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/filler_op.h (18431B)

#ifndef CAFFE2_OPERATORS_FILLER_OP_H_
#define CAFFE2_OPERATORS_FILLER_OP_H_

#include "caffe2/core/context.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/math.h"

namespace caffe2 {

// FillerOp takes in either zero or one input.
//
// If the number of input is 1, the shape will be identical to that of the input
// at run time with optional additional dimensions appended at the end as
// specified by "extra_shape" argument. In that case the "shape" parameter
// should not be set.
//
// If the number of inputs is 0, the full shape must be provided via "shape"
// argument
template <class Context>
class FillerOp : public Operator<Context> {
 public:
  template <class... Args>
  explicit FillerOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        shape_(this->template GetRepeatedArgument<int64_t>("shape")),
        extra_shape_(ToVectorint64_t(
            this->template GetRepeatedArgument<int>("extra_shape"))),
        input_as_shape_(
            this->template GetSingleArgument<bool>("input_as_shape", false)) {
    if (InputSize()) {
      if (shape_.size() != 0) {
        CAFFE_THROW(
            "Cannot set the shape argument and pass in an input at "
            "the same time");
      }
    } else {
      if (!extra_shape_.empty()) {
        CAFFE_THROW("Cannot set extra_shape when there is no input");
      }
      if (input_as_shape_) {
        CAFFE_THROW("An input must be given if input_as_shape is true");
      }
      if (shape_.size() == 0 &&
          this->template HasSingleArgumentOfType<int>("shape")) {
        CAFFE_THROW("Fill 'shape' argument was a scalar, list expected");
      }
    }
  }

virtual ~FillerOp() {}
  USE_OPERATOR_CONTEXT_FUNCTIONS;

bool RunOnDevice() override {
    auto* output = Operator<Context>::Output(0);
    if (InputSize()) {
      auto shape = vector<int64_t>{};
      if (input_as_shape_) {
        if (this->InputIsTensorType(0, CPU)) {
          // originally, shape input must be in CPU context
          auto& input = this->template Input<Tensor>(0, CPU);
          CAFFE_ENFORCE_EQ(
              input.dim(),
              1,
              "When input_as_shape is true, the input must be a 1D tensor of "
              "data type int64_t");
          CAFFE_ENFORCE(input.numel() > 0);
          auto* shape_data = input.template data<int64_t>();
          shape.insert(shape.end(), shape_data, shape_data + input.dim32(0));
        } else {
          // in ONNX case, we allow shape to be in CUDA context
          auto& input = Input(0);
          CAFFE_ENFORCE_EQ(
              input.dim(),
              1,
              "When input_as_shape is true, the input must be a 1D tensor of "
              "data type int64_t");
          CAFFE_ENFORCE(input.numel() > 0);
          auto* shape_data = input.template data<int64_t>();
          std::unique_ptr<int64_t[]> shape_data_copy =
              std::make_unique<int64_t[]>(input.dim32(0));
          context_.template CopyToCPU<int64_t>(
              input.dim32(0), shape_data, shape_data_copy.get());
          shape.insert(
              shape.end(),
              shape_data_copy.get(),
              shape_data_copy.get() + input.dim32(0));
        }
      } else {
        auto& input = Input(0);
        shape.insert(shape.end(), input.sizes().begin(), input.sizes().end());
      }
      shape.insert(shape.end(), extra_shape_.begin(), extra_shape_.end());
      output->Resize(shape);
      shape_ = shape;
    } else {
      output->Resize(shape_);
    }
    return Fill(output);
  }

virtual bool Fill(Tensor* output) = 0;

protected:
  vector<int64_t> shape_;
  vector<int64_t> extra_shape_;
  bool input_as_shape_;
};

template <typename T, class Context>
class UniformFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit UniformFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...),
        min_(this->template GetSingleArgument<T>("min", 0)),
        max_(this->template GetSingleArgument<T>("max", 1)) {
    if (InputSize() == 3) {
      CAFFE_ENFORCE(
          !this->template HasSingleArgumentOfType<T>("min"),
          "Cannot set both min arg and min input blob");
      CAFFE_ENFORCE(
          !this->template HasSingleArgumentOfType<T>("max"),
          "Cannot set both max arg and max input blob");
    } else {
      CAFFE_ENFORCE_LT(
          min_, max_, "Max value should be bigger than min value.");
    }
  }

bool Fill(Tensor* output) override {
    T min = min_;
    T max = max_;
    if (InputSize() == 3) {
      CAFFE_ENFORCE_EQ(1, Input(1).numel(), "min blob must be scalar");
      CAFFE_ENFORCE_EQ(1, Input(2).numel(), "max blob must be scalar");
      min = *Input(1).template data<T>();
      max = *Input(2).template data<T>();
      if (min > max) {
        auto shape = output->sizes().vec();
        shape[0] = 0;
        output->Resize(shape);
        output->template mutable_data<T>();
        return true;
      }
    }
    math::RandUniform<T, Context>(
        output->numel(),
        min,
        max,
        output->template mutable_data<T>(),
        &context_);
    return true;
  }

private:
  T min_;
  T max_;
};

template <class Context>
class UniqueUniformFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit UniqueUniformFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...) {
    TensorProto_DataType dtype =
        static_cast<TensorProto_DataType>(this->template GetSingleArgument<int>(
            "dtype", TensorProto_DataType_INT32));

switch (dtype) {
      case TensorProto_DataType_INT32:
        CheckRange<int>();
        body_ = &UniqueUniformFillOp::FillWithType<int>;
        break;
      case TensorProto_DataType_INT64:
        CheckRange<int64_t>();
        body_ = &UniqueUniformFillOp::FillWithType<int64_t>;
        break;
      case TensorProto_DataType_UNDEFINED:
        CAFFE_THROW(
            "UniqueUniformFill op cannot have undefined 'dtype' argument");
      // break;
      default:
        CAFFE_THROW("Unexpected 'dtype' argument value: ", dtype);
    }
  }

bool Fill(Tensor* output) override {
    return (this->*body_)(output);
  }

private:
  template <typename T>
  void CheckRange() {
    CAFFE_ENFORCE(this->template HasSingleArgumentOfType<T>("min"));
    CAFFE_ENFORCE(this->template HasSingleArgumentOfType<T>("max"));
    CAFFE_ENFORCE_LT(
        this->template GetSingleArgument<T>("min", 0),
        this->template GetSingleArgument<T>("max", 0),
        "Max value should be bigger than min value.");
  }

template <typename T>
  bool FillWithType(Tensor* output) {
    T min = this->template GetSingleArgument<T>("min", 0);
    T max = this->template GetSingleArgument<T>("max", 0);

const T* avoid_data = nullptr;
    size_t avoid_size = 0;
    if (InputSize() >= 2) {
      auto& avoid = Input(1);
      avoid_data = avoid.template data<T>();
      avoid_size = avoid.numel();
    }
    math::RandUniformUnique<T, Context>(
        output->numel(),
        min,
        max,
        output->template mutable_data<T>(),
        avoid_size,
        avoid_data,
        &context_);
    return true;
  }

bool (UniqueUniformFillOp::*body_)(Tensor* output);
};

template <class Context>
class ConstantFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit ConstantFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...) {
    TensorProto_DataType dtype =
        static_cast<TensorProto_DataType>(this->template GetSingleArgument<int>(
            "dtype", TensorProto_DataType_FLOAT));

if (!OperatorBase::HasArgument("dtype") &&
        OperatorBase::HasArgument("value")) {
      // If 'dtype' is not provided, infer type based on the type of 'value'
      // Currently, single argument contains either float, int64 or bytes
      if (this->template HasSingleArgumentOfType<float>("value")) {
        dtype = TensorProto_DataType_FLOAT;
      } else if (this->template HasSingleArgumentOfType<int64_t>("value")) {
        dtype = TensorProto_DataType_INT64;
      } else {
        CAFFE_THROW("Argument 'value' is of unexpected type");
      }
      VLOG(1) << "Argument 'dtype' is not provided. Assume the data type is "
              << "the same as that of argument 'value': " << dtype;
    }

switch (dtype) {
      case TensorProto_DataType_FLOAT:
        body_ = &ConstantFillOp::FillWithType<float>;
        break;
      case TensorProto_DataType_DOUBLE:
        body_ = &ConstantFillOp::FillWithType<double>;
        break;
      case TensorProto_DataType_BOOL:
        body_ = &ConstantFillOp::FillWithType<bool>;
        break;
      case TensorProto_DataType_INT8:
        body_ = &ConstantFillOp::FillWithType<int8_t>;
        break;
      case TensorProto_DataType_INT16:
        body_ = &ConstantFillOp::FillWithType<int16_t>;
        break;
      case TensorProto_DataType_INT32:
        body_ = &ConstantFillOp::FillWithType<int>;
        break;
      case TensorProto_DataType_INT64:
        body_ = &ConstantFillOp::FillWithType<int64_t>;
        break;
      case TensorProto_DataType_UINT8:
        body_ = &ConstantFillOp::FillWithType<uint8_t>;
        break;
      case TensorProto_DataType_UINT16:
        body_ = &ConstantFillOp::FillWithType<uint16_t>;
        break;
      case TensorProto_DataType_STRING:
        body_ = &ConstantFillOp::FillWithString;
        break;
      case TensorProto_DataType_UNDEFINED:
        CAFFE_THROW("ConstantFill op cannot have undefined 'dtype' argument");
      // break;
      default:
        CAFFE_THROW("Unexpected 'dtype' argument value: ", dtype);
    }
  }

bool Fill(Tensor* output) override {
    return (this->*body_)(output);
  }

template <typename T>
  bool FillWithType(Tensor* output) {
    T value = this->template GetSingleArgument<T>("value", 0);
    if (InputSize() == 2) {
      auto& value_vec = Input(1);
      if (value_vec) {
        CAFFE_ENFORCE_EQ(
            value_vec.size(), 1, "value vector must have 1 element");
        value = value_vec.template data<T>()[0];
      }
    }

auto* data = output->template mutable_data<T>();
    if (output->numel()) {
      math::Set<T, Context>(output->numel(), value, data, &context_);
    }
    return true;
  }

bool FillWithString(Tensor* output) {
    CAFFE_ENFORCE_LT(
        InputSize(), 2, "constant fill string from tensor is not supported");
    auto value = this->template GetSingleArgument<std::string>("value", "");
    auto* data = output->template mutable_data<std::string>();
    for (int i = 0; i < output->numel(); ++i) {
      data[i] = value;
    }
    return true;
  }

private:
  bool (ConstantFillOp::*body_)(Tensor* output);
};

template <class Context>
class DiagonalFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit DiagonalFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...) {
    TensorProto_DataType dtype =
        static_cast<TensorProto_DataType>(this->template GetSingleArgument<int>(
            "dtype", TensorProto_DataType_FLOAT));

switch (dtype) {
      case TensorProto_DataType_FLOAT:
        body_ = &DiagonalFillOp::FillWithType<float>;
        break;
      case TensorProto_DataType_DOUBLE:
        body_ = &DiagonalFillOp::FillWithType<double>;
        break;
      case TensorProto_DataType_BOOL:
        body_ = &DiagonalFillOp::FillWithType<bool>;
        break;
      case TensorProto_DataType_INT8:
        body_ = &DiagonalFillOp::FillWithType<int8_t>;
        break;
      case TensorProto_DataType_INT16:
        body_ = &DiagonalFillOp::FillWithType<int16_t>;
        break;
      case TensorProto_DataType_INT32:
        body_ = &DiagonalFillOp::FillWithType<int>;
        break;
      case TensorProto_DataType_INT64:
        body_ = &DiagonalFillOp::FillWithType<int64_t>;
        break;
      case TensorProto_DataType_UINT8:
        body_ = &DiagonalFillOp::FillWithType<uint8_t>;
        break;
      case TensorProto_DataType_UINT16:
        body_ = &DiagonalFillOp::FillWithType<uint16_t>;
        break;
      case TensorProto_DataType_UNDEFINED:
        CAFFE_THROW("Cannot have undefined 'dtype' argument");
      default:
        CAFFE_THROW("Unexpected 'dtype' argument value: ", dtype);
    }
  }

bool Fill(Tensor* output) override {
    return (this->*body_)(output);
  }

template <typename T>
  bool FillWithType(Tensor* output);

private:
  void VerifyOutputShape(Tensor* output) {
    CAFFE_ENFORCE(output->dim() >= 2, "Input shape must be >= 2D");
  }

int64_t GetStepSize(Tensor* output) {
    int64_t step;
    if (output->dim() == 2) {
      step = output->size(1) + 1;
    } else {
      int64_t prev_i = output->size(0);
      for (auto i : output->sizes()) {
        if (i != prev_i) {
          CAFFE_THROW("All dimensions of input must be of equal length");
        }
      }
      vector<int64_t> cumprod(output->dim());
      auto dims = output->sizes();
      std::partial_sum(
          dims.begin(),
          dims.end() - 1,
          cumprod.begin(),
          std::multiplies<int64_t>());
      step = 1 +
          std::accumulate(
                 cumprod.begin(), cumprod.end(), static_cast<int64_t>(0));
      VLOG(0) << step;
    }
    return step;
  }

bool (DiagonalFillOp::*body_)(Tensor* output);
};

template <typename T, class Context>
class GaussianFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit GaussianFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...),
        mean_(this->template GetSingleArgument<float>("mean", 0)),
        std_(this->template GetSingleArgument<float>("std", 1)) {
    DCHECK_GT(std_, 0) << "Standard deviation should be nonnegative.";
  }

bool Fill(Tensor* output) override {
    math::RandGaussian<T, Context>(
        output->numel(),
        mean_,
        std_,
        output->template mutable_data<T>(),
        &context_);
    return true;
  }

private:
  T mean_;
  T std_;
};

template <typename T, class Context>
class XavierFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit XavierFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...) {}

bool Fill(Tensor* output) override {
    const int fan_in = output->numel() / output->dim32(0);
    T scale = std::sqrt(T(3) / fan_in);
    math::RandUniform<T, Context>(
        output->numel(),
        -scale,
        scale,
        output->template mutable_data<T>(),
        &context_);
    return true;
  }
};

template <typename T, class Context>
class MSRAFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit MSRAFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...) {}

bool Fill(Tensor* output) override {
    const int fan_out = output->numel() / output->dim32(1);
    T scale = std::sqrt(T(2) / fan_out);
    math::RandGaussian<T, Context>(
        output->numel(),
        0.0,
        scale,
        output->template mutable_data<T>(),
        &context_);
    return true;
  }
};

// This is mostly used just as a debugging purpose stuff: it fills a tensor
// sequentially with values 0, 1, 2..., which can then be used to check e.g.
// reshape operations by allowing one to read the indices more easily.
template <typename T, class Context>
class RangeFillOp final : public FillerOp<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit RangeFillOp(Args&&... args)
      : FillerOp<Context>(std::forward<Args>(args)...) {}

bool Fill(Tensor* output) override;
};

template <class Context>
class LengthsRangeFillOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  USE_SIMPLE_CTOR_DTOR(LengthsRangeFillOp);

bool RunOnDevice() override {
    auto& input = Input(0);

auto* input_data = input.template data<int32_t>();

CAFFE_ENFORCE_EQ(input.dim(), 1, "Input must be a vector.");

auto len_sum = std::accumulate(input_data, input_data + input.numel(), 0);

auto* output = Output(0, {len_sum}, at::dtype<int32_t>());
    auto* output_data = output->template mutable_data<int32_t>();

int32_t offset = 0;
    for (int i = 0; i < input.numel(); ++i) {
      auto len = input_data[i];
      auto start = output_data + offset;
      std::iota(
          start,
          start + len,
          0); // make the third argument the arg of this operator
      offset += len;
    }
    return true;
  }
};

template <int VALUE_TYPE = TensorProto_DataType_FLOAT>
inline std::vector<TensorShape> FillerTensorInference(
    const OperatorDef& def,
    const vector<TensorShape>& in) {
  vector<TensorShape> out(1);
  ArgumentHelper helper(def);
  out[0].set_data_type(static_cast<TensorProto_DataType>(
      helper.GetSingleArgument<int>("dtype", VALUE_TYPE)));

if (in.size()) {
    // TODO
    bool input_as_shape =
        helper.GetSingleArgument<bool>("input_as_shape", false);
    if (input_as_shape) {
      out[0].set_unknown_shape(true);
      return out;
    }
    for (auto d : in[0].dims()) {
      out[0].add_dims(d);
    }
  } else {
    auto shape = helper.GetRepeatedArgument<int64_t>("shape");
    for (auto d : shape) {
      out[0].add_dims(d);
    }
  }
  return out;
}

} // namespace caffe2

#endif // CAFFE2_OPERATORS_FILLER_OP_H_

cmd: