diff --git a/doc/fluid/api/detection.rst b/doc/fluid/api/detection.rst new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/doc/fluid/api/gen_doc.sh b/doc/fluid/api/gen_doc.sh index 0f0539355559446fd91f659d61b636db214b5a40..3ee299f5aceb32465574277c25d962360183f1cc 100755 --- a/doc/fluid/api/gen_doc.sh +++ b/doc/fluid/api/gen_doc.sh @@ -1,5 +1,5 @@ #!/bin/bash -python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst +python gen_doc.py layers --submodules control_flow device io nn ops tensor detection > layers.rst for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer do diff --git a/doc/fluid/api/io.rst b/doc/fluid/api/io.rst index 3e956f8302d261b52f9f76ff8eb4a01f9c6381f8..dd9d88b669957c22cd0a07fa4b7e219e2d6e5d61 100644 --- a/doc/fluid/api/io.rst +++ b/doc/fluid/api/io.rst @@ -59,21 +59,3 @@ get_inference_program .. autofunction:: paddle.fluid.io.get_inference_program :noindex: -save_checkpoint ---------------- - -.. autofunction:: paddle.fluid.io.save_checkpoint - :noindex: - -load_checkpoint ---------------- - -.. autofunction:: paddle.fluid.io.load_checkpoint - :noindex: - -clean_checkpoint ----------------- - -.. autofunction:: paddle.fluid.io.clean_checkpoint - :noindex: - diff --git a/doc/fluid/api/layers.rst b/doc/fluid/api/layers.rst index f78e6db3268e44d5f30d83508f07c4ed68106e48..ba33c0d7d650e76711040c40ab9e5fdcf11c3a6c 100644 --- a/doc/fluid/api/layers.rst +++ b/doc/fluid/api/layers.rst @@ -181,12 +181,6 @@ Print .. autofunction:: paddle.fluid.layers.Print :noindex: -is_empty --------- - -.. autofunction:: paddle.fluid.layers.is_empty - :noindex: - device ====== @@ -261,19 +255,6 @@ double_buffer .. autofunction:: paddle.fluid.layers.double_buffer :noindex: -random_data_generator ---------------------- - -.. autofunction:: paddle.fluid.layers.random_data_generator - :noindex: - -Preprocessor ------------- - -.. autoclass:: paddle.fluid.layers.Preprocessor - :members: - :noindex: - nn == @@ -613,30 +594,6 @@ roi_pool .. autofunction:: paddle.fluid.layers.roi_pool :noindex: -dice_loss ---------- - -.. autofunction:: paddle.fluid.layers.dice_loss - :noindex: - -resize_bilinear ---------------- - -.. autofunction:: paddle.fluid.layers.resize_bilinear - :noindex: - -gather ------- - -.. autofunction:: paddle.fluid.layers.gather - :noindex: - -random_crop ------------ - -.. autofunction:: paddle.fluid.layers.random_crop - :noindex: - ops === @@ -784,12 +741,6 @@ sum .. autofunction:: paddle.fluid.layers.sum :noindex: -shape ------ - -.. autofunction:: paddle.fluid.layers.shape - :noindex: - sigmoid ------- @@ -1039,3 +990,54 @@ zeros .. autofunction:: paddle.fluid.layers.zeros :noindex: +detection +========= + +multi_box_head +-------------- + +.. autofunction:: paddle.fluid.layers.multi_box_head + :noindex: + +bipartite_match +--------------- + +.. autofunction:: paddle.fluid.layers.bipartite_match + :noindex: + +target_assign +------------- + +.. autofunction:: paddle.fluid.layers.target_assign + :noindex: + +detection_output +---------------- + +.. autofunction:: paddle.fluid.layers.detection_output + :noindex: + +ssd_loss +-------- + +.. autofunction:: paddle.fluid.layers.ssd_loss + :noindex: + +detection_map +------------- + +.. autofunction:: paddle.fluid.layers.detection_map + :noindex: + +iou_similarity +-------------- + +.. autofunction:: paddle.fluid.layers.iou_similarity + :noindex: + +box_coder +--------- + +.. autofunction:: paddle.fluid.layers.box_coder + :noindex: + diff --git a/doc/fluid/api/optimizer.rst b/doc/fluid/api/optimizer.rst index 6ad44bb6905b6e3f2b6e4aeb3701ced5d18e2005..79a0995fce303518d989693976c4e92e05795ca2 100644 --- a/doc/fluid/api/optimizer.rst +++ b/doc/fluid/api/optimizer.rst @@ -89,13 +89,6 @@ DecayedAdagradOptimizer :members: :noindex: -RMSPropOptimizer ----------------- - -.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer - :members: - :noindex: - Adadelta -------- diff --git a/doc/fluid/api/profiler.rst b/doc/fluid/api/profiler.rst index 39fda65863471a78895503184848a754828b71a1..74d102dcb0db35766c34e3d14939a8aa5861686b 100644 --- a/doc/fluid/api/profiler.rst +++ b/doc/fluid/api/profiler.rst @@ -23,15 +23,3 @@ profiler .. autofunction:: paddle.fluid.profiler.profiler :noindex: -start_profiler --------------- - -.. autofunction:: paddle.fluid.profiler.start_profiler - :noindex: - -stop_profiler -------------- - -.. autofunction:: paddle.fluid.profiler.stop_profiler - :noindex: - diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 6bc770580640f242cfce6a9838f00210f785010a..6286dda4a54991b7a1042aed9886fdcb694198ba 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -84,7 +84,7 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor) cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog) if(WITH_DISTRIBUTE) - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) else() diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc index 0e4a56d4a45a732cfcf43b09228bc0c44df5924c..8206cc9890160da756efb13c991020f09b20126a 100644 --- a/paddle/fluid/operators/batch_norm_mkldnn_op.cc +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -19,10 +19,17 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; +using batch_norm_bwd = mkldnn::batch_normalization_backward; +using batch_norm_fwd = mkldnn::batch_normalization_forward; +using framework::DataLayout; +using framework::Tensor; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; using paddle::platform::MKLDNNDeviceContext; using paddle::platform::MKLDNNMemDesc; -using mkldnn::memory; +using platform::to_void_cast; template using EigenArrayMap = @@ -64,21 +71,12 @@ void run_batch_norm_op(Args &&... args) { mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } -template -inline void *cast_const_to_void(const T *t) { - return static_cast(const_cast(t)); -} } // namespace template class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { - auto data_layout_str = ctx.Attr("data_layout"); - auto data_layout = framework::StringToDataLayout(data_layout_str); - PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, - "MKLDNN batch normalization handles only NCHW data layout"); - const float epsilon = ctx.Attr("epsilon"); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); @@ -99,41 +97,53 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { const auto *scale = ctx.Input("Scale"); const auto *shift = ctx.Input("Bias"); - y->mutable_data(ctx.GetPlace()); - mean_out->mutable_data(ctx.GetPlace()); - variance_out->mutable_data(ctx.GetPlace()); + PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN && + x->format() != memory::format::format_undef, + "Wrong layout/format set for Input x tensor"); + + const T *x_data = x->data(); + const T *mean_data = mean->data(); + const T *variance_data = variance->data(); + T *y_data = y->mutable_data(ctx.GetPlace()); + T *mean_out_data = mean_out->mutable_data(ctx.GetPlace()); + T *variance_out_data = variance_out->mutable_data(ctx.GetPlace()); + T *batch_mean_data = nullptr; + T *batch_variance_data = nullptr; if (!is_test) { - batch_mean->mutable_data(ctx.GetPlace()); - batch_variance->mutable_data(ctx.GetPlace()); + batch_mean_data = batch_mean->mutable_data(ctx.GetPlace()); + batch_variance_data = batch_variance->mutable_data(ctx.GetPlace()); } auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring : mkldnn::prop_kind::forward_training; - auto dims = paddle::framework::vectorize2int(x->dims()); - - auto src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - - auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; - auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine}; - - auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data())}; - auto dst = mkldnn::memory{dst_pd, y->data()}; + auto src_tz = paddle::framework::vectorize2int(x->dims()); + auto scale_tz = paddle::framework::vectorize2int(scale->dims()); + PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1"); + const unsigned int ic = scale_tz[0]; unsigned flags = mkldnn::use_scale_shift; if (is_test) flags |= mkldnn::use_global_stats; + // create mkldnn memory from input x tensor + auto src_memory = + memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine}, + to_void_cast(x_data)); + + // create primitive descriptor for batch norm forward using bn_fwd_types = bn_type_traits; - auto batch_norm_fwd_desc = - bn_fwd_types::op_desc{propagation, src_md, epsilon, flags}; - auto batch_norm_fwd_pd = - bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ + propagation, src_memory.get_primitive_desc().desc(), epsilon, flags}; + std::shared_ptr batch_norm_fwd_pd = + std::shared_ptr( + new batch_norm_fwd::primitive_desc(batch_norm_fwd_desc, + mkldnn_engine)); - const unsigned int ic = dims[1]; + // Save the pd to be used in backward pass + const std::string key = ctx.op().Output("SavedMean"); + const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; + dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd); // MKLDNN requires a single piece of memory for scale and shift/bias data const size_t scaleshift_size = 2 * ic; @@ -143,73 +153,58 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { copy_to_weights(scale->data(), scale->data() + ic, shift->data(), shift->data() + ic, &scaleshift_data); - auto scaleshift_memory = mkldnn::memory{ - batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + // crate mkldnn memory for weights(scale/shift) + auto scaleshift_memory = memory(batch_norm_fwd_pd->weights_primitive_desc(), + scaleshift_data.data()); - if (is_test) { - auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), - cast_const_to_void(mean->data())}; + // create mkldnn memory for output y tensor + auto dst_memory = memory(batch_norm_fwd_pd->dst_primitive_desc(), y_data); + if (is_test) { + // create mkldnn memory for stats (as input) + auto mean_memory = memory(batch_norm_fwd_pd->mean_primitive_desc(), + to_void_cast(mean_data)); auto variance_memory = - mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), - cast_const_to_void(variance->data())}; + memory(batch_norm_fwd_pd->variance_primitive_desc(), + to_void_cast(variance_data)); run_batch_norm_op( - batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory, + *batch_norm_fwd_pd, src_memory, + (const mkldnn::primitive::at &)mean_memory, (const mkldnn::primitive::at &)variance_memory, scaleshift_memory, - dst); + dst_memory); } else { + // create mkldnn memory for stats (as output) auto mean_memory = - mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), - cast_const_to_void(batch_mean->data())}; - - auto variance_memory = - mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), - cast_const_to_void(batch_variance->data())}; + memory(batch_norm_fwd_pd->mean_primitive_desc(), batch_mean_data); + auto variance_memory = memory( + batch_norm_fwd_pd->variance_primitive_desc(), batch_variance_data); - run_batch_norm_op(batch_norm_fwd_pd, src, - scaleshift_memory, dst, + run_batch_norm_op(*batch_norm_fwd_pd, src_memory, + scaleshift_memory, dst_memory, mean_memory, variance_memory); } if (!is_test) { - const unsigned int in = dims[0]; - const unsigned int sample_size = x->numel() / in / ic; - - // saved_xx is use just in this batch of data - EigenVectorArrayMap saved_mean_e( - batch_mean->mutable_data(ctx.GetPlace()), ic); - EigenVectorArrayMap saved_variance_e( - batch_variance->mutable_data(ctx.GetPlace()), ic); - saved_mean_e.setZero(); - saved_variance_e.setZero(); - - const unsigned int x_arr_size = in * ic; - ConstEigenArrayMap x_arr(x->data(), sample_size, x_arr_size); - for (unsigned int nc = 0; nc < x_arr_size; ++nc) { - saved_mean_e(nc % ic) += x_arr.col(nc).sum(); - } - saved_mean_e /= in * sample_size; - for (unsigned int nc = 0; nc < x_arr_size; ++nc) { - saved_variance_e(nc % ic) += - (x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm(); - } - saved_variance_e /= in * sample_size; - - ConstEigenVectorArrayMap mean_arr{mean->data(), ic}; - ConstEigenVectorArrayMap variance_arr{variance->data(), ic}; - - EigenVectorArrayMap running_mean_arr( - mean_out->mutable_data(ctx.GetPlace()), ic); - EigenVectorArrayMap running_var_arr( - variance_out->mutable_data(ctx.GetPlace()), ic); + // mkldnn only compute stats for current batch + // so we need compute momentum stats via Eigen lib + EigenVectorArrayMap batch_mean_e(batch_mean_data, ic); + EigenVectorArrayMap batch_variance_e(batch_variance_data, ic); + ConstEigenVectorArrayMap mean_e(mean_data, ic); + ConstEigenVectorArrayMap variance_e{variance_data, ic}; + + EigenVectorArrayMap running_mean_e(mean_out_data, ic); + EigenVectorArrayMap running_variance_e(variance_out_data, ic); auto one_minus_momentum = 1. - momentum; - running_mean_arr = - mean_arr * momentum + saved_mean_e * one_minus_momentum; - running_var_arr = - variance_arr * momentum + saved_variance_e * one_minus_momentum; + running_mean_e = mean_e * momentum + batch_mean_e * one_minus_momentum; + running_variance_e = + variance_e * momentum + batch_variance_e * one_minus_momentum; } + + y->set_layout(DataLayout::kMKLDNN); + y->set_format( + (memory::format)dst_memory.get_primitive_desc().desc().data.format); } }; @@ -217,11 +212,6 @@ template class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { public: void Compute(const paddle::framework::ExecutionContext &ctx) const override { - auto data_layout_str = ctx.Attr("data_layout"); - auto data_layout = framework::StringToDataLayout(data_layout_str); - PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, - "MKLDNN batch normalization handles only NCHW data layout"); - auto &dev_ctx = ctx.template device_context(); auto mkldnn_engine = dev_ctx.GetEngine(); @@ -238,88 +228,132 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto *diff_scale = ctx.Output(framework::GradVarName("Scale")); auto *diff_shift = ctx.Output(framework::GradVarName("Bias")); - diff_x->mutable_data(ctx.GetPlace()); - diff_scale->mutable_data(ctx.GetPlace()); - diff_shift->mutable_data(ctx.GetPlace()); + PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN && + diff_y->format() != memory::format::format_undef, + "Wrong layout/format set for Input diff_y tensor"); + + const T *x_data = x->data(); + const T *diff_y_data = diff_y->data(); + const T *batch_mean_data = batch_mean->data(); + const T *batch_variance_data = batch_variance->data(); + const T *scale_data = scale->data(); + const T *shift_data = shift->data(); + T *diff_x_data = diff_x->mutable_data(ctx.GetPlace()); + T *diff_scale_data = diff_scale->mutable_data(ctx.GetPlace()); + T *diff_shift_data = diff_shift->mutable_data(ctx.GetPlace()); + + auto src_tz = paddle::framework::vectorize2int(x->dims()); + auto diff_src_tz = src_tz; + auto dst_tz = src_tz; + auto diff_dst_tz = dst_tz; + auto scale_tz = paddle::framework::vectorize2int(scale->dims()); + PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1"); + + const unsigned int ic = scale_tz[0]; + + // Retrieve bn_fwd_pd from device context + const std::string key = ctx.op().Input("SavedMean"); + const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; + auto batch_norm_fwd_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_batch_norm_fwd_pd)); + PADDLE_ENFORCE(batch_norm_fwd_pd != nullptr, + "Fail to find batch_norm_fwd_pd in device context"); - auto dims = paddle::framework::vectorize2int(x->dims()); - unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats; + using bn_bwd_types = bn_type_traits; - auto src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto diff_src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto diff_dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + // create mkldnn memory from input diff_y tensor + auto user_diff_dst_memory = + memory({{{diff_dst_tz}, memory::data_type::f32, diff_y->format()}, + mkldnn_engine}, + to_void_cast(diff_y_data)); - using bn_bwd_types = bn_type_traits; - using bn_fwd_types = bn_type_traits; + // create mkldnn memory from input x tensor + auto src_memory = + memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine}, + to_void_cast(x_data)); - auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ - mkldnn::prop_kind::forward_training, src_md, epsilon, flags}; - auto batch_norm_fwd_pd = - bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + // for diff_dst, try to use same format as dst in forward pass + auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc(); + auto diff_dst_md = diff_dst_pd.desc(); + // create primitive descriptor for batch norm backward + unsigned flags = mkldnn::use_scale_shift; auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ - mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags}; + mkldnn::prop_kind::backward, diff_dst_md, + src_memory.get_primitive_desc().desc(), epsilon, flags}; auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ - batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd}; - - auto src = mkldnn::memory{{src_md, mkldnn_engine}, - cast_const_to_void(x->data())}; - - auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(), - cast_const_to_void(batch_mean->data())}; - - auto variance = - mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(), - cast_const_to_void(batch_variance->data())}; - - auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine}, - cast_const_to_void(diff_y->data())}; + batch_norm_bwd_desc, mkldnn_engine, *batch_norm_fwd_pd}; + + // reorder user_diff_dst if it's not in preferred format + auto diff_dst_memory = user_diff_dst_memory; + primitive reorder_diff_dst; + bool is_diff_dst_reordered = false; + if (diff_dst_pd != user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = memory(diff_dst_pd); + reorder_diff_dst = reorder(user_diff_dst_memory, diff_dst_memory); + is_diff_dst_reordered = true; + } - const unsigned int ic = dims[1]; + // create mkldnn memory for input tensors (src/mean/variance) + auto mean_memory = memory(batch_norm_bwd_pd.mean_primitive_desc(), + to_void_cast(batch_mean_data)); + auto variance_memory = memory(batch_norm_bwd_pd.variance_primitive_desc(), + to_void_cast(batch_variance_data)); + // MKLDNN requires a single piece of memory for scale and shift/bias data const size_t scaleshift_size = 2 * ic; std::vector scaleshift_data; scaleshift_data.reserve(scaleshift_size); - copy_to_weights(scale->data(), scale->data() + ic, shift->data(), - shift->data() + ic, &scaleshift_data); + copy_to_weights(scale_data, scale_data + ic, shift_data, shift_data + ic, + &scaleshift_data); - auto scaleshift_memory = mkldnn::memory{ - batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + // create mkldnn memory for input tensors (scale/shift) + auto scaleshift_memory = memory(batch_norm_bwd_pd.weights_primitive_desc(), + scaleshift_data.data()); + // create mkldnn memory for output diff weights (combined scale/shift) std::vector diff_scaleshift_data; diff_scaleshift_data.reserve(scaleshift_size); - copy_to_weights(diff_scale->data(), diff_scale->data() + ic, - diff_shift->data(), diff_shift->data() + ic, - &diff_scaleshift_data); - auto diff_scaleshift_memory = - mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(), - diff_scaleshift_data.data()}; - - auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine}, - static_cast(diff_x->data())}; - - run_batch_norm_op( - batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory, - diff_src, diff_scaleshift_memory); - + memory(batch_norm_bwd_pd.diff_weights_primitive_desc(), + diff_scaleshift_data.data()); + + // here assume diff_src is in the same format of src + auto diff_src_memory = memory(src_memory.get_primitive_desc(), diff_x_data); + + // finally create batch_norm backward primitive + auto batch_norm_bwd_prim = + batch_norm_bwd(batch_norm_bwd_pd, src_memory, mean_memory, + variance_memory, diff_dst_memory, scaleshift_memory, + diff_src_memory, diff_scaleshift_memory); + + // execute optional reorder and batch_norm backward primitive + std::vector pipeline; + if (is_diff_dst_reordered) pipeline.push_back(reorder_diff_dst); + pipeline.push_back(batch_norm_bwd_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + // copy back diff sacle/shift to output tensors (diff scale/shift) + diff_scaleshift_data.resize(scaleshift_size); auto it = std::begin(diff_scaleshift_data); - std::copy(it, std::next(it, ic), diff_scale->data()); + std::copy(it, std::next(it, ic), diff_scale_data); std::copy(std::next(it, ic), std::end(diff_scaleshift_data), - diff_shift->data()); + diff_shift_data); + + // set layout/format of output tensors + diff_x->set_layout(DataLayout::kMKLDNN); + diff_x->set_format((memory::format)diff_src_memory.get_primitive_desc() + .desc() + .data.format); } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace, +REGISTER_OP_KERNEL(batch_norm, MKLDNN, ::paddle::platform::CPUPlace, ops::BatchNormMKLDNNOpKernel); -REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace, +REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, ::paddle::platform::CPUPlace, ops::BatchNormMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 92fbb9adaf6a6a335abee3c9443d4b1d6097021b..625ca2d7c4c70d1098b0fb28380d8d1eb24cb338 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -110,19 +110,19 @@ class BatchNormOp : public framework::OperatorWithKernel { ctx.Input("Variance")->type()), "Variance input should be of float type"); - framework::LibraryType library_{framework::LibraryType::kPlain}; // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; framework::DataLayout layout = framework::DataLayout::kAnyLayout; - #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; + library = framework::LibraryType::kMKLDNN; layout = framework::DataLayout::kMKLDNN; } #endif + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, - library_); + library); } }; @@ -370,19 +370,21 @@ class BatchNormGradOp : public framework::OperatorWithKernel { PADDLE_THROW("can't find Y@GRAD"); } - framework::LibraryType library_{framework::LibraryType::kPlain}; // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout_ = framework::DataLayout::kAnyLayout; + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; - layout_ = framework::DataLayout::kMKLDNN; + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif + return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout_, library_); + layout, library); } }; diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc index c202eed354c5f1a91e93e1c3919d1bfebc1bc401..40dc7c9a0b6a40f2419ace3ce7e0e5e82bc95c1a 100644 --- a/paddle/fluid/platform/cpu_info.cc +++ b/paddle/fluid/platform/cpu_info.cc @@ -59,9 +59,10 @@ inline size_t CpuTotalPhysicalMemory() { size_t CpuMaxAllocSize() { // For distributed systems, it requires configuring and limiting // the fraction of memory to use. - return std::min(static_cast(FLAGS_fraction_of_cpu_memory_to_use * - CpuTotalPhysicalMemory()), - FLAGS_initial_cpu_memory_in_mb * 1 << 20); + return std::min( + static_cast(FLAGS_fraction_of_cpu_memory_to_use * + CpuTotalPhysicalMemory()), + static_cast(FLAGS_initial_cpu_memory_in_mb * 1 << 20)); } size_t CpuMinChunkSize() { diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index bbd35aaecba27ea9fd66b9be585a972690980ab8..f6438c82ac207d0e38d8be5e9d6252b28e72826e 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -382,7 +382,7 @@ class Operator(object): 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv', 'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine', 'ncclInit', 'channel_create', 'channel_close', 'channel_send', - 'channel_recv', 'select' + 'channel_recv', 'select', 'gen_nccl_id' } def __init__(self, diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 7a4dc3d7be1d4c3e0ddd8b639378b72aa9fef6e5..183231dab5c40a6d5a08f238fe5b1ad334efda9e 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -261,9 +261,10 @@ def embedding(input, return tmp -# TODO(qijun): expose H0 and C0 def dynamic_lstm(input, size, + h_0=None, + c_0=None, param_attr=None, bias_attr=None, use_peepholes=True, @@ -324,6 +325,13 @@ def dynamic_lstm(input, (T X 4D), where T is the total time steps in this mini-batch, D is the hidden size. size(int): 4 * hidden size. + h_0(Variable): The initial hidden state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size and D is the hidden size. + c_0(Variable): The initial cell state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size. `h_0` and `c_0` can be NULL but only at the same time. + param_attr(ParamAttr|None): The parameter attribute for the learnable hidden-hidden weights. @@ -387,12 +395,20 @@ def dynamic_lstm(input, cell = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype) batch_cell_pre_act = helper.create_tmp_variable(dtype) + inputs = {'Input': input, 'Weight': weight, 'Bias': bias} + batch_size = input.shape[0] + if h_0: + assert h_0.shape == (batch_size, size), \ + 'The shape of h0 should be (batch_size, %d)' % size + inputs['H0'] = h_0 + if c_0: + assert c_0.shape == (batch_size, size), \ + 'The shape of c0 should be (batch_size, %d)' % size + inputs['C0'] = c_0 helper.append_op( type='lstm', - inputs={'Input': input, - 'Weight': weight, - 'Bias': bias}, + inputs=inputs, outputs={ 'Hidden': hidden, 'Cell': cell, @@ -677,11 +693,13 @@ def dynamic_gru(input, attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) bias = helper.create_parameter( attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True) + batch_size = input.shape[0] inputs = {'Input': input, 'Weight': weight, 'Bias': bias} if h_0 != None: assert h_0.shape == ( - size, size), 'The shape of h0 should be(%d, %d)' % (size, size) - inputs['h0'] = h_0 + batch_size, size + ), 'The shape of h0 should be(batch_size, %d)' % size + inputs['H0'] = h_0 hidden = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype)