FM / operators

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators

Name	Size	Mode	Actions
abs_op.h	705	0644	edit dl rm
accumulate_op.h	1073	0644	edit dl rm
accuracy_op.h	652	0644	edit dl rm
acos_op.h	711	0644	edit dl rm
activation_ops_cudnn.h	4122	0644	edit dl rm
affine_channel_op.h	3450	0644	edit dl rm
alias_with_name.h	1234	0644	edit dl rm
apmeter_op.h	1027	0644	edit dl rm
arg_ops.h	2319	0644	edit dl rm
asin_op.h	711	0644	edit dl rm
assert_op.h	1335	0644	edit dl rm
async_net_barrier_op.h	904	0644	edit dl rm
atan_op.h	711	0644	edit dl rm
batch_box_cox_op.h	2287	0644	edit dl rm
batch_bucketize_op.h	720	0644	edit dl rm
batch_gather_ops.h	5264	0644	edit dl rm
batch_matmul_op.h	9602	0644	edit dl rm
batch_moments_op.h	3364	0644	edit dl rm
batch_permutation_op.h	954	0644	edit dl rm
batch_sparse_to_dense_op.h	6147	0644	edit dl rm
bbox_transform_op.h	2668	0644	edit dl rm
bisect_percentile_op.h	4921	0644	edit dl rm
boolean_mask_ops.h	2665	0644	edit dl rm
boolean_unmask_ops.h	378	0644	edit dl rm
box_with_nms_limit_op.h	4960	0644	edit dl rm
bucketize_op.h	1361	0644	edit dl rm
byte_weight_dequant_op.h	1722	0644	edit dl rm
cast_op.h	1393	0644	edit dl rm
cbrt_op.h	723	0644	edit dl rm
cc_bmm_bg_op.h	3894	0644	edit dl rm
ceil_op.h	782	0644	edit dl rm
channel_backprop_stats_op.h	737	0644	edit dl rm
channel_shuffle_op.h	1902	0644	edit dl rm
channel_stats_op.h	1807	0644	edit dl rm
clip_op.h	1639	0644	edit dl rm
collect_and_distribute_fpn_rpn_proposals_op.h	6875	0644	edit dl rm
concat_split_op.h	11850	0644	edit dl rm
conditional_op.h	487	0644	edit dl rm
conv_op.h	3125	0644	edit dl rm
conv_op_cache_cudnn.h	1935	0644	edit dl rm
conv_op_impl.h	28729	0644	edit dl rm
conv_op_shared.h	672	0644	edit dl rm
conv_pool_op_base.h	32109	0644	edit dl rm
conv_transpose_op.h	1727	0644	edit dl rm
conv_transpose_op_impl.h	18264	0644	edit dl rm
conv_transpose_op_mobile.h	1470	0644	edit dl rm
conv_transpose_op_mobile_impl.h	19587	0644	edit dl rm
conv_transpose_unpool_op_base.h	10303	0644	edit dl rm
copy_op.h	1296	0644	edit dl rm
copy_rows_to_tensor_op.h	2599	0644	edit dl rm
cosh_op.h	711	0644	edit dl rm
cosine_embedding_criterion_op.h	1127	0644	edit dl rm
cos_op.h	705	0644	edit dl rm
counter_ops.h	4596	0644	edit dl rm
create_scope_op.h	5232	0644	edit dl rm
cross_entropy_op.h	4420	0644	edit dl rm
ctc_beam_search_decoder_op.h	1102	0644	edit dl rm
ctc_greedy_decoder_op.h	817	0644	edit dl rm
cube_op.h	723	0644	edit dl rm
dataset_ops.h	5501	0644	edit dl rm
data_couple.h	464	0644	edit dl rm
deform_conv_op.h	3543	0644	edit dl rm
deform_conv_op_impl.h	13171	0644	edit dl rm
dense_vector_to_id_list_op.h	1797	0644	edit dl rm
distance_op.h	8419	0644	edit dl rm
do_op.h	6981	0644	edit dl rm
dropout_op.h	1516	0644	edit dl rm
elementwise_add_op.h	2024	0644	edit dl rm
elementwise_div_op.h	1224	0644	edit dl rm
elementwise_linear_op.h	1170	0644	edit dl rm
elementwise_logical_ops.h	5083	0644	edit dl rm
elementwise_mul_op.h	1224	0644	edit dl rm
elementwise_ops.h	19115	0644	edit dl rm
elementwise_ops_utils.h	1008	0644	edit dl rm
elementwise_op_test.h	9237	0644	edit dl rm
elementwise_sub_op.h	2025	0644	edit dl rm
elu_op.h	875	0644	edit dl rm
enforce_finite_op.h	2303	0644	edit dl rm
ensure_clipped_op.h	1608	0644	edit dl rm
ensure_cpu_output_op.h	1465	0644	edit dl rm
erf_op.h	751	0644	edit dl rm
expand_op.h	3877	0644	edit dl rm
expand_squeeze_dims_op.h	3451	0644	edit dl rm
exp_op.h	425	0644	edit dl rm
fc_inference.h	775	0644	edit dl rm
feature_maps_ops.h	32437	0644	edit dl rm
feed_blob_op.h	802	0644	edit dl rm
filler_op.h	18431	0644	edit dl rm
find_duplicate_elements_op.h	1563	0644	edit dl rm
find_op.h	2055	0644	edit dl rm
flatten_op.h	1525	0644	edit dl rm
flexible_top_k.h	936	0644	edit dl rm
floor_op.h	788	0644	edit dl rm
free_op.h	777	0644	edit dl rm
fully_connected_op.h	9351	0644	edit dl rm
fused_rowwise_8bit_conversion_ops.h	6601	0644	edit dl rm
fused_rowwise_nbitfake_conversion_ops.h	4375	0644	edit dl rm
fused_rowwise_nbit_conversion_ops.h	8723	0644	edit dl rm
fused_rowwise_random_quantization_ops.h	2607	0644	edit dl rm
gather_fused_8bit_rowwise_op.h	2179	0644	edit dl rm
gather_op.h	7505	0644	edit dl rm
gather_ranges_to_dense_op.h	8188	0644	edit dl rm
gelu_op.h	1452	0644	edit dl rm
generate_proposals_op.h	6256	0644	edit dl rm
generate_proposals_op_util_boxes.h	14309	0644	edit dl rm
generate_proposals_op_util_nms.h	26214	0644	edit dl rm
generate_proposals_op_util_nms_gpu.h	2128	0644	edit dl rm
given_tensor_byte_string_to_uint8_fill_op.h	2150	0644	edit dl rm
given_tensor_fill_op.h	3002	0644	edit dl rm
glu_op.h	1458	0644	edit dl rm
group_norm_op.h	8967	0644	edit dl rm
gru_unit_op.h	6626	0644	edit dl rm
half_float_ops.h	2732	0644	edit dl rm
hard_sigmoid_op.h	994	0644	edit dl rm
heatmap_max_keypoint_op.h	939	0644	edit dl rm
histogram_op.h	2421	0644	edit dl rm
h_softmax_op.h	4954	0644	edit dl rm
if_op.h	1764	0644	edit dl rm
im2col_op.h	8943	0644	edit dl rm
index_hash_ops.h	2232	0644	edit dl rm
index_ops.h	3155	0644	edit dl rm
inference_lstm_op.h	9881	0644	edit dl rm
instance_norm_op.h	7441	0644	edit dl rm
integral_image_op.h	923	0644	edit dl rm
is_empty_op.h	558	0644	edit dl rm
jsd_op.h	721	0644	edit dl rm
key_split_ops.h	1400	0644	edit dl rm
layer_norm_op.h	8098	0644	edit dl rm
leaky_relu_op.h	1111	0644	edit dl rm
lengths_pad_op.h	2574	0644	edit dl rm
lengths_reducer_fused_8bit_rowwise_ops.h	5532	0644	edit dl rm
lengths_reducer_fused_nbit_rowwise_ops.h	23465	0644	edit dl rm
lengths_reducer_ops.h	23315	0644	edit dl rm
lengths_reducer_rowwise_8bit_ops.h	6180	0644	edit dl rm
lengths_tile_op.h	582	0644	edit dl rm
lengths_top_k_op.h	1358	0644	edit dl rm
length_split_op.h	2259	0644	edit dl rm
listwise_l2r_op.h	1677	0644	edit dl rm
load_save_op.h	14091	0644	edit dl rm
load_save_op_util.h	1642	0644	edit dl rm
locally_connected_op.h	3872	0644	edit dl rm
locally_connected_op_impl.h	26495	0644	edit dl rm
locally_connected_op_util.h	1332	0644	edit dl rm
local_response_normalization_op.h	2804	0644	edit dl rm
log1p_op.h	717	0644	edit dl rm
logit_op.h	1129	0644	edit dl rm
log_op.h	431	0644	edit dl rm
loss_op.h	1058	0644	edit dl rm
lpnorm_op.h	1279	0644	edit dl rm
lstm_unit_op.h	6733	0644	edit dl rm
lstm_utils.h	9424	0644	edit dl rm
map_ops.h	8011	0644	edit dl rm
margin_ranking_criterion_op.h	1113	0644	edit dl rm
matmul_op.h	2843	0644	edit dl rm
max_pool_with_index_gpu.h	1155	0644	edit dl rm
mean_op.h	3252	0644	edit dl rm
merge_id_lists_op.h	2570	0644	edit dl rm
minmax_ops.h	3829	0644	edit dl rm
mish_op.h	794	0644	edit dl rm
mod_op.h	984	0644	edit dl rm
moments_op.h	4051	0644	edit dl rm
multi_class_accuracy_op.h	539	0644	edit dl rm
negate_gradient_op.h	566	0644	edit dl rm
negative_op.h	451	0644	edit dl rm
ngram_ops.h	2644	0644	edit dl rm
normalize_l1_op.h	1075	0644	edit dl rm
normalize_op.h	3013	0644	edit dl rm
no_default_engine_op.h	1063	0644	edit dl rm
numpy_tile_op.h	3643	0644	edit dl rm
one_hot_ops.h	2562	0644	edit dl rm
onnx_while_op.h	10655	0644	edit dl rm
operator_fallback_gpu.h	4155	0644	edit dl rm
op_utils_cudnn.h	2112	0644	edit dl rm
order_switch_ops.h	2149	0644	edit dl rm
pack_rnn_sequence_op.h	3074	0644	edit dl rm
pack_segments.h	2729	0644	edit dl rm
pad_op.h	2902	0644	edit dl rm
partition_ops.h	9958	0644	edit dl rm
percentile_op.h	1009	0644	edit dl rm
perplexity_op.h	447	0644	edit dl rm
piecewise_linear_transform_op.h	8281	0644	edit dl rm
pool_op.h	8525	0644	edit dl rm
pool_op_util.h	1105	0644	edit dl rm
pow_op.h	4677	0644	edit dl rm
prefetch_op.h	4661	0644	edit dl rm
prelu_op.h	1067	0644	edit dl rm
prepend_dim_op.h	2760	0644	edit dl rm
quantile_op.h	4120	0644	edit dl rm
quant_decode_op.h	5370	0644	edit dl rm
rank_loss_op.h	820	0644	edit dl rm
reciprocal_op.h	721	0644	edit dl rm
reducer_functors.h	24556	0644	edit dl rm
reduce_front_back_max_ops.h	4399	0644	edit dl rm
reduce_front_back_sum_mean_ops.h	5337	0644	edit dl rm
reduce_ops.h	9962	0644	edit dl rm
reduction_ops.h	5944	0644	edit dl rm
relu_n_op.h	990	0644	edit dl rm
relu_op.h	624	0644	edit dl rm
remove_data_blocks_op.h	2651	0644	edit dl rm
replace_nan_op.h	1170	0644	edit dl rm
reshape_op.h	5723	0644	edit dl rm
resize_3d_op.h	2677	0644	edit dl rm
resize_op.h	2307	0644	edit dl rm
reverse_packed_segs_op.h	2772	0644	edit dl rm
rmac_regions_op.h	708	0644	edit dl rm
rms_norm_op.h	2968	0644	edit dl rm
roi_align_gradient_op.h	1486	0644	edit dl rm
roi_align_op.h	2857	0644	edit dl rm
roi_align_rotated_gradient_op.h	1369	0644	edit dl rm
roi_align_rotated_op.h	1636	0644	edit dl rm
roi_pool_op.h	2503	0644	edit dl rm
rowmul_op.h	1947	0644	edit dl rm
rsqrt_op.h	729	0644	edit dl rm
scale_blobs_op.h	1458	0644	edit dl rm
scale_op.h	1019	0644	edit dl rm
segment_reduction_op.h	71022	0644	edit dl rm
self_binning_histogram_op.h	6258	0644	edit dl rm
selu_op.h	1545	0644	edit dl rm
sequence_ops.h	8264	0644	edit dl rm
shape_op.h	1638	0644	edit dl rm
sigmoid_op.h	639	0644	edit dl rm
sinh_op.h	711	0644	edit dl rm
sinusoid_position_encoding_op.h	2834	0644	edit dl rm
sin_op.h	705	0644	edit dl rm
slice_op.h	10071	0644	edit dl rm
softmax_op.h	1174	0644	edit dl rm
softmax_utils.h	447	0644	edit dl rm
softmax_with_loss_op.h	2883	0644	edit dl rm
softplus_op.h	781	0644	edit dl rm
softsign_op.h	675	0644	edit dl rm
space_batch_op.h	6848	0644	edit dl rm
sparse_dropout_with_replacement_op.h	1122	0644	edit dl rm
sparse_itemwise_dropout_with_replacement_op.h	1163	0644	edit dl rm
sparse_lp_regularizer_op.h	1130	0644	edit dl rm
sparse_normalize_op.h	834	0644	edit dl rm
sparse_to_dense_mask_op.h	10051	0644	edit dl rm
sparse_to_dense_op.h	3977	0644	edit dl rm
spatial_batch_norm_op.h	15175	0644	edit dl rm
spatial_softmax_with_loss_op.h	2182	0644	edit dl rm
sqrt_op.h	448	0644	edit dl rm
sqr_op.h	431	0644	edit dl rm
square_root_divide_op.h	1857	0644	edit dl rm
stats_put_ops.h	2813	0644	edit dl rm
stop_gradient.h	548	0644	edit dl rm
string_ops.h	2067	0644	edit dl rm
stump_func_op.h	2112	0644	edit dl rm
summarize_op.h	1875	0644	edit dl rm
swish_op.h	772	0644	edit dl rm
tanh_op.h	723	0644	edit dl rm
tan_op.h	705	0644	edit dl rm
tensor_protos_db_input.h	3633	0644	edit dl rm
text_file_reader_utils.h	2900	0644	edit dl rm
thresholded_relu_op.h	1137	0644	edit dl rm
tile_op.h	8741	0644	edit dl rm
top_k.h	1061	0644	edit dl rm
transpose_op.h	2082	0644	edit dl rm
tt_linear_op.h	6501	0644	edit dl rm
unique_ops.h	1666	0644	edit dl rm
unsafe_coalesce.h	2481	0644	edit dl rm
upsample_op.h	2246	0644	edit dl rm
utility_ops.h	49994	0644	edit dl rm
variable_length_sequence_padding.h	1378	0644	edit dl rm
weighted_multi_sampling_op.h	602	0644	edit dl rm
weighted_sample_op.h	739	0644	edit dl rm
while_op.h	1961	0644	edit dl rm
zero_gradient_op.h	347	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/inference_lstm_op.h (9881B)

#ifndef LSTM_OP_H_
#define LSTM_OP_H_

#include <algorithm>
#include <sstream>
#include <unordered_map>
#include <vector>
#include "caffe2/core/blob_serialization.h"
#include "caffe2/core/export_caffe2_op_to_c10.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor.h"
#include "caffe2/utils/eigen_utils.h"
#include "caffe2/utils/math.h"
#include "lstm_utils.h"

C10_DECLARE_EXPORT_CAFFE2_OP_TO_C10(LSTMOp);

namespace caffe2 {
namespace {

using t_tuple = std::tuple<Tensor, Tensor>;

struct CellParams {
  CellParams(
      const Tensor& _w_ih,
      const Tensor& _w_hh,
      const Tensor& _b_ih,
      const Tensor& _b_hh,
      CPUContext* _context) {
    initParams(_w_ih, _w_hh, _b_ih, _b_hh, _context);
  }

CellParams(const CellParams& rhs) {
    initParams(rhs.w_ih, rhs.w_hh, rhs.b_ih, rhs.b_hh, rhs.context);
  }

CellParams& operator=(const CellParams& rhs) {
    initParams(rhs.w_ih, rhs.w_hh, rhs.b_ih, rhs.b_hh, rhs.context);
    return *this;
  }

void initParams(
      const Tensor& _w_ih,
      const Tensor& _w_hh,
      const Tensor& _b_ih,
      const Tensor& _b_hh,
      CPUContext* _context) {
    w_ih = copy_ctor(_w_ih);
    w_hh = copy_ctor(_w_hh);
    b_ih = copy_ctor(_b_ih);
    b_hh = copy_ctor(_b_hh);
    context = _context;
  }

Tensor w_ih;
  Tensor w_hh;
  Tensor b_ih; /* optional */
  Tensor b_hh; /* optional */
  CPUContext* context;

Tensor linear_ih(const Tensor& input) const {
    return linear(input, w_ih, b_ih, context);
  }
  Tensor linear_hh(const Tensor& h) const {
    return linear(h, w_hh, b_hh, context);
  }
};

struct LSTMCell {
  explicit LSTMCell(CPUContext* context) : context_(context) {}
  t_tuple operator()(
      const Tensor& input,
      const t_tuple& hidden,
      const CellParams& params) const {
    const auto& hx = std::get<0>(hidden);
    const auto& cx = std::get<1>(hidden);
    auto linear_ih = params.linear_ih(input);
    auto linear_hh = params.linear_hh(hx);
    auto gates = add(linear_ih, linear_hh, context_);
    auto chunked_gates = chunk(gates, 4, 1, context_);
    auto ingate = sigmoid(chunked_gates[0]);
    auto forgetgate = sigmoid(chunked_gates[1]);
    auto cellgate = tanh(chunked_gates[2], context_);
    auto outgate = sigmoid(chunked_gates[3]);

auto cy =
        add(mul(forgetgate, cx, context_),
            mul(ingate, cellgate, context_),
            context_);
    auto hy = mul(outgate, tanh(cy, context_), context_);
    return std::make_tuple(std::move(hy), std::move(cy));
  }
  CPUContext* context_;
};

template <typename output_type, typename hidden_type>
struct LayerOutput {
  output_type outputs;
  hidden_type final_hidden;

LayerOutput(const output_type& _outputs, const hidden_type& _hidden) {
    outputs = copy_ctor(_outputs);
    final_hidden = copy_ctor(_hidden);
  }
};

template <typename hidden_type, typename param_type>
struct Layer {
  using output_type = LayerOutput<Tensor, hidden_type>;
  virtual ~Layer() {}
  virtual output_type operator()(
      const Tensor& input,
      const hidden_type& input_hidden,
      const param_type& params) const = 0;
};

struct FullLSTMLayer : Layer<t_tuple, CellParams> {
  FullLSTMLayer(LSTMCell& cell, CPUContext* context)
      : cell_(cell), context_(context) {}

LayerOutput<std::vector<Tensor>, t_tuple> operator()(
      const std::vector<Tensor>& step_inputs,
      const std::tuple<Tensor, Tensor>& input_hidden,
      const CellParams& params) const {
    std::vector<Tensor> step_outputs;
    auto hidden = copy_ctor(input_hidden);

for (size_t i = 0; i < step_inputs.size(); i++) {
      hidden = cell_(step_inputs[i], hidden, params);
      step_outputs.push_back(copy_ctor(std::get<0>(hidden)));
    }

return {step_outputs, hidden};
  }

LayerOutput<Tensor, t_tuple> operator()(
      const Tensor& inputs,
      const std::tuple<Tensor, Tensor>& input_hidden,
      const CellParams& params) const override {
    auto unstacked_output =
        (*this)(unbind(inputs, 0, context_), input_hidden, params);
    return {stack(unstacked_output.outputs, 0, context_),
            unstacked_output.final_hidden};
  }
  LSTMCell cell_;
  CPUContext* context_;
};

struct FullBidirectionalLSTMLayer
    : Layer<std::pair<t_tuple, t_tuple>, std::pair<CellParams, CellParams>> {
  using bidir_hidden_type = std::pair<t_tuple, t_tuple>;
  using param_type = std::pair<CellParams, CellParams>;
  using output_type = LayerOutput<Tensor, bidir_hidden_type>;

FullBidirectionalLSTMLayer(LSTMCell& cell, CPUContext* context)
      : layer_(cell, context), context_(context) {}

output_type operator()(
      const Tensor& input,
      const bidir_hidden_type& input_hidden,
      const param_type& params) const override {
    std::vector<Tensor> outputs;
    auto step_inputs = unbind(input, 0, context_);
    auto fw_result = layer_(step_inputs, input_hidden.first, params.first);
    auto fw_output = stack(fw_result.outputs, 0, context_);
    outputs.push_back(copy_ctor(fw_output));
    auto rev_step_inputs = reverse(std::move(step_inputs));
    auto rev_result =
        layer_(rev_step_inputs, input_hidden.second, params.second);
    std::reverse(rev_result.outputs.begin(), rev_result.outputs.end());
    auto rev_output = stack(rev_result.outputs, 0, context_);
    outputs.push_back(copy_ctor(rev_output));
    return {cat(outputs, fw_output.dim() - 1, context_),
            std::make_pair(
                std::move(fw_result.final_hidden),
                std::move(rev_result.final_hidden))};
  }

inline std::vector<Tensor> reverse(std::vector<Tensor>&& x) const {
    std::reverse(x.begin(), x.end());
    return std::move(x);
  }

private:
  FullLSTMLayer layer_;
  CPUContext* context_;
};

template <typename hidden_type, typename weight_type>
LayerOutput<Tensor, std::vector<hidden_type>> apply_layer_stack(
    const Layer<hidden_type, weight_type>& layer,
    const Tensor& input,
    const std::vector<hidden_type>& hiddens,
    const std::vector<weight_type>& weights,
    int64_t num_layers) {
  CAFFE_ENFORCE(
      num_layers == hiddens.size(),
      "Expected more hidden states in stacked_rnn");
  CAFFE_ENFORCE(
      num_layers == weights.size(), "Expected more weights in stacked_rnn");

auto layer_input = input.UnsafeSharedInstance();
  auto hidden_it = hiddens.begin();
  auto weight_it = weights.begin();
  std::vector<hidden_type> final_hiddens(num_layers);
  for (int64_t l = 0; l < num_layers; ++l) {
    auto layer_output = layer(layer_input, *(hidden_it++), *(weight_it++));
    final_hiddens.at(l) = std::move(layer_output.final_hidden);
    layer_input = std::move(layer_output.outputs);
  }
  return {layer_input, final_hiddens};
}

std::tuple<Tensor, Tensor, Tensor> _lstm_impl(
    const Tensor& input,
    const std::vector<CellParams>& params,
    const Tensor& hx,
    const Tensor& cx,
    int64_t num_layers,
    bool bidirectional,
    CPUContext* context) {
  using stack_output = LayerOutput<Tensor, std::vector<t_tuple>>;
  auto layer_hx = unbind(hx, 0, context);
  auto layer_cx = unbind(cx, 0, context);
  int64_t total_layers = layer_hx.size();
  std::vector<std::tuple<Tensor, Tensor>> hiddens;
  hiddens.reserve(total_layers);
  for (int64_t i = 0; i < total_layers; ++i) {
    hiddens.emplace_back(std::move(layer_hx[i]), std::move(layer_cx[i]));
  }
  LSTMCell cell(context);
  std::shared_ptr<stack_output> stack_output_ptr;
  if (bidirectional) {
    auto bidir_result = apply_layer_stack(
        FullBidirectionalLSTMLayer{cell, context},
        input,
        pair_vec(hiddens),
        pair_vec(params),
        num_layers);
    stack_output_ptr.reset(new stack_output(
        bidir_result.outputs,
        unpair_vec(std::move(bidir_result.final_hidden))));
  } else {
    auto result = apply_layer_stack(
        FullLSTMLayer{cell, context}, input, hiddens, params, num_layers);
    stack_output_ptr = std::make_shared<stack_output>(std::move(result));
  }

std::vector<Tensor> hy, cy;
  hy.reserve(total_layers);
  cy.reserve(total_layers);
  for (auto& hidden : stack_output_ptr->final_hidden) {
    hy.push_back(std::move(std::get<0>(hidden)));
    cy.push_back(std::move(std::get<1>(hidden)));
  }
  return std::make_tuple(
      std::move(stack_output_ptr->outputs),
      stack(hy, 0, context),
      stack(cy, 0, context));
}

// Parses a flat list of parameter tensors into a list of CellParams
std::vector<CellParams> gather_params(
    const std::vector<Tensor>& params,
    bool has_biases,
    CPUContext* context) {
  Tensor undefined;
  std::vector<CellParams> result;
  if (has_biases) {
    CAFFE_ENFORCE_EQ(
        params.size() % 4, 0, "got an incorrect number of LSTM parameters");
    for (size_t i = 0; i < params.size(); i += 4) {
      result.emplace_back(
          params[i], params[i + 1], params[i + 2], params[i + 3], context);
    }
  } else {
    CAFFE_ENFORCE_EQ(
        params.size() % 2, 0, "got an incorrect number of LSTM parameters");
    for (size_t i = 0; i < params.size(); i += 2) {
      result.emplace_back(
          params[i], params[i + 1], undefined, undefined, context);
    }
  }
  return result;
}

class InferenceLSTMOp : public Operator<CPUContext> {
 public:
  template <class... Args>
  explicit InferenceLSTMOp(Args&&... args)
      : Operator(std::forward<Args>(args)...),
        num_layers_(this->template GetSingleArgument<int64_t>("num_layers", 1)),
        bidirectional_(
            this->template GetSingleArgument<bool>("bidirectional", false)),
        has_biases_(this->template GetSingleArgument<bool>("has_biases", true)),
        batch_first_(
            this->template GetSingleArgument<bool>("batch_first", false)) {}

bool RunOnDevice() override;

protected:
  int64_t num_layers_;
  bool bidirectional_;
  bool has_biases_;
  bool batch_first_;
};

} // namespace
} // namespace caffe2
#endif // LSTM_OP_H_

cmd: