_EXPECT_OK(
iterator->GetNext(iterator_ctx.get(), &next, &end_of_sequence));
reshuffled_out_tensors.insert(reshuffled_out_tensors.end(), next.begin(),
next.end());
// For the forever-repeat case, we test only a finite number of steps of
// the infinite sequence.
if (count_value == -1 && reshuffled_out_tensors.size() ==
test_case.expected_shuffle_outputs.size()) {
break;
}
}
TF_EXPECT_OK(ExpectEqual(shuffled_out_tensors,
test_case.expected_shuffle_outputs,
/*compare_order*/ true));
TF_EXPECT_OK(ExpectEqual(reshuffled_out_tensors,
test_case.expected_reshuffle_outputs,
/*compare_order*/ true));
} }
TEST_F(ShuffleDatasetOpTest, DatasetNodeName) { TEST_P(ParameterizedShuffleDatasetOpTest, DatasetNodeName) {
int thread_num = 2, cpu_num = 2; int thread_num = 2, cpu_num = 2;
TestCase test_case = TestCase1(); TestCase test_case = GetParam();
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -226,28 +425,36 @@ TEST_F(ShuffleDatasetOpTest, DatasetNodeName) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
EXPECT_EQ(shuffle_dataset->node_name(), kNodeName); if (count_value == 1) {
EXPECT_EQ(dataset->node_name(), kShuffleNodeName);
} else {
EXPECT_EQ(dataset->node_name(), kShuffleAndRepeatNodeName);
}
} }
TEST_F(ShuffleDatasetOpTest, DatasetTypeString) { TEST_P(ParameterizedShuffleDatasetOpTest, DatasetTypeString) {
int thread_num = 2, cpu_num = 2; int thread_num = 2, cpu_num = 2;
TestCase test_case = TestCase1(); TestCase test_case = GetParam();
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -261,16 +468,21 @@ TEST_F(ShuffleDatasetOpTest, DatasetTypeString) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
EXPECT_EQ(shuffle_dataset->type_string(), kOpName); if (count_value == 1) {
EXPECT_EQ(dataset->type_string(), kShuffleOpName);
} else {
EXPECT_EQ(dataset->type_string(), kShuffleAndRepeatOpName);
}
} }
TEST_P(ParameterizedShuffleDatasetOpTest, DatasetOutputDtypes) { TEST_P(ParameterizedShuffleDatasetOpTest, DatasetOutputDtypes) {
@ -279,10 +491,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, DatasetOutputDtypes) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -296,16 +511,17 @@ TEST_P(ParameterizedShuffleDatasetOpTest, DatasetOutputDtypes) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
TF_EXPECT_OK(VerifyTypesMatch(shuffle_dataset->output_dtypes(), TF_EXPECT_OK(VerifyTypesMatch(dataset->output_dtypes(),
test_case.expected_output_dtypes)); test_case.expected_output_dtypes));
} }
@ -315,10 +531,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, DatasetOutputShapes) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -332,16 +551,17 @@ TEST_P(ParameterizedShuffleDatasetOpTest, DatasetOutputShapes) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
TF_EXPECT_OK(VerifyShapesCompatible(shuffle_dataset->output_shapes(), TF_EXPECT_OK(VerifyShapesCompatible(dataset->output_shapes(),
test_case.expected_output_shapes)); test_case.expected_output_shapes));
} }
@ -351,10 +571,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, Cardinality) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -368,17 +591,17 @@ TEST_P(ParameterizedShuffleDatasetOpTest, Cardinality) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
TF_EXPECT_OK(VerifyShapesCompatible(shuffle_dataset->output_shapes(), EXPECT_EQ(dataset->Cardinality(), test_case.expected_cardinality);
test_case.expected_output_shapes));
} }
TEST_P(ParameterizedShuffleDatasetOpTest, DatasetSave) { TEST_P(ParameterizedShuffleDatasetOpTest, DatasetSave) {
@ -387,10 +610,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, DatasetSave) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -404,20 +630,21 @@ TEST_P(ParameterizedShuffleDatasetOpTest, DatasetSave) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
std::unique_ptr<SerializationContext> serialization_context; std::unique_ptr<SerializationContext> serialization_context;
TF_ASSERT_OK(CreateSerializationContext(&serialization_context)); TF_ASSERT_OK(CreateSerializationContext(&serialization_context));
VariantTensorData data; VariantTensorData data;
VariantTensorDataWriter writer(&data); VariantTensorDataWriter writer(&data);
TF_ASSERT_OK(shuffle_dataset->Save(serialization_context.get(), &writer)); TF_ASSERT_OK(dataset->Save(serialization_context.get(), &writer));
TF_ASSERT_OK(writer.Flush()); TF_ASSERT_OK(writer.Flush());
} }
@ -427,10 +654,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, IteratorOutputDtypes) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -444,21 +674,21 @@ TEST_P(ParameterizedShuffleDatasetOpTest, IteratorOutputDtypes) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
std::unique_ptr<IteratorContext> iterator_ctx; std::unique_ptr<IteratorContext> iterator_ctx;
TF_ASSERT_OK( TF_ASSERT_OK(CreateIteratorContext(dataset_context.get(), &iterator_ctx));
CreateIteratorContext(shuffle_dataset_context.get(), &iterator_ctx));
std::unique_ptr<IteratorBase> iterator; std::unique_ptr<IteratorBase> iterator;
TF_ASSERT_OK( TF_ASSERT_OK(
shuffle_dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator)); dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator));
TF_EXPECT_OK(VerifyTypesMatch(iterator->output_dtypes(), TF_EXPECT_OK(VerifyTypesMatch(iterator->output_dtypes(),
test_case.expected_output_dtypes)); test_case.expected_output_dtypes));
@ -470,10 +700,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, IteratorOutputShapes) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -487,36 +720,39 @@ TEST_P(ParameterizedShuffleDatasetOpTest, IteratorOutputShapes) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
std::unique_ptr<IteratorContext> iterator_ctx; std::unique_ptr<IteratorContext> iterator_ctx;
TF_ASSERT_OK( TF_ASSERT_OK(CreateIteratorContext(dataset_context.get(), &iterator_ctx));
CreateIteratorContext(shuffle_dataset_context.get(), &iterator_ctx));
std::unique_ptr<IteratorBase> iterator; std::unique_ptr<IteratorBase> iterator;
TF_ASSERT_OK( TF_ASSERT_OK(
shuffle_dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator)); dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator));
TF_EXPECT_OK(VerifyShapesCompatible(iterator->output_shapes(), TF_EXPECT_OK(VerifyShapesCompatible(iterator->output_shapes(),
test_case.expected_output_shapes)); test_case.expected_output_shapes));
} }
TEST_F(ShuffleDatasetOpTest, IteratorOutputPrefix) { TEST_P(ParameterizedShuffleDatasetOpTest, IteratorOutputPrefix) {
int thread_num = 2, cpu_num = 2; int thread_num = 2, cpu_num = 2;
TestCase test_case = TestCase1(); TestCase test_case = GetParam();
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -530,23 +766,27 @@ TEST_F(ShuffleDatasetOpTest, IteratorOutputPrefix) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
std::unique_ptr<IteratorContext> iterator_ctx; std::unique_ptr<IteratorContext> iterator_ctx;
TF_ASSERT_OK( TF_ASSERT_OK(CreateIteratorContext(dataset_context.get(), &iterator_ctx));
CreateIteratorContext(shuffle_dataset_context.get(), &iterator_ctx));
std::unique_ptr<IteratorBase> iterator; std::unique_ptr<IteratorBase> iterator;
TF_ASSERT_OK( TF_ASSERT_OK(
shuffle_dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator)); dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator));
EXPECT_EQ(iterator->prefix(), "Iterator::Shuffle"); if (count_value == 1) {
EXPECT_EQ(iterator->prefix(), "Iterator::Shuffle");
} else {
EXPECT_EQ(iterator->prefix(), "Iterator::ShuffleAndRepeat");
}
} }
TEST_P(ParameterizedShuffleDatasetOpTest, Roundtrip) { TEST_P(ParameterizedShuffleDatasetOpTest, Roundtrip) {
@ -555,10 +795,13 @@ TEST_P(ParameterizedShuffleDatasetOpTest, Roundtrip) {
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; Tensor count = test_case.count;
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( int64 count_value = count.flat<int64>()(0);
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, std::unique_ptr<OpKernel> dataset_kernel;
test_case.expected_output_shapes, &shuffle_dataset_kernel)); TF_ASSERT_OK(
CreateDatasetOpKernel(count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes,
test_case.expected_output_shapes, &dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
@ -572,21 +815,21 @@ TEST_P(ParameterizedShuffleDatasetOpTest, Roundtrip) {
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* dataset;
TF_ASSERT_OK(CreateDataset(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
shuffle_dataset_context.get(), &shuffle_dataset)); CreateDataset(dataset_kernel.get(), dataset_context.get(), &dataset));
core::ScopedUnref scoped_unref_shuffle_dataset(shuffle_dataset); core::ScopedUnref scoped_unref_dataset(dataset);
std::unique_ptr<IteratorContext> iterator_ctx; std::unique_ptr<IteratorContext> iterator_ctx;
TF_ASSERT_OK( TF_ASSERT_OK(CreateIteratorContext(dataset_context.get(), &iterator_ctx));
CreateIteratorContext(shuffle_dataset_context.get(), &iterator_ctx));
std::unique_ptr<IteratorBase> iterator; std::unique_ptr<IteratorBase> iterator;
TF_ASSERT_OK( TF_ASSERT_OK(
shuffle_dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator)); dataset->MakeIterator(iterator_ctx.get(), "Iterator", &iterator));
std::unique_ptr<SerializationContext> serialization_ctx; std::unique_ptr<SerializationContext> serialization_ctx;
TF_ASSERT_OK(CreateSerializationContext(&serialization_ctx)); TF_ASSERT_OK(CreateSerializationContext(&serialization_ctx));
@ -602,7 +845,7 @@ TEST_P(ParameterizedShuffleDatasetOpTest, Roundtrip) {
TF_EXPECT_OK(writer.Flush()); TF_EXPECT_OK(writer.Flush());
VariantTensorDataReader reader(&data); VariantTensorDataReader reader(&data);
TF_EXPECT_OK(RestoreIterator(iterator_ctx.get(), &reader, "Iterator", TF_EXPECT_OK(RestoreIterator(iterator_ctx.get(), &reader, "Iterator",
*shuffle_dataset, &iterator)); *dataset, &iterator));
while (cur_iteration <= breakpoint) { while (cur_iteration <= breakpoint) {
std::vector<Tensor> next; std::vector<Tensor> next;
@ -613,51 +856,58 @@ TEST_P(ParameterizedShuffleDatasetOpTest, Roundtrip) {
} }
} }
// When `buffer_size = 1`, the output sequence of `ShuffleDataset` will be in TF_EXPECT_OK(ExpectEqual(out_tensors, test_case.expected_shuffle_outputs,
// order, so we need to consider the element sequence when evaluating the /*compare_order*/ true));
// result for this case.
bool expect_items_equal = test_case.buffer_size.flat<int64>()(0) > 1;
TF_EXPECT_OK(ExpectEqual(out_tensors, test_case.expected_outputs,
/*expect_items_equal*/ expect_items_equal));
} }
INSTANTIATE_TEST_SUITE_P(ShuffleDatasetOpTest, INSTANTIATE_TEST_SUITE_P(ShuffleDatasetOpTest,
ParameterizedShuffleDatasetOpTest, ParameterizedShuffleDatasetOpTest,
::testing::ValuesIn(std::vector<TestCase>( ::testing::ValuesIn(std::vector<TestCase>(
{TestCase1(), TestCase2(), TestCase3()}))); {TestCase1(), TestCase2(), TestCase3(),
TestCase4(), TestCase5(), TestCase6(),
TestCase7(), TestCase8()})));
TEST_F(ShuffleDatasetOpTest, InvalidBufferSize) { TEST_F(ShuffleDatasetOpTest, InvalidArguments) {
int thread_num = 2, cpu_num = 2; int thread_num = 2, cpu_num = 2;
TestCase test_case = InvalidBufferSizeTestCase(); std::vector<TestCase> test_cases = {
InvalidBufferSizeTestCaseForShuffleDataset(),
InvalidBufferSizeTestCaseForShuffleAndRepeatDataset(),
InvalidCountTestCaseForShuffleAndRepeatDataset()};
TF_ASSERT_OK(InitThreadPool(thread_num)); TF_ASSERT_OK(InitThreadPool(thread_num));
TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num)); TF_ASSERT_OK(InitFunctionLibraryRuntime({}, cpu_num));
std::unique_ptr<OpKernel> shuffle_dataset_kernel; for (const auto& test_case : test_cases) {
TF_ASSERT_OK(CreateShuffleDatasetOpKernel( Tensor count = test_case.count;
test_case.reshuffle_each_iteration, test_case.expected_output_dtypes, int64 count_value = count.flat<int64>()(0);
test_case.expected_output_shapes, &shuffle_dataset_kernel)); std::unique_ptr<OpKernel> dataset_kernel;
TF_ASSERT_OK(CreateDatasetOpKernel(
count_value, test_case.reshuffle_each_iteration,
test_case.expected_output_dtypes, test_case.expected_output_shapes,
&dataset_kernel));
DatasetBase* range_dataset; DatasetBase* range_dataset;
TF_ASSERT_OK(CreateRangeDataset<int64>( TF_ASSERT_OK(CreateRangeDataset<int64>(
test_case.range_data_param.start, test_case.range_data_param.end, test_case.range_data_param.start, test_case.range_data_param.end,
test_case.range_data_param.step, "range", &range_dataset)); test_case.range_data_param.step, "range", &range_dataset));
Tensor range_dataset_tensor(DT_VARIANT, TensorShape({})); Tensor range_dataset_tensor(DT_VARIANT, TensorShape({}));
TF_ASSERT_OK( TF_ASSERT_OK(
StoreDatasetInVariantTensor(range_dataset, &range_dataset_tensor)); StoreDatasetInVariantTensor(range_dataset, &range_dataset_tensor));
Tensor buffer_size = test_case.buffer_size; Tensor buffer_size = test_case.buffer_size;
Tensor seed = test_case.seed; Tensor seed = test_case.seed;
Tensor seed2 = test_case.seed2; Tensor seed2 = test_case.seed2;
gtl::InlinedVector<TensorValue, 4> inputs( gtl::InlinedVector<TensorValue, 4> inputs(
{&range_dataset_tensor, &buffer_size, &seed, &seed2}); {&range_dataset_tensor, &buffer_size, &seed, &seed2});
if (count_value != 1) inputs.push_back(&count);
std::unique_ptr<OpKernelContext> shuffle_dataset_context; std::unique_ptr<OpKernelContext> dataset_context;
TF_ASSERT_OK(CreateShuffleDatasetContext(shuffle_dataset_kernel.get(), TF_ASSERT_OK(
&inputs, &shuffle_dataset_context)); CreateDatasetContext(dataset_kernel.get(), &inputs, &dataset_context));
DatasetBase* shuffle_dataset; DatasetBase* shuffle_dataset;
EXPECT_EQ(CreateDataset(shuffle_dataset_kernel.get(), EXPECT_EQ(CreateDataset(dataset_kernel.get(), dataset_context.get(),
shuffle_dataset_context.get(), &shuffle_dataset) &shuffle_dataset)
.code(), .code(),
tensorflow::error::INVALID_ARGUMENT); tensorflow::error::INVALID_ARGUMENT);
}
} }
} // namespace } // namespace