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/spatial_batch_norm_op.h (15175B)

#ifndef CAFFE2_OPERATORS_SPATIAL_BATCH_NORM_OP_H_
#define CAFFE2_OPERATORS_SPATIAL_BATCH_NORM_OP_H_

#include <algorithm>
#include <array>
#include <functional>
#include <limits>
#include <string>
#include <vector>

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

namespace caffe2 {

template <class Context>
class SpatialBNOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;

template <class... Args>
  explicit SpatialBNOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        OP_SINGLE_ARG(bool, OpSchema::Arg_IsTest, is_test_, false),
        OP_SINGLE_ARG(double, "epsilon", epsilon_, 1e-5),
        OP_SINGLE_ARG(float, "momentum", momentum_, 0.9f),
        order_(StringToStorageOrder(
            this->template GetSingleArgument<std::string>("order", "NCHW"))),
        OP_SINGLE_ARG(int, "num_batches", num_batches_, 1) {
    CAFFE_ENFORCE_NE(
        order_,
        StorageOrder::UNKNOWN,
        "order should be either \"NCHW\" or \"NHWC\".");
    CAFFE_ENFORCE(
        (is_test_ && OutputSize() == 1) || (!is_test_ && OutputSize() == 5));
    CAFFE_ENFORCE_GT(epsilon_, 0);
    CAFFE_ENFORCE_GE(momentum_, 0);
    CAFFE_ENFORCE_LE(momentum_, 1);
  }

virtual ~SpatialBNOp() = default;

bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<float>>::call(this, Input(0));
  }

template <typename T>
  bool DoRunWithType() {
    const auto& X = Input(INPUT);
    const auto& scale = Input(SCALE);
    const auto& bias = Input(BIAS);

const int ndim = X.dim();
    CAFFE_ENFORCE_GE(ndim, 2);
    const int N = X.dim32(0);
    const int C =
        (order_ == StorageOrder::NCHW ? X.dim32(1) : X.dim32(ndim - 1));
    const std::vector<int> X_dims(X.sizes().cbegin(), X.sizes().cend());
    CAFFE_ENFORCE_NE(C, 0);
    const int HxW =
        std::accumulate(
            X_dims.cbegin() + 1, X_dims.cend(), 1, std::multiplies<int>()) /
        C;
    CAFFE_ENFORCE_EQ(scale.numel(), C);
    CAFFE_ENFORCE_EQ(bias.numel(), C);

auto* Y = Output(OUTPUT, X.sizes(), at::dtype<T>());
    const T* X_data = X.template data<T>();
    const T* scale_data = scale.template data<T>();
    const T* bias_data = bias.template data<T>();
    T* Y_data = Y->template mutable_data<T>();
    ReinitializeTensor(
        &alpha_, {C}, at::dtype<T>().device(Context::GetDeviceType()));
    ReinitializeTensor(
        &beta_, {C}, at::dtype<T>().device(Context::GetDeviceType()));
    T* alpha_data = alpha_.template mutable_data<T>();
    T* beta_data = beta_.template mutable_data<T>();
    if (is_test_) {
      if (N == 0) {
        return true;
      }
      const auto& mean = Input(EST_MEAN);
      const auto& var = Input(EST_VAR);
      CAFFE_ENFORCE_EQ(mean.numel(), C);
      CAFFE_ENFORCE_EQ(var.numel(), C);
      ComputeFusedParam<T>(
          C,
          scale_data,
          bias_data,
          mean.template data<T>(),
          var.template data<T>(),
          alpha_data,
          beta_data);
    } else {
      auto* saved_mean = Output(SAVED_MEAN, {C}, at::dtype<T>());
      auto* saved_rstd = Output(SAVED_INV_STD, {C}, at::dtype<T>());
      T* saved_mean_data = saved_mean->template mutable_data<T>();
      T* saved_rstd_data = saved_rstd->template mutable_data<T>();

// Enforce Alias
      CAFFE_ENFORCE(
          IsInputOutputAlias(3, 1), "Input 3 and Output 1 should be alias.");
      CAFFE_ENFORCE(
          IsInputOutputAlias(4, 2), "Input 4 and Output 2 should be alias.");

Tensor* running_mean = nullptr;
      Tensor* running_var = nullptr;
      const auto& mean = Input(EST_MEAN);
      const auto& var = Input(EST_VAR);
      if (mean.numel() != C) {
        running_mean = Output(RUNNING_MEAN, {C}, at::dtype<T>());
        C10_LOG_EVERY_MS(WARNING, 1000)
            << "[Depreacated] Running mean is not initialized in "
               "SpatialBatchNorm Op";
        math::Set<T, Context>(
            C, T(0), running_mean->template mutable_data<T>(), &context_);
      } else {
        running_mean = Output(RUNNING_MEAN, {C}, at::dtype<T>());
      }
      if (var.numel() != C) {
        running_var = Output(RUNNING_VAR, {C}, at::dtype<T>());
        math::Set<T, Context>(
            C, T(0), running_var->template mutable_data<T>(), &context_);
        C10_LOG_EVERY_MS(WARNING, 1000)
            << "[Deprecated] Running variance is not initialized in "
               "SpatialBatchNorm Op";
      } else {
        running_var = Output(RUNNING_VAR, {C}, at::dtype<T>());
      }

T* running_mean_data = running_mean->template mutable_data<T>();
      T* running_var_data = running_var->template mutable_data<T>();
      if (N == 0) {
        math::Set<T, Context>(C, T(0), saved_mean_data, &context_);
        math::Set<T, Context>(C, T(0), saved_rstd_data, &context_);
        return true;
      }
      if (num_batches_ > 1) {
        const auto& batch_mean_sum = Input(BATCH_MEAN_SUM);
        const auto& batch_var_sum = Input(BATCH_VAR_SUM);
        CAFFE_ENFORCE_EQ(batch_mean_sum.numel(), C);
        CAFFE_ENFORCE_EQ(batch_var_sum.numel(), C);
        ComputeBatchMoments<T>(
            N,
            C,
            HxW,
            batch_mean_sum.template data<T>(),
            batch_var_sum.template data<T>(),
            saved_mean_data,
            saved_rstd_data);
      } else {
        if (order_ == StorageOrder::NCHW) {
          const std::array<int, 3> X_dims_arr = {N, C, HxW};
          const std::array<int, 3> Y_dims_arr = {1, C, 1};
          math::Moments<T, Context>(
              3,
              X_dims_arr.data(),
              Y_dims_arr.data(),
              X_data,
              saved_mean_data,
              saved_rstd_data,
              &context_);
        } else {
          const std::array<int, 2> X_dims_arr = {N * HxW, C};
          const std::array<int, 2> Y_dims_arr = {1, C};
          math::Moments<T, Context>(
              2,
              X_dims_arr.data(),
              Y_dims_arr.data(),
              X_data,
              saved_mean_data,
              saved_rstd_data,
              &context_);
        }
      }
      ComputeRunningMomentsAndFusedParam<T>(
          C,
          num_batches_ * N * HxW,
          scale_data,
          bias_data,
          saved_mean_data,
          saved_rstd_data,
          running_mean_data,
          running_var_data,
          saved_rstd_data,
          alpha_data,
          beta_data);
    }
    if (order_ == StorageOrder::NCHW) {
      math::AffineChannel<T, Context, StorageOrder::NCHW>(
          N, C, HxW, X_data, alpha_data, beta_data, Y_data, &context_);
    } else {
      math::AffineChannel<T, Context, StorageOrder::NHWC>(
          N, C, HxW, X_data, alpha_data, beta_data, Y_data, &context_);
    }

return true;
  }

protected:
  template <typename T>
  void ComputeFusedParam(
      const int C,
      const T* scale,
      const T* bias,
      const T* mean,
      const T* var,
      T* alpha,
      T* beta) {
    EigenVectorArrayMap<T> alpha_arr(alpha, C);
    EigenVectorArrayMap<T> beta_arr(beta, C);
    alpha_arr = ConstEigenVectorArrayMap<T>(scale, C) *
        (ConstEigenVectorArrayMap<T>(var, C) + static_cast<T>(epsilon_))
            .rsqrt();
    beta_arr = ConstEigenVectorArrayMap<T>(bias, C) -
        alpha_arr * ConstEigenVectorArrayMap<T>(mean, C);
  }

template <typename T>
  void ComputeBatchMoments(
      const int N,
      const int C,
      const int HxW,
      const T* batch_mean_sum,
      const T* batch_var_sum,
      T* mean,
      T* var) {
    const T scale = T(1) / static_cast<T>(num_batches_ * N * HxW);
    EigenVectorArrayMap<T> mean_arr(mean, C);
    EigenVectorArrayMap<T> var_arr(var, C);
    mean_arr = ConstEigenVectorArrayMap<T>(batch_mean_sum, C) * scale;
    var_arr = ConstEigenVectorArrayMap<T>(batch_var_sum, C) * scale -
        mean_arr.square();
  }

template <typename T>
  void ComputeRunningMomentsAndFusedParam(
      const int C,
      const int reduce_size,
      const T* scale,
      const T* bias,
      const T* mean,
      const T* var,
      T* running_mean,
      T* running_var,
      T* rstd,
      T* alpha,
      T* beta) {
    const T a = T(1) - static_cast<T>(momentum_);
    const T b = static_cast<T>(momentum_);
    const T unbias_scale = reduce_size == 1
        ? std::numeric_limits<T>::infinity()
        : static_cast<T>(reduce_size) / static_cast<T>(reduce_size - 1);
    math::Axpby<T, T, Context>(C, a, mean, b, running_mean, &context_);
    math::Axpby<T, T, Context>(
        C, a * unbias_scale, var, b, running_var, &context_);
    math::InvStd<T, Context>(C, static_cast<T>(epsilon_), var, rstd, &context_);
    EigenVectorArrayMap<T> alpha_arr(alpha, C);
    EigenVectorArrayMap<T> beta_arr(beta, C);
    alpha_arr = ConstEigenVectorArrayMap<T>(scale, C) *
        ConstEigenVectorArrayMap<T>(rstd, C);
    beta_arr = ConstEigenVectorArrayMap<T>(bias, C) -
        alpha_arr * ConstEigenVectorArrayMap<T>(mean, C);
  }

const bool is_test_;
  double epsilon_;
  const float momentum_;
  const StorageOrder order_;
  const int num_batches_;

Tensor alpha_;
  Tensor beta_;

INPUT_TAGS(
      INPUT,
      SCALE,
      BIAS,
      EST_MEAN,
      EST_VAR,
      BATCH_MEAN_SUM,
      BATCH_VAR_SUM);
  OUTPUT_TAGS(OUTPUT, RUNNING_MEAN, RUNNING_VAR, SAVED_MEAN, SAVED_INV_STD);
};

template <class Context>
class SpatialBNGradientOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;

template <class... Args>
  explicit SpatialBNGradientOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        OP_SINGLE_ARG(double, "epsilon", epsilon_, 1e-5),
        order_(StringToStorageOrder(
            this->template GetSingleArgument<string>("order", "NCHW"))),
        OP_SINGLE_ARG(int, "num_batches", num_batches_, 1) {
    CAFFE_ENFORCE_NE(
        order_,
        StorageOrder::UNKNOWN,
        "order should be either \"NCHW\" or \"NHWC\".");
    CAFFE_ENFORCE(InputSize() == 5 || InputSize() == 7);
    CAFFE_ENFORCE_EQ(OutputSize(), 3);
  }

virtual ~SpatialBNGradientOp() = default;

bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<float>>::call(this, Input(0));
  }

template <typename T>
  bool DoRunWithType() {
    const auto& X = Input(INPUT);
    const auto& dY = Input(OUTPUT_GRAD);
    const auto& scale = Input(SCALE);
    const auto& mean = Input(SAVED_MEAN);
    const auto& rstd = Input(SAVED_INV_STD);
    const int ndim = X.dim();
    CAFFE_ENFORCE_GE(ndim, 3);
    const int N = X.dim32(0);
    const int C =
        (order_ == StorageOrder::NCHW ? X.dim32(1) : X.dim32(ndim - 1));
    const std::vector<int> X_dims(X.sizes().cbegin(), X.sizes().cend());
    const int HxW =
        std::accumulate(
            X_dims.cbegin() + 1, X_dims.cend(), 1, std::multiplies<int>()) /
        C;
    CAFFE_ENFORCE_EQ(scale.numel(), C);
    CAFFE_ENFORCE_EQ(mean.numel(), C);
    CAFFE_ENFORCE_EQ(rstd.numel(), C);

auto* dX = Output(INPUT_GRAD, X.sizes(), at::dtype<T>());
    at::IntArrayRef dscale_sizes, dbias_sizes;
    if (num_batches_ == 1) {
      dscale_sizes = scale.sizes();
      dbias_sizes = scale.sizes();
    } else {
      const auto& dscale_sum = Input(AGGREGATE_SCALE_GRAD);
      const auto& dbias_sum = Input(AGGREGATE_BIAS_GRAD);
      // Note: previously there was alias check to decide whether to call
      // ResizeLike or not, since we only call Resize when the size does not
      // match the size of cached Tensor, this check is not necessary
      dscale_sizes = dscale_sum.sizes();
      dbias_sizes = dbias_sum.sizes();
    }
    auto* dscale = Output(SCALE_GRAD, dscale_sizes, at::dtype<T>());
    auto* dbias = Output(BIAS_GRAD, dbias_sizes, at::dtype<T>());
    const T* X_data = X.template data<T>();
    const T* dY_data = dY.template data<T>();
    const T* scale_data = scale.template data<T>();
    const T* mean_data = mean.template data<T>();
    const T* rstd_data = rstd.template data<T>();
    T* dX_data = dX->template mutable_data<T>();
    T* dscale_data = dscale->template mutable_data<T>();
    T* dbias_data = dbias->template mutable_data<T>();

if (N == 0) {
      math::Set<T, Context>(C, T(0), dscale_data, &context_);
      math::Set<T, Context>(C, T(0), dbias_data, &context_);
      return true;
    }
    ReinitializeTensor(
        &alpha_, {C}, at::dtype<T>().device(Context::GetDeviceType()));
    ReinitializeTensor(
        &beta_, {C}, at::dtype<T>().device(Context::GetDeviceType()));
    ReinitializeTensor(
        &gamma_, {C}, at::dtype<T>().device(Context::GetDeviceType()));
    T* alpha_data = alpha_.template mutable_data<T>();
    T* beta_data = beta_.template mutable_data<T>();
    T* gamma_data = gamma_.template mutable_data<T>();
    if (num_batches_ > 1) {
      const auto& dscale_sum = Input(AGGREGATE_SCALE_GRAD);
      const auto& dbias_sum = Input(AGGREGATE_BIAS_GRAD);
      ComputeMultiBatchScaleBiasGradientsAndFusedParams<T>(
          N,
          C,
          HxW,
          scale_data,
          mean_data,
          rstd_data,
          dscale_sum.template data<T>(),
          dbias_sum.template data<T>(),
          dscale_data,
          dbias_data,
          alpha_data,
          beta_data,
          gamma_data);
    } else {
      ComputeScaleBiasGradientsAndFusedParams<T>(
          N,
          C,
          HxW,
          dY_data,
          X_data,
          scale_data,
          mean_data,
          rstd_data,
          dscale_data,
          dbias_data,
          alpha_data,
          beta_data,
          gamma_data,
          dX_data);
    }
    ComputeXGradient<T>(
        N, C, HxW, dY_data, X_data, alpha_data, beta_data, gamma_data, dX_data);

return true;
  }

protected:
  template <typename T>
  void ComputeMultiBatchScaleBiasGradientsAndFusedParams(
      const int N,
      const int C,
      const int HxW,
      const T* scale,
      const T* mean,
      const T* rstd,
      const T* dscale_sum,
      const T* dbias_sum,
      T* dscale,
      T* dbias,
      T* alpha,
      T* beta,
      T* gamma);

template <typename T>
  void ComputeScaleBiasGradientsAndFusedParams(
      const int N,
      const int C,
      const int HxW,
      const T* dY,
      const T* X,
      const T* scale,
      const T* mean,
      const T* rstd,
      T* dscale,
      T* dbias,
      T* alpha,
      T* beta,
      T* gamma,
      T* scratch);

template <typename T>
  void ComputeXGradient(
      const int N,
      const int C,
      const int HxW,
      const T* dY,
      const T* X,
      const T* alpha,
      const T* beta,
      const T* gamma,
      T* dX);

double epsilon_;
  const StorageOrder order_;
  const int num_batches_;

Tensor alpha_;
  Tensor beta_;
  Tensor gamma_;
  Tensor ones_;

INPUT_TAGS(
      INPUT,
      SCALE,
      OUTPUT_GRAD,
      SAVED_MEAN,
      SAVED_INV_STD,
      AGGREGATE_SCALE_GRAD,
      AGGREGATE_BIAS_GRAD);
  OUTPUT_TAGS(INPUT_GRAD, SCALE_GRAD, BIAS_GRAD);
};

} // namespace caffe2

#endif // CAFFE2_OPERATORS_SPATIAL_BATCH_NORM_OP_H_

cmd: