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

Add support for batch-major input in the unidirectional LSTM Op.

Browse Source 
PiperOrigin-RevId: 217406579
This commit is contained in:
A. Unique TensorFlower 2018-10-16 16:14:24 -07:00 committed by TensorFlower Gardener
parent e6440a80c8
commit a3f855aca2
10 changed files with 342 additions and 144 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/contrib/lite/c/builtin_op_data.h
View File

@ -187,10 +187,13 @@ typedef struct {
} TfLiteLSTMParams;
typedef struct {
// Parameters for the LSTM kernel.
// Parameters needed for the underlying LSTM.
TfLiteFusedActivation activation;
float cell_clip;
float proj_clip;
// If set to true then the first dimension is time, otherwise batch.
bool time_major;
} TfLiteUnidirectionalSequenceLSTMParams;
typedef struct {

2
tensorflow/contrib/lite/core/api/flatbuffer_conversions.cc
View File

@ -399,11 +399,11 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type,
parse_activation(seq_lstm_params->fused_activation_function());
params->cell_clip = seq_lstm_params->cell_clip();
params->proj_clip = seq_lstm_params->proj_clip();
params->time_major = seq_lstm_params->time_major();
}
*builtin_data = reinterpret_cast<void*>(params);
break;
}
case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM: {
auto params =
allocator->AllocatePOD<TfLiteBidirectionalSequenceLSTMParams>();

30
tensorflow/contrib/lite/kernels/bidirectional_sequence_lstm.cc
View File

@ -876,6 +876,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
params->merge_outputs ? fw_recurrent_to_output_weights->dims->data[1] : 0;
const auto actual_bw_output = params->merge_outputs ? fw_output : bw_output;
// TODO(mirkov): add batch_major support (http://b/117326122).
switch (fw_input_to_output_weights->type) {
case kTfLiteFloat32: {
TfLiteStatus fw_pass_status = lstm_eval::EvalFloat(
@ -889,8 +890,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
fw_aux_input_to_output_weights, fw_input_gate_bias,
fw_forget_gate_bias, fw_cell_bias, fw_output_gate_bias,
fw_projection_weights, fw_projection_bias, &lstm_params,
/*forward_sequence=*/true, /*output_offset=*/0, fw_scratch_buffer,
fw_activation_state, fw_cell_state, fw_output);
/*forward_sequence=*/true, /*time_major=*/true, /*output_offset=*/0,
fw_scratch_buffer, fw_activation_state, fw_cell_state, fw_output);
TF_LITE_ENSURE_OK(context, fw_pass_status);
TfLiteStatus bw_pass_status = lstm_eval::EvalFloat(
@ -904,8 +905,9 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
bw_aux_input_to_output_weights, bw_input_gate_bias,
bw_forget_gate_bias, bw_cell_bias, bw_output_gate_bias,
bw_projection_weights, bw_projection_bias, &lstm_params,
/*forward_sequence=*/false, bw_output_offset, bw_scratch_buffer,
bw_activation_state, bw_cell_state, actual_bw_output);
/*forward_sequence=*/false, /*time_major=*/true, bw_output_offset,
bw_scratch_buffer, bw_activation_state, bw_cell_state,
actual_bw_output);
TF_LITE_ENSURE_OK(context, bw_pass_status);
return kTfLiteOk;
}
@ -942,11 +944,11 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
fw_aux_input_to_output_weights, fw_input_gate_bias,
fw_forget_gate_bias, fw_cell_bias, fw_output_gate_bias,
fw_projection_weights, fw_projection_bias, &lstm_params,
/*forward_sequence=*/true, /*output_offset=*/0, fw_scratch_buffer,
scaling_factors, prod_scaling_factors, recovered_cell_weights,
input_quantized, aux_input_quantized, fw_activation_state_quantized,
fw_cell_state_quantized, fw_activation_state, fw_cell_state,
fw_output);
/*forward_sequence=*/true, /*time_major=*/true, /*output_offset=*/0,
fw_scratch_buffer, scaling_factors, prod_scaling_factors,
recovered_cell_weights, input_quantized, aux_input_quantized,
fw_activation_state_quantized, fw_cell_state_quantized,
fw_activation_state, fw_cell_state, fw_output);
TF_LITE_ENSURE_OK(context, fw_pass_status);
TfLiteStatus bw_pass_status = lstm_eval::EvalHybrid(
@ -960,11 +962,11 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
bw_aux_input_to_output_weights, bw_input_gate_bias,
bw_forget_gate_bias, bw_cell_bias, bw_output_gate_bias,
bw_projection_weights, bw_projection_bias, &lstm_params,
/*forward_sequence=*/false, bw_output_offset, bw_scratch_buffer,
scaling_factors, prod_scaling_factors, recovered_cell_weights,
input_quantized, aux_input_quantized, bw_activation_state_quantized,
bw_cell_state_quantized, bw_activation_state, bw_cell_state,
actual_bw_output);
/*forward_sequence=*/false, /*time_major=*/true, bw_output_offset,
bw_scratch_buffer, scaling_factors, prod_scaling_factors,
recovered_cell_weights, input_quantized, aux_input_quantized,
bw_activation_state_quantized, bw_cell_state_quantized,
bw_activation_state, bw_cell_state, actual_bw_output);
TF_LITE_ENSURE_OK(context, bw_pass_status);
return kTfLiteOk;
}

6
tensorflow/contrib/lite/kernels/lstm.cc
View File

@ -497,6 +497,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
/*aux_input_to_output_weights=*/nullptr, input_gate_bias,
forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
projection_bias, params, /*forward_sequence=*/true,
/*time_major=*/true,
/*output_offset=*/0, scratch_buffer, activation_state, cell_state,
output);
}
@ -524,8 +525,9 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
/*aux_input_to_output_weights=*/nullptr, input_gate_bias,
forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
projection_bias, params, /*forward_sequence=*/true,
/*output_offset=*/0, scratch_buffer, scaling_factors,
prod_scaling_factors, recovered_cell_weights, input_quantized,
/*time_major=*/true, /*output_offset=*/0, scratch_buffer,
scaling_factors, prod_scaling_factors, recovered_cell_weights,
input_quantized,
/*aux_input_quantized=*/nullptr, activation_state_quantized,
cell_state_quantized, activation_state, cell_state, output);
}

227
tensorflow/contrib/lite/kernels/lstm_eval.cc
View File

@ -710,9 +710,10 @@ TfLiteStatus EvalFloat(
const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
const TfLiteLSTMParams* params, bool forward_sequence, int output_offset,
TfLiteTensor* scratch_buffer, TfLiteTensor* activation_state,
TfLiteTensor* cell_state, TfLiteTensor* output) {
const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
int output_offset, TfLiteTensor* scratch_buffer,
TfLiteTensor* activation_state, TfLiteTensor* cell_state,
TfLiteTensor* output) {
TF_LITE_ASSERT(input->dims->size >= 2 && input->dims->size <= 3);
const int max_time = (input->dims->size == 2) ? 1 : input->dims->data[0];
const int n_batch = input->dims->data[input->dims->size - 2];
@ -777,36 +778,71 @@ TfLiteStatus EvalFloat(
aux_input_to_output_weights_ptr = aux_input_to_output_weights->data.f;
}
// Loop through the sequence.
const int output_batch_leading_dim =
output->dims->data[output->dims->size - 1];
const int input_step = n_batch * n_input;
const int output_step = n_batch * output_batch_leading_dim;
for (int t = 0; t < max_time; t++) {
// If this is the forward_sequence, step forward, otherwise step backwards.
const int t_rel = forward_sequence ? t : max_time - t - 1;
const float* input_ptr = input->data.f + t_rel * input_step;
if (aux_input) {
aux_input_ptr = aux_input->data.f + t_rel * input_step;
}
float* output_ptr_time =
output->data.f + t_rel * output_step + output_offset;
if (time_major) {
// Loop through the sequence.
const int input_step = n_batch * n_input;
const int output_step = n_batch * output_batch_leading_dim;
for (int t = 0; t < max_time; t++) {
// If this is the forward_sequence, step forward, otherwise step
// backwards.
const int t_rel = forward_sequence ? t : max_time - t - 1;
const float* input_ptr = input->data.f + t_rel * input_step;
if (aux_input) {
aux_input_ptr = aux_input->data.f + t_rel * input_step;
}
float* output_ptr_time =
output->data.f + t_rel * output_step + output_offset;
LstmStepWithAuxInput(
input_ptr, input_to_input_weights_ptr, input_to_forget_weights->data.f,
input_to_cell_weights->data.f, input_to_output_weights->data.f,
aux_input_ptr, aux_input_to_input_weights_ptr,
aux_input_to_forget_weights_ptr, aux_input_to_cell_weights_ptr,
aux_input_to_output_weights_ptr, recurrent_to_input_weights_ptr,
recurrent_to_forget_weights->data.f, recurrent_to_cell_weights->data.f,
recurrent_to_output_weights->data.f, cell_to_input_weights_ptr,
cell_to_forget_weights_ptr, cell_to_output_weights_ptr,
input_gate_bias_ptr, forget_gate_bias->data.f, cell_bias->data.f,
output_gate_bias->data.f, projection_weights_ptr, projection_bias_ptr,
params, n_batch, n_cell, n_input, aux_input_size, n_output,
output_batch_leading_dim, activation_state->data.f, cell_state->data.f,
input_gate_scratch, forget_gate_scratch, cell_scratch,
output_gate_scratch, output_ptr_time);
LstmStepWithAuxInput(
input_ptr, input_to_input_weights_ptr,
input_to_forget_weights->data.f, input_to_cell_weights->data.f,
input_to_output_weights->data.f, aux_input_ptr,
aux_input_to_input_weights_ptr, aux_input_to_forget_weights_ptr,
aux_input_to_cell_weights_ptr, aux_input_to_output_weights_ptr,
recurrent_to_input_weights_ptr, recurrent_to_forget_weights->data.f,
recurrent_to_cell_weights->data.f,
recurrent_to_output_weights->data.f, cell_to_input_weights_ptr,
cell_to_forget_weights_ptr, cell_to_output_weights_ptr,
input_gate_bias_ptr, forget_gate_bias->data.f, cell_bias->data.f,
output_gate_bias->data.f, projection_weights_ptr, projection_bias_ptr,
params, n_batch, n_cell, n_input, aux_input_size, n_output,
output_batch_leading_dim, activation_state->data.f,
cell_state->data.f, input_gate_scratch, forget_gate_scratch,
cell_scratch, output_gate_scratch, output_ptr_time);
}
} else {
for (int b = 0; b < n_batch; b++) {
const int input_step = n_input;
const int output_step = output_batch_leading_dim;
for (int t = 0; t < max_time; t++) {
// If this is the forward_sequence, step forward, otherwise step
// backwards.
const int t_rel = forward_sequence ? t : max_time - t - 1;
const float* input_ptr = input->data.f + t_rel * input_step;
float* output_ptr_time =
output->data.f + t_rel * output_step + output_offset;
LstmStepWithAuxInput(
input_ptr, input_to_input_weights_ptr,
input_to_forget_weights->data.f, input_to_cell_weights->data.f,
input_to_output_weights->data.f, aux_input_ptr,
aux_input_to_input_weights_ptr, aux_input_to_forget_weights_ptr,
aux_input_to_cell_weights_ptr, aux_input_to_output_weights_ptr,
recurrent_to_input_weights_ptr, recurrent_to_forget_weights->data.f,
recurrent_to_cell_weights->data.f,
recurrent_to_output_weights->data.f, cell_to_input_weights_ptr,
cell_to_forget_weights_ptr, cell_to_output_weights_ptr,
input_gate_bias_ptr, forget_gate_bias->data.f, cell_bias->data.f,
output_gate_bias->data.f, projection_weights_ptr,
projection_bias_ptr, params, /*n_batch=*/1, n_cell, n_input,
aux_input_size, n_output, output_batch_leading_dim,
activation_state->data.f, cell_state->data.f, input_gate_scratch,
forget_gate_scratch, cell_scratch, output_gate_scratch,
output_ptr_time);
}
}
}
return kTfLiteOk;
}
@ -830,13 +866,13 @@ TfLiteStatus EvalHybrid(
const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
const TfLiteLSTMParams* params, bool forward_sequence, int output_offset,
TfLiteTensor* scratch_buffer, TfLiteTensor* scaling_factors,
TfLiteTensor* prod_scaling_factors, TfLiteTensor* recovered_cell_weights,
TfLiteTensor* input_quantized, TfLiteTensor* aux_input_quantized,
TfLiteTensor* output_state_quantized, TfLiteTensor* cell_state_quantized,
TfLiteTensor* output_state, TfLiteTensor* cell_state,
TfLiteTensor* output) {
const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
int output_offset, TfLiteTensor* scratch_buffer,
TfLiteTensor* scaling_factors, TfLiteTensor* prod_scaling_factors,
TfLiteTensor* recovered_cell_weights, TfLiteTensor* input_quantized,
TfLiteTensor* aux_input_quantized, TfLiteTensor* output_state_quantized,
TfLiteTensor* cell_state_quantized, TfLiteTensor* output_state,
TfLiteTensor* cell_state, TfLiteTensor* output) {
TF_LITE_ASSERT(input->dims->size >= 2 && input->dims->size <= 3);
const int max_time = (input->dims->size == 2) ? 1 : input->dims->data[0];
const int n_batch = input->dims->data[input->dims->size - 2];
@ -990,45 +1026,90 @@ TfLiteStatus EvalHybrid(
aux_input_to_output_weights->params.scale;
}
// Feed the sequence into the LSTM step-by-step.
const int output_batch_leading_dim =
output->dims->data[output->dims->size - 1];
const int input_step = n_batch * n_input;
const int output_step = n_batch * output_batch_leading_dim;
for (int t = 0; t < max_time; t++) {
// If this is the forward_sequence, step forward, otherwise step backwards.
const int t_rel = forward_sequence ? t : max_time - t - 1;
const float* input_ptr = input->data.f + t_rel * input_step;
if (aux_input) {
aux_input_ptr = aux_input->data.f + t_rel * input_step;
}
float* output_ptr = output->data.f + t_rel * output_step + output_offset;
if (time_major) {
// Feed the sequence into the LSTM step-by-step.
const int input_step = n_batch * n_input;
const int output_step = n_batch * output_batch_leading_dim;
for (int t = 0; t < max_time; t++) {
// If this is the forward_sequence, step forward, otherwise step
// backwards.
const int t_rel = forward_sequence ? t : max_time - t - 1;
const float* input_ptr = input->data.f + t_rel * input_step;
if (aux_input) {
aux_input_ptr = aux_input->data.f + t_rel * input_step;
}
float* output_ptr = output->data.f + t_rel * output_step + output_offset;
LstmStepWithAuxInput(
input_ptr, input_to_input_weights_ptr, input_to_input_weights_scale,
input_to_forget_weights_ptr, input_to_forget_weights_scale,
input_to_cell_weights_ptr, input_to_cell_weights_scale,
input_to_output_weights_ptr, input_to_output_weights_scale,
aux_input_ptr, aux_input_to_input_weights_ptr,
aux_input_to_input_weights_scale, aux_input_to_forget_weights_ptr,
aux_input_to_forget_weights_scale, aux_input_to_cell_weights_ptr,
aux_input_to_cell_weights_scale, aux_input_to_output_weights_ptr,
aux_input_to_output_weights_scale, recurrent_to_input_weights_ptr,
recurrent_to_input_weights_scale, recurrent_to_forget_weights_ptr,
recurrent_to_forget_weights_scale, recurrent_to_cell_weights_ptr,
recurrent_to_cell_weights_scale, recurrent_to_output_weights_ptr,
recurrent_to_output_weights_scale, cell_to_input_weights_ptr,
cell_to_input_weights_scale, cell_to_forget_weights_ptr,
cell_to_forget_weights_scale, cell_to_output_weights_ptr,
cell_to_output_weights_scale, input_gate_bias_ptr, forget_gate_bias_ptr,
cell_bias_ptr, output_gate_bias_ptr, projection_weights_ptr,
projection_weights_scale, projection_bias_ptr, params, n_batch, n_cell,
n_input, aux_input_size, n_output, output_batch_leading_dim,
input_gate_scratch, forget_gate_scratch, cell_scratch,
output_gate_scratch, scaling_factors_ptr, prod_scaling_factors_ptr,
recovered_cell_weights_ptr, quantized_input_ptr,
quantized_aux_input_ptr, quantized_output_state_ptr,
quantized_cell_state_ptr, output_state_ptr, cell_state_ptr, output_ptr);
LstmStepWithAuxInput(
input_ptr, input_to_input_weights_ptr, input_to_input_weights_scale,
input_to_forget_weights_ptr, input_to_forget_weights_scale,
input_to_cell_weights_ptr, input_to_cell_weights_scale,
input_to_output_weights_ptr, input_to_output_weights_scale,
aux_input_ptr, aux_input_to_input_weights_ptr,
aux_input_to_input_weights_scale, aux_input_to_forget_weights_ptr,
aux_input_to_forget_weights_scale, aux_input_to_cell_weights_ptr,
aux_input_to_cell_weights_scale, aux_input_to_output_weights_ptr,
aux_input_to_output_weights_scale, recurrent_to_input_weights_ptr,
recurrent_to_input_weights_scale, recurrent_to_forget_weights_ptr,
recurrent_to_forget_weights_scale, recurrent_to_cell_weights_ptr,
recurrent_to_cell_weights_scale, recurrent_to_output_weights_ptr,
recurrent_to_output_weights_scale, cell_to_input_weights_ptr,
cell_to_input_weights_scale, cell_to_forget_weights_ptr,
cell_to_forget_weights_scale, cell_to_output_weights_ptr,
cell_to_output_weights_scale, input_gate_bias_ptr,
forget_gate_bias_ptr, cell_bias_ptr, output_gate_bias_ptr,
projection_weights_ptr, projection_weights_scale, projection_bias_ptr,
params, n_batch, n_cell, n_input, aux_input_size, n_output,
output_batch_leading_dim, input_gate_scratch, forget_gate_scratch,
cell_scratch, output_gate_scratch, scaling_factors_ptr,
prod_scaling_factors_ptr, recovered_cell_weights_ptr,
quantized_input_ptr, quantized_aux_input_ptr,
quantized_output_state_ptr, quantized_cell_state_ptr,
output_state_ptr, cell_state_ptr, output_ptr);
}
} else {
for (int b = 0; b < n_batch; b++) {
const int input_step = n_input;
const int output_step = output_batch_leading_dim;
for (int t = 0; t < max_time; t++) {
// If this is the forward_sequence, step forward, otherwise step
// backwards.
const int t_rel = forward_sequence ? t : max_time - t - 1;
const float* input_ptr = input->data.f + t_rel * input_step;
float* output_ptr =
output->data.f + t_rel * output_step + output_offset;
LstmStepWithAuxInput(
input_ptr, input_to_input_weights_ptr, input_to_input_weights_scale,
input_to_forget_weights_ptr, input_to_forget_weights_scale,
input_to_cell_weights_ptr, input_to_cell_weights_scale,
input_to_output_weights_ptr, input_to_output_weights_scale,
aux_input_ptr, aux_input_to_input_weights_ptr,
aux_input_to_input_weights_scale, aux_input_to_forget_weights_ptr,
aux_input_to_forget_weights_scale, aux_input_to_cell_weights_ptr,
aux_input_to_cell_weights_scale, aux_input_to_output_weights_ptr,
aux_input_to_output_weights_scale, recurrent_to_input_weights_ptr,
recurrent_to_input_weights_scale, recurrent_to_forget_weights_ptr,
recurrent_to_forget_weights_scale, recurrent_to_cell_weights_ptr,
recurrent_to_cell_weights_scale, recurrent_to_output_weights_ptr,
recurrent_to_output_weights_scale, cell_to_input_weights_ptr,
cell_to_input_weights_scale, cell_to_forget_weights_ptr,
cell_to_forget_weights_scale, cell_to_output_weights_ptr,
cell_to_output_weights_scale, input_gate_bias_ptr,
forget_gate_bias_ptr, cell_bias_ptr, output_gate_bias_ptr,
projection_weights_ptr, projection_weights_scale,
projection_bias_ptr, params, n_batch, n_cell, n_input,
aux_input_size, n_output, output_batch_leading_dim,
input_gate_scratch, forget_gate_scratch, cell_scratch,
output_gate_scratch, scaling_factors_ptr, prod_scaling_factors_ptr,
recovered_cell_weights_ptr, quantized_input_ptr,
quantized_aux_input_ptr, quantized_output_state_ptr,
quantized_cell_state_ptr, output_state_ptr, cell_state_ptr,
output_ptr);
}
}
}
return kTfLiteOk;

20
tensorflow/contrib/lite/kernels/lstm_eval.h
View File

@ -42,9 +42,10 @@ TfLiteStatus EvalFloat(
const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
const TfLiteLSTMParams* params, bool forward_sequence, int output_offset,
TfLiteTensor* scratch_buffer, TfLiteTensor* activation_state,
TfLiteTensor* cell_state, TfLiteTensor* output);
const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
int output_offset, TfLiteTensor* scratch_buffer,
TfLiteTensor* activation_state, TfLiteTensor* cell_state,
TfLiteTensor* output);
TfLiteStatus EvalHybrid(
const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights,
@ -65,12 +66,13 @@ TfLiteStatus EvalHybrid(
const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
const TfLiteLSTMParams* params, bool forward_sequence, int output_offset,
TfLiteTensor* scratch_buffer, TfLiteTensor* scaling_factors,
TfLiteTensor* prod_scaling_factors, TfLiteTensor* recovered_cell_weights,
TfLiteTensor* input_quantized, TfLiteTensor* aux_input_quantized,
TfLiteTensor* output_state_quantized, TfLiteTensor* cell_state_quantized,
TfLiteTensor* output_state, TfLiteTensor* cell_state, TfLiteTensor* output);
const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
int output_offset, TfLiteTensor* scratch_buffer,
TfLiteTensor* scaling_factors, TfLiteTensor* prod_scaling_factors,
TfLiteTensor* recovered_cell_weights, TfLiteTensor* input_quantized,
TfLiteTensor* aux_input_quantized, TfLiteTensor* output_state_quantized,
TfLiteTensor* cell_state_quantized, TfLiteTensor* output_state,
TfLiteTensor* cell_state, TfLiteTensor* output);
} // namespace lstm_eval
} // namespace builtin

18
tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc
View File

@ -260,8 +260,11 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
const TfLiteTensor* input = GetInput(context, node, kInputTensor);
TF_LITE_ENSURE_EQ(context, input->type, kTfLiteFloat32);
TF_LITE_ENSURE(context, input->dims->size > 1);
const int max_time = input->dims->data[0];
const int n_batch = input->dims->data[1];
const auto* params =
reinterpret_cast<TfLiteUnidirectionalSequenceLSTMParams*>(
node->builtin_data);
const bool time_major = params->time_major;
const int n_batch = time_major ? input->dims->data[1] : input->dims->data[0];
const int n_input = input->dims->data[2];
const TfLiteTensor* input_to_output_weights =
@ -296,10 +299,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
TF_LITE_ENSURE_EQ(context, NumElements(cell_state), n_batch * n_cell);
// Resize the output tensors.
TfLiteIntArray* output_size = TfLiteIntArrayCreate(3);
output_size->data[0] = max_time;
output_size->data[1] = n_batch;
output_size->data[2] = n_output;
TfLiteIntArray* output_size = TfLiteIntArrayCopy(input->dims);
output_size->data[input->dims->size - 1] = n_output;
TF_LITE_ENSURE_OK(context,
context->ResizeTensor(context, output, output_size));
@ -436,6 +437,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
const auto* params =
reinterpret_cast<TfLiteUnidirectionalSequenceLSTMParams*>(
node->builtin_data);
const bool time_major = params->time_major;
const TfLiteTensor* input = GetInput(context, node, kInputTensor);
const TfLiteTensor* input_to_input_weights =
@ -506,7 +508,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
/*aux_input_to_cell_weights=*/nullptr,
/*aux_input_to_output_weights=*/nullptr, input_gate_bias,
forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
projection_bias, &lstm_params, /*forward_sequence=*/true,
projection_bias, &lstm_params, /*forward_sequence=*/true, time_major,
/*output_offset=*/0, scratch_buffer, activation_state, cell_state,
output);
}
@ -533,7 +535,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
/*aux_input_to_cell_weights=*/nullptr,
/*aux_input_to_output_weights=*/nullptr, input_gate_bias,
forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
projection_bias, &lstm_params, /*forward_sequence=*/true,
projection_bias, &lstm_params, /*forward_sequence=*/true, time_major,
/*output_offset=*/0, scratch_buffer, scaling_factors,
prod_scaling_factors, recovered_cell_weights, input_quantized,
/*aux_input_quantized=*/nullptr, activation_state_quantized,

152
tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm_test.cc
View File

@ -32,7 +32,7 @@ using ::testing::ElementsAreArray;
class UnidirectionalLSTMOpModel : public SingleOpModel {
public:
UnidirectionalLSTMOpModel(int n_batch, int n_input, int n_cell, int n_output,
int sequence_length, bool use_cifg,
int sequence_length, bool time_major, bool use_cifg,
bool use_peephole, bool use_projection_weights,
bool use_projection_bias, float cell_clip,
float proj_clip,
@ -110,12 +110,12 @@ class UnidirectionalLSTMOpModel : public SingleOpModel {
output_ = AddOutput(TensorType_FLOAT32);
SetBuiltinOp(
BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM,
BuiltinOptions_UnidirectionalSequenceLSTMOptions,
CreateUnidirectionalSequenceLSTMOptions(
builder_, ActivationFunctionType_TANH, cell_clip, proj_clip)
.Union());
SetBuiltinOp(BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM,
BuiltinOptions_UnidirectionalSequenceLSTMOptions,
CreateUnidirectionalSequenceLSTMOptions(
builder_, ActivationFunctionType_TANH, cell_clip,
proj_clip, time_major)
.Union());
BuildInterpreter(input_shapes);
}
@ -241,12 +241,12 @@ class HybridUnidirectionalLSTMOpModel : public UnidirectionalLSTMOpModel {
public:
HybridUnidirectionalLSTMOpModel(
int n_batch, int n_input, int n_cell, int n_output, int sequence_length,
bool use_cifg, bool use_peephole, bool use_projection_weights,
bool use_projection_bias, float cell_clip, float proj_clip,
const std::vector<std::vector<int>>& input_shapes)
bool time_major, bool use_cifg, bool use_peephole,
bool use_projection_weights, bool use_projection_bias, float cell_clip,
float proj_clip, const std::vector<std::vector<int>>& input_shapes)
: UnidirectionalLSTMOpModel(
n_batch, n_input, n_cell, n_output, sequence_length, use_cifg,
use_peephole, use_projection_weights, use_projection_bias,
n_batch, n_input, n_cell, n_output, sequence_length, time_major,
use_cifg, use_peephole, use_projection_weights, use_projection_bias,
cell_clip, proj_clip, input_shapes, TensorType_UINT8) {}
void SetInputToInputWeights(const std::vector<float>& f) {
@ -326,21 +326,32 @@ class BaseLstmTest : public ::testing::Test {
// Compares output up to tolerance to the result of the lstm given the input.
void VerifyGoldens(const std::vector<std::vector<float>>& input,
const std::vector<std::vector<float>>& output,
UnidirectionalLSTMOpModel* lstm, float tolerance = 1e-5) {
UnidirectionalLSTMOpModel* lstm, float tolerance = 1e-5,
bool time_major = true) {
const int num_batches = input.size();
EXPECT_GT(num_batches, 0);
const int num_inputs = lstm->num_inputs();
EXPECT_GT(num_inputs, 0);
const int input_sequence_size = input[0].size() / num_inputs;
EXPECT_GT(input_sequence_size, 0);
// Feed the whole sequence as input.
for (int i = 0; i < input_sequence_size; ++i) {
for (int b = 0; b < num_batches; ++b) {
const float* batch_start = input[b].data() + i * num_inputs;
const float* batch_end = batch_start + num_inputs;
if (time_major) {
// Feed the whole sequence as input.
for (int i = 0; i < input_sequence_size; ++i) {
for (int b = 0; b < num_batches; ++b) {
const float* batch_start = input[b].data() + i * num_inputs;
const float* batch_end = batch_start + num_inputs;
lstm->SetInput(((i * num_batches) + b) * lstm->num_inputs(),
batch_start, batch_end);
lstm->SetInput(((i * num_batches) + b) * num_inputs, batch_start,
batch_end);
}
}
} else {
for (int b = 0; b < num_batches; ++b) {
const float* batch_start = input[b].data();
const float* batch_end = batch_start + input_sequence_size * num_inputs;
lstm->SetInput(b * input_sequence_size * num_inputs, batch_start,
batch_end);
}
}
@ -349,15 +360,25 @@ class BaseLstmTest : public ::testing::Test {
const int num_outputs = lstm->num_outputs();
EXPECT_GT(num_outputs, 0);
std::vector<float> expected;
for (int i = 0; i < input_sequence_size; ++i) {
for (int b = 0; b < num_batches; ++b) {
const float* golden_start_batch = output[b].data() + i * num_outputs;
const float* golden_end_batch = golden_start_batch + num_outputs;
expected.insert(expected.end(), golden_start_batch, golden_end_batch);
if (time_major) {
for (int i = 0; i < input_sequence_size; ++i) {
for (int b = 0; b < num_batches; ++b) {
const float* golden_start_batch = output[b].data() + i * num_outputs;
const float* golden_end_batch = golden_start_batch + num_outputs;
expected.insert(expected.end(), golden_start_batch, golden_end_batch);
}
}
} else {
for (int b = 0; b < num_batches; ++b) {
const float* golden_batch_start = output[b].data();
const float* golden_batch_end =
golden_batch_start + input_sequence_size * num_outputs;
expected.insert(expected.end(), golden_batch_start, golden_batch_end);
}
}
EXPECT_THAT(lstm->GetOutput(),
ElementsAreArray(ArrayFloatNear(expected, tolerance)));
}
@ -422,7 +443,7 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) {
UnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*use_cifg=*/false, /*use_peephole=*/false,
/*time_major=*/true, /*use_cifg=*/false, /*use_peephole=*/false,
/*use_projection_weights=*/false,
/*use_projection_bias=*/false,
/*cell_clip=*/0.0, /*proj_clip=*/0.0,
@ -473,6 +494,73 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) {
VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm);
}
TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest,
LstmBlackBoxTestBatchMajor) {
const int n_batch = 1;
const int n_input = 2;
// n_cell and n_output have the same size when there is no projection.
const int n_cell = 4;
const int n_output = 4;
const int sequence_length = 3;
UnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*time_major=*/true, /*use_cifg=*/false, /*use_peephole=*/false,
/*use_projection_weights=*/false,
/*use_projection_bias=*/false,
/*cell_clip=*/0.0, /*proj_clip=*/0.0,
{
{sequence_length, n_batch, n_input}, // input tensor
{n_cell, n_input}, // input_to_input_weight tensor
{n_cell, n_input}, // input_to_forget_weight tensor
{n_cell, n_input}, // input_to_cell_weight tensor
{n_cell, n_input}, // input_to_output_weight tensor
{n_cell, n_output}, // recurrent_to_input_weight tensor
{n_cell, n_output}, // recurrent_to_forget_weight tensor
{n_cell, n_output}, // recurrent_to_cell_weight tensor
{n_cell, n_output}, // recurrent_to_output_weight tensor
{0}, // cell_to_input_weight tensor
{0}, // cell_to_forget_weight tensor
{0}, // cell_to_output_weight tensor
{n_cell}, // input_gate_bias tensor
{n_cell}, // forget_gate_bias tensor
{n_cell}, // cell_bias tensor
{n_cell}, // output_gate_bias tensor
{0, 0}, // projection_weight tensor
{0}, // projection_bias tensor
{n_batch, n_output}, // activation_state tensor
{n_batch, n_cell}, // cell_state tensor
});
lstm.SetInputToInputWeights(input_to_input_weights_);
lstm.SetInputToCellWeights(input_to_cell_weights_);
lstm.SetInputToForgetWeights(input_to_forget_weights_);
lstm.SetInputToOutputWeights(input_to_output_weights_);
lstm.SetInputGateBias(input_gate_bias_);
lstm.SetCellBias(cell_gate_bias_);
lstm.SetForgetGateBias(forget_gate_bias_);
lstm.SetOutputGateBias(output_gate_bias_);
lstm.SetRecurrentToInputWeights(recurrent_to_input_weights_);
lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_);
lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_);
lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_);
// Reshuffle input and output to batch major format.
std::vector<std::vector<float>> input;
std::vector<std::vector<float>> output;
VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm, /*tolerance=*/1e-5,
/*time_major=*/false);
}
TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, HybridLstmBlackBoxTest) {
const int n_batch = 1;
const int n_input = 2;
@ -483,7 +571,7 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, HybridLstmBlackBoxTest) {
HybridUnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*use_cifg=*/false, /*use_peephole=*/false,
/*time_major=*/true, /*use_cifg=*/false, /*use_peephole=*/false,
/*use_projection_weights=*/false,
/*use_projection_bias=*/false, /*cell_clip=*/0.0, /*proj_clip=*/0.0,
{
@ -591,7 +679,7 @@ TEST_F(CifgPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) {
UnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*use_cifg=*/true, /*use_peephole=*/true,
/*time_major=*/true, /*use_cifg=*/true, /*use_peephole=*/true,
/*use_projection_weights=*/false,
/*use_projection_bias=*/false,
/*cell_clip=*/0.0, /*proj_clip=*/0.0,
@ -652,7 +740,7 @@ TEST_F(CifgPeepholeNoProjectionNoClippingLstmTest, HybridLstmBlackBoxTest) {
HybridUnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*use_cifg=*/true, /*use_peephole=*/true,
/*time_major=*/true, /*use_cifg=*/true, /*use_peephole=*/true,
/*use_projection_weights=*/false,
/*use_projection_bias=*/false,
/*cell_clip=*/0.0, /*proj_clip=*/0.0,
@ -1311,7 +1399,7 @@ TEST_F(NoCifgPeepholeProjectionClippingLstmTest, LstmBlackBoxTest) {
UnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*use_cifg=*/false, /*use_peephole=*/true,
/*time_major=*/true, /*use_cifg=*/false, /*use_peephole=*/true,
/*use_projection_weights=*/true,
/*use_projection_bias=*/false,
/*cell_clip=*/0.0, /*proj_clip=*/0.0,
@ -1377,7 +1465,7 @@ TEST_F(NoCifgPeepholeProjectionClippingLstmTest, HybridLstmBlackBoxTest) {
HybridUnidirectionalLSTMOpModel lstm(
n_batch, n_input, n_cell, n_output, sequence_length,
/*use_cifg=*/false, /*use_peephole=*/true,
/*time_major=*/true, /*use_cifg=*/false, /*use_peephole=*/true,
/*use_projection_weights=*/true,
/*use_projection_bias=*/false,
/*cell_clip=*/0.0, /*proj_clip=*/0.0,

3
tensorflow/contrib/lite/schema/schema.fbs
View File

@ -407,6 +407,9 @@ table UnidirectionalSequenceLSTMOptions {
fused_activation_function:ActivationFunctionType;
cell_clip: float; // Optional, 0.0 means no clipping
proj_clip: float; // Optional, 0.0 means no clipping
// If true then first dimension is sequence, otherwise batch.
time_major:bool;
}
table BidirectionalSequenceLSTMOptions {

23
tensorflow/contrib/lite/schema/schema_generated.h
View File

@ -3534,10 +3534,12 @@ struct UnidirectionalSequenceLSTMOptionsT : public flatbuffers::NativeTable {
ActivationFunctionType fused_activation_function;
float cell_clip;
float proj_clip;
bool time_major;
UnidirectionalSequenceLSTMOptionsT()
: fused_activation_function(ActivationFunctionType_NONE),
cell_clip(0.0f),
proj_clip(0.0f) {
proj_clip(0.0f),
time_major(false) {
}
};
@ -3546,7 +3548,8 @@ struct UnidirectionalSequenceLSTMOptions FLATBUFFERS_FINAL_CLASS : private flatb
enum {
VT_FUSED_ACTIVATION_FUNCTION = 4,
VT_CELL_CLIP = 6,
VT_PROJ_CLIP = 8
VT_PROJ_CLIP = 8,
VT_TIME_MAJOR = 10
};
ActivationFunctionType fused_activation_function() const {
return static_cast<ActivationFunctionType>(GetField<int8_t>(VT_FUSED_ACTIVATION_FUNCTION, 0));
@ -3557,11 +3560,15 @@ struct UnidirectionalSequenceLSTMOptions FLATBUFFERS_FINAL_CLASS : private flatb
float proj_clip() const {
return GetField<float>(VT_PROJ_CLIP, 0.0f);
}
bool time_major() const {
return GetField<uint8_t>(VT_TIME_MAJOR, 0) != 0;
}
bool Verify(flatbuffers::Verifier &verifier) const {
return VerifyTableStart(verifier) &&
VerifyField<int8_t>(verifier, VT_FUSED_ACTIVATION_FUNCTION) &&
VerifyField<float>(verifier, VT_CELL_CLIP) &&
VerifyField<float>(verifier, VT_PROJ_CLIP) &&
VerifyField<uint8_t>(verifier, VT_TIME_MAJOR) &&
verifier.EndTable();
}
UnidirectionalSequenceLSTMOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const;
@ -3581,6 +3588,9 @@ struct UnidirectionalSequenceLSTMOptionsBuilder {
void add_proj_clip(float proj_clip) {
fbb_.AddElement<float>(UnidirectionalSequenceLSTMOptions::VT_PROJ_CLIP, proj_clip, 0.0f);
}
void add_time_major(bool time_major) {
fbb_.AddElement<uint8_t>(UnidirectionalSequenceLSTMOptions::VT_TIME_MAJOR, static_cast<uint8_t>(time_major), 0);
}
explicit UnidirectionalSequenceLSTMOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb)
: fbb_(_fbb) {
start_ = fbb_.StartTable();
@ -3597,10 +3607,12 @@ inline flatbuffers::Offset<UnidirectionalSequenceLSTMOptions> CreateUnidirection
flatbuffers::FlatBufferBuilder &_fbb,
ActivationFunctionType fused_activation_function = ActivationFunctionType_NONE,
float cell_clip = 0.0f,
float proj_clip = 0.0f) {
float proj_clip = 0.0f,
bool time_major = false) {
UnidirectionalSequenceLSTMOptionsBuilder builder_(_fbb);
builder_.add_proj_clip(proj_clip);
builder_.add_cell_clip(cell_clip);
builder_.add_time_major(time_major);
builder_.add_fused_activation_function(fused_activation_function);
return builder_.Finish();
}
@ -8060,6 +8072,7 @@ inline void UnidirectionalSequenceLSTMOptions::UnPackTo(UnidirectionalSequenceLS
{ auto _e = fused_activation_function(); _o->fused_activation_function = _e; };
{ auto _e = cell_clip(); _o->cell_clip = _e; };
{ auto _e = proj_clip(); _o->proj_clip = _e; };
{ auto _e = time_major(); _o->time_major = _e; };
}
inline flatbuffers::Offset<UnidirectionalSequenceLSTMOptions> UnidirectionalSequenceLSTMOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const UnidirectionalSequenceLSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) {
@ -8073,11 +8086,13 @@ inline flatbuffers::Offset<UnidirectionalSequenceLSTMOptions> CreateUnidirection
auto _fused_activation_function = _o->fused_activation_function;
auto _cell_clip = _o->cell_clip;
auto _proj_clip = _o->proj_clip;
auto _time_major = _o->time_major;
return tflite::CreateUnidirectionalSequenceLSTMOptions(
_fbb,
_fused_activation_function,
_cell_clip,
_proj_clip);
_proj_clip,
_time_major);
}
inline BidirectionalSequenceLSTMOptionsT *BidirectionalSequenceLSTMOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const {