/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/
/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators
NameSizeModeActions
abs_op.h7050644editdlrm
accumulate_op.h10730644editdlrm
accuracy_op.h6520644editdlrm
acos_op.h7110644editdlrm
activation_ops_cudnn.h41220644editdlrm
affine_channel_op.h34500644editdlrm
alias_with_name.h12340644editdlrm
apmeter_op.h10270644editdlrm
arg_ops.h23190644editdlrm
asin_op.h7110644editdlrm
assert_op.h13350644editdlrm
async_net_barrier_op.h9040644editdlrm
atan_op.h7110644editdlrm
batch_box_cox_op.h22870644editdlrm
batch_bucketize_op.h7200644editdlrm
batch_gather_ops.h52640644editdlrm
batch_matmul_op.h96020644editdlrm
batch_moments_op.h33640644editdlrm
batch_permutation_op.h9540644editdlrm
batch_sparse_to_dense_op.h61470644editdlrm
bbox_transform_op.h26680644editdlrm
bisect_percentile_op.h49210644editdlrm
boolean_mask_ops.h26650644editdlrm
boolean_unmask_ops.h3780644editdlrm
box_with_nms_limit_op.h49600644editdlrm
bucketize_op.h13610644editdlrm
byte_weight_dequant_op.h17220644editdlrm
cast_op.h13930644editdlrm
cbrt_op.h7230644editdlrm
cc_bmm_bg_op.h38940644editdlrm
ceil_op.h7820644editdlrm
channel_backprop_stats_op.h7370644editdlrm
channel_shuffle_op.h19020644editdlrm
channel_stats_op.h18070644editdlrm
clip_op.h16390644editdlrm
collect_and_distribute_fpn_rpn_proposals_op.h68750644editdlrm
concat_split_op.h118500644editdlrm
conditional_op.h4870644editdlrm
conv_op.h31250644editdlrm
conv_op_cache_cudnn.h19350644editdlrm
conv_op_impl.h287290644editdlrm
conv_op_shared.h6720644editdlrm
conv_pool_op_base.h321090644editdlrm
conv_transpose_op.h17270644editdlrm
conv_transpose_op_impl.h182640644editdlrm
conv_transpose_op_mobile.h14700644editdlrm
conv_transpose_op_mobile_impl.h195870644editdlrm
conv_transpose_unpool_op_base.h103030644editdlrm
copy_op.h12960644editdlrm
copy_rows_to_tensor_op.h25990644editdlrm
cosh_op.h7110644editdlrm
cosine_embedding_criterion_op.h11270644editdlrm
cos_op.h7050644editdlrm
counter_ops.h45960644editdlrm
create_scope_op.h52320644editdlrm
cross_entropy_op.h44200644editdlrm
ctc_beam_search_decoder_op.h11020644editdlrm
ctc_greedy_decoder_op.h8170644editdlrm
cube_op.h7230644editdlrm
dataset_ops.h55010644editdlrm
data_couple.h4640644editdlrm
deform_conv_op.h35430644editdlrm
deform_conv_op_impl.h131710644editdlrm
dense_vector_to_id_list_op.h17970644editdlrm
distance_op.h84190644editdlrm
do_op.h69810644editdlrm
dropout_op.h15160644editdlrm
elementwise_add_op.h20240644editdlrm
elementwise_div_op.h12240644editdlrm
elementwise_linear_op.h11700644editdlrm
elementwise_logical_ops.h50830644editdlrm
elementwise_mul_op.h12240644editdlrm
elementwise_ops.h191150644editdlrm
elementwise_ops_utils.h10080644editdlrm
elementwise_op_test.h92370644editdlrm
elementwise_sub_op.h20250644editdlrm
elu_op.h8750644editdlrm
enforce_finite_op.h23030644editdlrm
ensure_clipped_op.h16080644editdlrm
ensure_cpu_output_op.h14650644editdlrm
erf_op.h7510644editdlrm
expand_op.h38770644editdlrm
expand_squeeze_dims_op.h34510644editdlrm
exp_op.h4250644editdlrm
fc_inference.h7750644editdlrm
feature_maps_ops.h324370644editdlrm
feed_blob_op.h8020644editdlrm
filler_op.h184310644editdlrm
find_duplicate_elements_op.h15630644editdlrm
find_op.h20550644editdlrm
flatten_op.h15250644editdlrm
flexible_top_k.h9360644editdlrm
floor_op.h7880644editdlrm
free_op.h7770644editdlrm
fully_connected_op.h93510644editdlrm
fused_rowwise_8bit_conversion_ops.h66010644editdlrm
fused_rowwise_nbitfake_conversion_ops.h43750644editdlrm
fused_rowwise_nbit_conversion_ops.h87230644editdlrm
fused_rowwise_random_quantization_ops.h26070644editdlrm
gather_fused_8bit_rowwise_op.h21790644editdlrm
gather_op.h75050644editdlrm
gather_ranges_to_dense_op.h81880644editdlrm
gelu_op.h14520644editdlrm
generate_proposals_op.h62560644editdlrm
generate_proposals_op_util_boxes.h143090644editdlrm
generate_proposals_op_util_nms.h262140644editdlrm
generate_proposals_op_util_nms_gpu.h21280644editdlrm
given_tensor_byte_string_to_uint8_fill_op.h21500644editdlrm
given_tensor_fill_op.h30020644editdlrm
glu_op.h14580644editdlrm
group_norm_op.h89670644editdlrm
gru_unit_op.h66260644editdlrm
half_float_ops.h27320644editdlrm
hard_sigmoid_op.h9940644editdlrm
heatmap_max_keypoint_op.h9390644editdlrm
histogram_op.h24210644editdlrm
h_softmax_op.h49540644editdlrm
if_op.h17640644editdlrm
im2col_op.h89430644editdlrm
index_hash_ops.h22320644editdlrm
index_ops.h31550644editdlrm
inference_lstm_op.h98810644editdlrm
instance_norm_op.h74410644editdlrm
integral_image_op.h9230644editdlrm
is_empty_op.h5580644editdlrm
jsd_op.h7210644editdlrm
key_split_ops.h14000644editdlrm
layer_norm_op.h80980644editdlrm
leaky_relu_op.h11110644editdlrm
lengths_pad_op.h25740644editdlrm
lengths_reducer_fused_8bit_rowwise_ops.h55320644editdlrm
lengths_reducer_fused_nbit_rowwise_ops.h234650644editdlrm
lengths_reducer_ops.h233150644editdlrm
lengths_reducer_rowwise_8bit_ops.h61800644editdlrm
lengths_tile_op.h5820644editdlrm
lengths_top_k_op.h13580644editdlrm
length_split_op.h22590644editdlrm
listwise_l2r_op.h16770644editdlrm
load_save_op.h140910644editdlrm
load_save_op_util.h16420644editdlrm
locally_connected_op.h38720644editdlrm
locally_connected_op_impl.h264950644editdlrm
locally_connected_op_util.h13320644editdlrm
local_response_normalization_op.h28040644editdlrm
log1p_op.h7170644editdlrm
logit_op.h11290644editdlrm
log_op.h4310644editdlrm
loss_op.h10580644editdlrm
lpnorm_op.h12790644editdlrm
lstm_unit_op.h67330644editdlrm
lstm_utils.h94240644editdlrm
map_ops.h80110644editdlrm
margin_ranking_criterion_op.h11130644editdlrm
matmul_op.h28430644editdlrm
max_pool_with_index_gpu.h11550644editdlrm
mean_op.h32520644editdlrm
merge_id_lists_op.h25700644editdlrm
minmax_ops.h38290644editdlrm
mish_op.h7940644editdlrm
mod_op.h9840644editdlrm
moments_op.h40510644editdlrm
multi_class_accuracy_op.h5390644editdlrm
negate_gradient_op.h5660644editdlrm
negative_op.h4510644editdlrm
ngram_ops.h26440644editdlrm
normalize_l1_op.h10750644editdlrm
normalize_op.h30130644editdlrm
no_default_engine_op.h10630644editdlrm
numpy_tile_op.h36430644editdlrm
one_hot_ops.h25620644editdlrm
onnx_while_op.h106550644editdlrm
operator_fallback_gpu.h41550644editdlrm
op_utils_cudnn.h21120644editdlrm
order_switch_ops.h21490644editdlrm
pack_rnn_sequence_op.h30740644editdlrm
pack_segments.h27290644editdlrm
pad_op.h29020644editdlrm
partition_ops.h99580644editdlrm
percentile_op.h10090644editdlrm
perplexity_op.h4470644editdlrm
piecewise_linear_transform_op.h82810644editdlrm
pool_op.h85250644editdlrm
pool_op_util.h11050644editdlrm
pow_op.h46770644editdlrm
prefetch_op.h46610644editdlrm
prelu_op.h10670644editdlrm
prepend_dim_op.h27600644editdlrm
quantile_op.h41200644editdlrm
quant_decode_op.h53700644editdlrm
rank_loss_op.h8200644editdlrm
reciprocal_op.h7210644editdlrm
reducer_functors.h245560644editdlrm
reduce_front_back_max_ops.h43990644editdlrm
reduce_front_back_sum_mean_ops.h53370644editdlrm
reduce_ops.h99620644editdlrm
reduction_ops.h59440644editdlrm
relu_n_op.h9900644editdlrm
relu_op.h6240644editdlrm
remove_data_blocks_op.h26510644editdlrm
replace_nan_op.h11700644editdlrm
reshape_op.h57230644editdlrm
resize_3d_op.h26770644editdlrm
resize_op.h23070644editdlrm
reverse_packed_segs_op.h27720644editdlrm
rmac_regions_op.h7080644editdlrm
rms_norm_op.h29680644editdlrm
roi_align_gradient_op.h14860644editdlrm
roi_align_op.h28570644editdlrm
roi_align_rotated_gradient_op.h13690644editdlrm
roi_align_rotated_op.h16360644editdlrm
roi_pool_op.h25030644editdlrm
rowmul_op.h19470644editdlrm
rsqrt_op.h7290644editdlrm
scale_blobs_op.h14580644editdlrm
scale_op.h10190644editdlrm
segment_reduction_op.h710220644editdlrm
self_binning_histogram_op.h62580644editdlrm
selu_op.h15450644editdlrm
sequence_ops.h82640644editdlrm
shape_op.h16380644editdlrm
sigmoid_op.h6390644editdlrm
sinh_op.h7110644editdlrm
sinusoid_position_encoding_op.h28340644editdlrm
sin_op.h7050644editdlrm
slice_op.h100710644editdlrm
softmax_op.h11740644editdlrm
softmax_utils.h4470644editdlrm
softmax_with_loss_op.h28830644editdlrm
softplus_op.h7810644editdlrm
softsign_op.h6750644editdlrm
space_batch_op.h68480644editdlrm
sparse_dropout_with_replacement_op.h11220644editdlrm
sparse_itemwise_dropout_with_replacement_op.h11630644editdlrm
sparse_lp_regularizer_op.h11300644editdlrm
sparse_normalize_op.h8340644editdlrm
sparse_to_dense_mask_op.h100510644editdlrm
sparse_to_dense_op.h39770644editdlrm
spatial_batch_norm_op.h151750644editdlrm
spatial_softmax_with_loss_op.h21820644editdlrm
sqrt_op.h4480644editdlrm
sqr_op.h4310644editdlrm
square_root_divide_op.h18570644editdlrm
stats_put_ops.h28130644editdlrm
stop_gradient.h5480644editdlrm
string_ops.h20670644editdlrm
stump_func_op.h21120644editdlrm
summarize_op.h18750644editdlrm
swish_op.h7720644editdlrm
tanh_op.h7230644editdlrm
tan_op.h7050644editdlrm
tensor_protos_db_input.h36330644editdlrm
text_file_reader_utils.h29000644editdlrm
thresholded_relu_op.h11370644editdlrm
tile_op.h87410644editdlrm
top_k.h10610644editdlrm
transpose_op.h20820644editdlrm
tt_linear_op.h65010644editdlrm
unique_ops.h16660644editdlrm
unsafe_coalesce.h24810644editdlrm
upsample_op.h22460644editdlrm
utility_ops.h499940644editdlrm
variable_length_sequence_padding.h13780644editdlrm
weighted_multi_sampling_op.h6020644editdlrm
weighted_sample_op.h7390644editdlrm
while_op.h19610644editdlrm
zero_gradient_op.h3470644editdlrm
Edit: /usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/fully_connected_op.h (9351B)
#ifndef CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_ #define CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_ #include #include "caffe2/core/context.h" #include "caffe2/core/operator.h" #include "caffe2/utils/conversions.h" #include "caffe2/utils/math.h" namespace caffe2 { // This is Caffe's InnerProductOp, with a name that fits its purpose better. template < class Context, class Engine = DefaultEngine, bool TransposeWeight = true> class FullyConnectedOp final : public Operator { public: USE_OPERATOR_CONTEXT_FUNCTIONS; template explicit FullyConnectedOp(Args&&... args) : Operator(std::forward(args)...), axis_(this->template GetSingleArgument("axis", 1)), axis_w_(this->template GetSingleArgument("axis_w", 1)), float16_compute_( this->template GetSingleArgument("float16_compute", false)) {} ~FullyConnectedOp() {} template < typename T_X, typename T_W, typename T_B, typename T_Y, typename MATH> bool DoRunWithType() { const auto& X = Input(0); const auto& W = Input(1); const auto& b = Input(2); CAFFE_ENFORCE(b.dim() == 1, b.dim()); // batch size const auto canonical_axis = X.canonical_axis_index(axis_); const auto M = X.size_to_dim(canonical_axis); const auto K = X.size_from_dim(canonical_axis); const auto canonical_axis_w = W.canonical_axis_index(axis_w_); const int N = TransposeWeight ? W.size_to_dim(canonical_axis_w) : W.size_from_dim(canonical_axis_w); auto dimErrorString = [&]() { return c10::str( "Dimension mismatch: ", "X: ", X.sizes(), ", W: ", W.sizes(), ", b: ", b.sizes(), ", axis: ", axis_, ", M: ", M, ", N: ", N, ", K: ", K); }; // Error checking CAFFE_ENFORCE(M == X.numel() / K, dimErrorString()); CAFFE_ENFORCE(K == W.numel() / N, dimErrorString()); CAFFE_ENFORCE(N == b.dim32(0), dimErrorString()); CAFFE_ENFORCE(N == b.numel(), dimErrorString()); Y_shape_cache_ = X.sizes().vec(); // This is an invariant of canonical_axis, so we can DCHECK. DCHECK_LE(canonical_axis + 1, Y_shape_cache_.size()); Y_shape_cache_.resize(canonical_axis + 1); Y_shape_cache_[canonical_axis] = N; auto* Y = Output(0, Y_shape_cache_, at::dtype()); CAFFE_ENFORCE(M * N == Y->numel(), dimErrorString()); if (X.numel() == 0) { // skip the rest of the computation if X is empty Y->template mutable_data(); return true; } // default to FLOAT as math.h does. TensorProto::DataType math_type = TensorProto_DataType_FLOAT; if (fp16_type()) { math_type = TensorProto_DataType_FLOAT16; } // W * x math::Gemm( CblasNoTrans, TransposeWeight ? CblasTrans : CblasNoTrans, M, N, K, 1, X.template data(), W.template data(), 0, Y->template mutable_data(), &context_, math_type); // Add bias term if (!bias_multiplier_.has_value()) { bias_multiplier_ = caffe2::empty({M}, at::dtype().device(Context::GetDeviceType())); math::Set( M, convert::To(1), bias_multiplier_->template mutable_data(), &context_); } else if (bias_multiplier_->numel() != M) { bias_multiplier_->Resize(M); math::Set( M, convert::To(1), bias_multiplier_->template mutable_data(), &context_); } math::Gemm( CblasNoTrans, CblasNoTrans, M, N, 1, 1, bias_multiplier_->template data(), b.template data(), 1, Y->template mutable_data(), &context_, math_type); return true; } bool RunOnDevice() override { return DoRunWithType< float, // X float, // W float, // B float, // Y float>(); // Math } protected: size_t axis_{1}; size_t axis_w_{1}; // A local vector to cache the output shape so we don't need to recreate // a vector object every time we run Run(). vector Y_shape_cache_; c10::optional bias_multiplier_; bool float16_compute_; }; template < class Context, class Engine = DefaultEngine, bool TransposeWeight = true> class FullyConnectedGradientOp : public Operator { public: USE_OPERATOR_CONTEXT_FUNCTIONS; template explicit FullyConnectedGradientOp(Args&&... args) : Operator(std::forward(args)...), axis_(this->template GetSingleArgument("axis", 1)), axis_w_(this->template GetSingleArgument("axis_w", 1)), float16_compute_( this->template GetSingleArgument("float16_compute", false)) {} ~FullyConnectedGradientOp() {} template < typename T_X, typename T_W, typename T_DY, typename T_B, typename T_DX, typename T_DW, typename T_DB, typename MATH> bool DoRunWithType() { const auto& X = Input(0); const auto& W = Input(1); const auto& dY = Input(2); // batch size const auto canonical_axis = X.canonical_axis_index(axis_); const int M = X.size_to_dim(canonical_axis); const int K = X.size_from_dim(canonical_axis); const auto canonical_axis_w = W.canonical_axis_index(axis_w_); const int N = TransposeWeight ? W.size_to_dim(canonical_axis_w) : W.size_from_dim(canonical_axis_w); auto dimErrorString = [&]() { return c10::str( "Dimension mismatch: ", "X: ", X.sizes(), ", W: ", W.sizes(), ", dY: ", dY.sizes(), ", axis: ", axis_, ", M: ", M, ", N: ", N, ", K: ", K); }; CAFFE_ENFORCE(M * K == X.numel(), dimErrorString()); CAFFE_ENFORCE(K * N == W.numel(), dimErrorString()); auto* dW = Output(0, W.sizes(), at::dtype()); auto* db = Output(1, {N}, at::dtype()); if (X.numel() == 0) { // generate a zero blob for db and dW when X is empty math::Set( db->numel(), convert::To(0), db->template mutable_data(), &context_); math::Set( dW->numel(), convert::To(0), dW->template mutable_data(), &context_); if (OutputSize() == 3) { Output(2, X.sizes(), at::dtype()); } return true; } // default to FLOAT as math.h does. TensorProto::DataType math_type = TensorProto_DataType_FLOAT; if (fp16_type()) { math_type = TensorProto_DataType_FLOAT16; } // Compute dW math::Gemm( CblasTrans, CblasNoTrans, TransposeWeight ? N : K, TransposeWeight ? K : N, M, 1, TransposeWeight ? dY.template data() : X.template data(), TransposeWeight ? X.template data() : dY.template data(), 0, dW->template mutable_data(), &context_, math_type); if (!bias_multiplier_.has_value()) { bias_multiplier_ = caffe2::empty({M}, at::dtype().device(Context::GetDeviceType())); math::Set( M, convert::To(1), bias_multiplier_->template mutable_data(), &context_); } else if (bias_multiplier_->numel() != M) { bias_multiplier_->Resize(M); math::Set( M, convert::To(1), bias_multiplier_->template mutable_data(), &context_); } // Compute dB math::Gemv( CblasTrans, M, N, 1, dY.template data(), bias_multiplier_->template data(), 0, db->template mutable_data(), &context_); // Compute dX if (OutputSize() == 3) { auto* dX = Output(2, X.sizes(), at::dtype()); math::Gemm( CblasNoTrans, TransposeWeight ? CblasNoTrans : CblasTrans, M, K, N, 1, dY.template data(), W.template data(), 0, dX->template mutable_data(), &context_, math_type); } return true; } bool RunOnDevice() override { return DoRunWithType< float, // X float, // W float, // dY float, // B float, // dX float, // dW float, // dB float>(); // Math } protected: size_t axis_{1}; size_t axis_w_{1}; c10::optional bias_multiplier_; bool float16_compute_; }; } // namespace caffe2 #endif // CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_