Skip to content

Commit aa0e357

Browse files
wawltorwawltor
authored
Merge pull request PaddlePaddle#15 from ZHUI/lstm_cpu
fix sequence length
2 parents 10d08ff + a044bcf commit aa0e357

File tree

2 files changed

+70
-29
lines changed

2 files changed

+70
-29
lines changed

‎paddle/fluid/operators/cudnn_lstm_op.h‎

Lines changed: 46 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -82,24 +82,27 @@ void create_mask_matrix(const framework::ExecutionContext& context,
8282
const bool& is_reverse) {
8383
const auto& seq_len_vec = GetDataFromTensor<int>(sequence_length);
8484
const int& table_width = mask_matrix->dims()[0];
85+
VLOG(2) << "INPUT MASK TENSOR SHAPE:" << mask_matrix->dims();
8586
Tensor temp;
8687
temp.Resize(
8788
framework::make_ddim({mask_matrix->dims()[1], mask_matrix->dims()[0]}));
8889
T* data_temp = temp.mutable_data<T>(context.GetPlace());
89-
std::memset(data_temp, 1, mask_matrix->numel() * sizeof(T));
90+
std::fill(data_temp, data_temp + mask_matrix->numel(), static_cast<T>(1.0));
9091
for (unsigned int i = 0; i < seq_len_vec.size(); i++) {
9192
// reset the mask matrix
9293
if (seq_len_vec[i] == table_width) {
9394
continue;
9495
}
9596
if (is_reverse) {
96-
std::memset(data_temp + i * table_width * sizeof(T), 0,
97-
(table_width - seq_len_vec[i]) * sizeof(T));
97+
std::fill(data_temp + i * table_width,
98+
data_temp + i * table_width + seq_len_vec[i],
99+
static_cast<T>(0));
98100
} else {
99-
std::memset(data_temp + (i * table_width + seq_len_vec[i]) * sizeof(T), 0,
100-
(table_width - seq_len_vec[i]) * sizeof(T));
101+
std::fill(data_temp + i * table_width + seq_len_vec[i],
102+
data_temp + (i + 1) * table_width, static_cast<T>(0));
101103
}
102104
}
105+
Print2DTensor<T>(&temp, "Original mask Tensor");
103106
// transpose the result for the mask
104107
mask_matrix->mutable_data<T>(context.GetPlace());
105108
std::vector<int> trans_vec;
@@ -125,8 +128,8 @@ void dropout_cpu_function_inplace(const framework::ExecutionContext& context,
125128
auto mask_data = mask->mutable_data<uint8_t>(context.GetPlace());
126129
// Special case when dropout_prob is 1.0
127130
if (dropout_prob == 1.0f) {
128-
std::memset(x_data, 0, size * sizeof(*x_data));
129-
std::memset(mask_data, 0, size * sizeof(*mask_data)); // NOLINT
131+
std::fill(x_data, x_data + size, static_cast<T>(0));
132+
std::fill(mask_data, mask_data + size, static_cast<T>(0));
130133
return;
131134
}
132135
auto engine = framework::GetCPURandomEngine(seed_number);
@@ -145,7 +148,7 @@ void dropout_cpu_function_inplace(const framework::ExecutionContext& context,
145148
}
146149
auto mask_data = mask->data<uint8_t>();
147150
if (dropout_prob == 1.0f) {
148-
std::memset(x_data, 0, size * sizeof(*x_data));
151+
std::fill(x_data, x_data + size, static_cast<T>(0));
149152
return;
150153
}
151154
for (size_t i = 0; i < size; ++i) {
@@ -300,15 +303,27 @@ struct LSTMCell : Cell<T> {
300303
cell_act, cand_act);
301304
framework::TensorCopy(*output, device_ctx->GetPlace(), *device_ctx, last_h);
302305
Print3DTensor<T>(last_h, "last_h");
303-
// auto eigen_output =
304-
// framework::EigenMatrix<T>::Reshape(*output, output->dims().size() -
305-
// 1);
306-
// auto eigen_mask = framework::EigenMatrix<T>::From(
307-
// mask_tensor, framework::make_ddim({mask_tensor.dims()[1], 1}));
308-
//// eigen_output.device(device_ctx->eigen_device()) =
309-
// eigen_output =
310-
// eigen_output *
311-
// eigen_mask.broadcast(Eigen::DSizes<int, 2>(1, output->dims()[1]));
306+
307+
auto eigen_init_h =
308+
framework::EigenMatrix<T>::Reshape(*init_h, init_h->dims().size() - 1);
309+
auto eigen_last_h =
310+
framework::EigenMatrix<T>::Reshape(*last_h, last_h->dims().size() - 1);
311+
312+
auto eigen_mask = framework::EigenMatrix<T>::From(
313+
mask_tensor, framework::make_ddim({mask_tensor.dims()[1], 1}));
314+
// eigen_output.device(device_ctx->eigen_device()) =
315+
auto eigen_mask_broadcast =
316+
eigen_mask.broadcast(Eigen::DSizes<int, 2>(1, output->dims()[1]));
317+
auto& place = *device_ctx->eigen_device();
318+
eigen_last_h.device(place) = eigen_last_h * eigen_mask_broadcast +
319+
eigen_init_h * (1 - eigen_mask_broadcast);
320+
321+
auto eigen_init_c =
322+
framework::EigenMatrix<T>::Reshape(*init_c, init_c->dims().size() - 1);
323+
auto eigen_last_c =
324+
framework::EigenMatrix<T>::Reshape(*last_c, last_c->dims().size() - 1);
325+
eigen_last_c.device(place) = eigen_last_c * eigen_mask_broadcast +
326+
eigen_init_c * (1 - eigen_mask_broadcast);
312327
}
313328
};
314329

@@ -367,7 +382,9 @@ struct Layer {
367382
framework::EigenMatrix<T>::Reshape(*output, output->dims().size() - 1);
368383
auto eigen_mask = framework::EigenMatrix<T>::From(
369384
mask_tensor, framework::make_ddim({mask_tensor.dims()[1], 1}));
370-
eigen_output =
385+
auto& place = *context.template device_context<platform::CPUDeviceContext>()
386+
.eigen_device();
387+
eigen_output.device(place) =
371388
eigen_output *
372389
eigen_mask.broadcast(Eigen::DSizes<int, 2>(1, output->dims()[1]));
373390
}
@@ -412,6 +429,7 @@ struct SingleLayer : Layer<T> {
412429
mask_matrix.Resize(framework::make_ddim({time_step, input->dims()[1]}));
413430
if (has_sequence_length) {
414431
create_mask_matrix<T>(context, sequence_length, &mask_matrix, false);
432+
Print2DTensor<T>(&mask_matrix, "Mask Matrix");
415433
mask_tensor_list = Unbind(mask_matrix);
416434
}
417435

@@ -447,9 +465,9 @@ struct SingleLayer : Layer<T> {
447465
init_c_holder, last_h_holder, last_c_holder, &output_tensors[i],
448466
mask_tensor_list[i]);
449467
}
450-
// if (has_sequence_length) {
451-
// this->postprocess(context, &output_tensors[i], mask_tensor_list[i]);
452-
//}
468+
if (has_sequence_length) {
469+
this->postprocess(context, &output_tensors[i], mask_tensor_list[i]);
470+
}
453471
}
454472
if (time_step % 2 == 0) {
455473
framework::TensorCopy(*last_h_holder, context.GetPlace(), dev_ctx,
@@ -717,7 +735,13 @@ class CudnnLSTMCPUKernel : public framework::OpKernel<T> {
717735
auto* weight = ctx.Input<Tensor>("W");
718736
auto* init_h = ctx.Input<Tensor>("InitH");
719737
auto* init_c = ctx.Input<Tensor>("InitC");
720-
auto* sequence_length = ctx.Input<Tensor>("SequenceLength");
738+
739+
bool has_seq_length = ctx.HasInput("SequenceLength");
740+
const Tensor* sequence_length = nullptr;
741+
if (has_seq_length) {
742+
sequence_length = ctx.Input<Tensor>("SequenceLength");
743+
}
744+
// auto* sequence_length = ctx.Input<Tensor>("SequenceLength");
721745
auto* last_h = ctx.Output<Tensor>("LastH");
722746
auto* last_c = ctx.Output<Tensor>("LastC");
723747
auto* output = ctx.Output<Tensor>("Out");

‎python/paddle/fluid/tests/unittests/test_lstm_cudnn_op.py‎

Lines changed: 24 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -359,6 +359,9 @@ def setUp(self):
359359
self.dtype = np.float64
360360
self.sequence_length = np.array([12, 11, 10, 9, 8], dtype=np.int32)
361361
self.num_layers = 1
362+
self.is_bidirec = False
363+
364+
self.set_attrs()
362365

363366
seq_length = 12
364367
batch_size = 5
@@ -405,10 +408,10 @@ def setUp(self):
405408
}
406409
self.attrs = {
407410
'dropout_prob': 0.0,
408-
'is_bidirec': False,
411+
'is_bidirec': self.is_bidirec,
409412
'input_size': input_size,
410413
'hidden_size': hidden_size,
411-
'num_layers': 1,
414+
'num_layers': self.num_layers,
412415
}
413416
self.outputs = {
414417
'Out': output,
@@ -433,13 +436,27 @@ def test_grad_with_place(self):
433436
['Out', 'LastH', 'LastC'])
434437

435438

436-
@unittest.skipIf(not core.is_compiled_with_cuda(),
437-
"core is not compiled with CUDA")
438-
class TestCUDNNLstmOp2(TestCUDNNLstmOp):
439-
def set_attrs(self):
440-
self.num_layers = 2
439+
#@unittest.skipIf(not core.is_compiled_with_cuda(),
440+
# "core is not compiled with CUDA")
441+
#class TestCUDNNLstmOp2(TestCUDNNLstmOp):
442+
# def set_attrs(self):
443+
# self.num_layers = 2
441444

442445

446+
class TestCUDNNLstmCpu(TestCUDNNLstmOp):
447+
def test_output_with_place(self):
448+
place = core.CPUPlace()
449+
self.check_output_with_place(
450+
place, no_check_set=['Reserve', 'StateOut'])
451+
452+
def test_grad_with_place(self):
453+
pass
454+
455+
456+
#class TestCUDNNLstmCpu2(TestCUDNNLstmCpu):
457+
# def set_attrs(self):
458+
# self.num_layers=2
459+
#
443460
@unittest.skipIf(not core.is_compiled_with_cuda(),
444461
"core is not compiled with CUDA")
445462
class TestCUDNNlstmAPI(unittest.TestCase):

0 commit comments

Comments
 (0)