From a32d420043e50d3bcc8564d768584a9b1f40b5af Mon Sep 17 00:00:00 2001 From: flame Date: Thu, 7 Mar 2019 20:39:06 +0800 Subject: [PATCH] cherry-pick from feature/anakin-engine: batch norm (#16110) * use anakin batch norm and scale implement fluid batch norm --- .../inference/anakin/convert/batch_norm.cc | 97 ++++++++++--------- .../anakin/convert/test_batch_norm_op.cc | 1 - 2 files changed, 50 insertions(+), 48 deletions(-) diff --git a/paddle/fluid/inference/anakin/convert/batch_norm.cc b/paddle/fluid/inference/anakin/convert/batch_norm.cc index 5ebcefb8a2..bb5d406f98 100644 --- a/paddle/fluid/inference/anakin/convert/batch_norm.cc +++ b/paddle/fluid/inference/anakin/convert/batch_norm.cc @@ -41,16 +41,15 @@ void BatchNormOpConverter::operator()(const framework::proto::OpDesc &op, auto output = op_desc.Output("Y").front(); auto op_name = op_desc.Type() + ":" + op_desc.Output("Y").front(); - engine_->AddOp(op_name, "Scale", {inputs["X"]}, {output}); - engine_->AddOpAttr(op_name, "bias_term", true); - engine_->AddOpAttr(op_name, "axis", 1); - engine_->AddOpAttr(op_name, "num_axes", 1); - bool is_test = boost::get(op_desc.GetAttr("is_test")); - PADDLE_ENFORCE(is_test); - float epsilon = boost::get(op_desc.GetAttr("epsilon")); - engine_->AddOpAttr(op_name, "epsilon", epsilon); + auto epsilon = boost::get(op_desc.GetAttr("epsilon")); + + auto bn_op_name = op_name + ":bn"; + auto bn_output = bn_op_name + "_output"; + engine_->AddOp(bn_op_name, "BatchNorm", {inputs["X"]}, {bn_output}); + engine_->AddOpAttr(bn_op_name, "epsilon", epsilon); + auto scale_op_name = op_name + ":scale"; auto get_lod_tensor = [this, &scope, &op_name](const std::string &var_name, framework::LoDTensor *tensor) { auto *v = scope.FindVar(var_name); @@ -69,50 +68,54 @@ void BatchNormOpConverter::operator()(const framework::proto::OpDesc &op, get_lod_tensor(inputs["Scale"], &scale_t); get_lod_tensor(inputs["Variance"], &variance_t); - auto *bias = bias_t.mutable_data(platform::CPUPlace()); - auto *mean = mean_t.mutable_data(platform::CPUPlace()); - auto *scale = scale_t.mutable_data(platform::CPUPlace()); - auto *variance = variance_t.mutable_data(platform::CPUPlace()); - - framework::LoDTensor combile_scale_t; - framework::LoDTensor combile_bias_t; - combile_scale_t.Resize(scale_t.dims()); - combile_bias_t.Resize(bias_t.dims()); - - auto *combile_scale = - combile_scale_t.mutable_data(platform::CPUPlace()); - auto *combile_bias = combile_bias_t.mutable_data(platform::CPUPlace()); - - size_t elem_num = combile_scale_t.memory_size() / sizeof(float); - for (size_t i = 0; i < elem_num; i++) { - combile_scale[i] = scale[i] / sqrtf(variance[i] + epsilon); - combile_bias[i] = bias[i] - mean[i] * combile_scale[i]; - } - - auto fill_shape = [](size_t n, std::vector *shape) { - shape->insert(shape->begin(), 1); - if (shape->size() < n) { - shape->insert(shape->end(), n - shape->size(), 1); + auto fill_shape = [](size_t n, std::vector shape) { + shape.insert(shape.begin(), 1); + if (shape.size() < n) { + shape.insert(shape.end(), n - shape.size(), 1); } + return shape; }; - auto scale_shape = framework::vectorize2int(combile_scale_t.dims()); - auto bias_shape = framework::vectorize2int(combile_bias_t.dims()); - fill_shape(4, &scale_shape); - fill_shape(4, &bias_shape); - Shape weight1_shape(scale_shape); - Shape weight2_shape(bias_shape); + Shape shape1(fill_shape(4, framework::vectorize2int(mean_t.dims()))); + Shape shape2(fill_shape(4, framework::vectorize2int(variance_t.dims()))); auto *weight1 = - GraphGlobalMem::Global().template new_block(weight1_shape); - auto *scale_data = static_cast(weight1->h_tensor().mutable_data()); - std::copy_n(combile_scale_t.data(), combile_scale_t.numel(), - scale_data); - engine_->AddOpAttr(op_name, "weight_1", *weight1); + GraphGlobalMem::Global().template new_block(shape1); + auto *mean_data = static_cast(weight1->h_tensor().mutable_data()); + std::copy_n(mean_t.data(), mean_t.numel(), mean_data); + engine_->AddOpAttr(bn_op_name, "weight_1", *weight1); auto *weight2 = - GraphGlobalMem::Global().template new_block(weight2_shape); - auto *bias_data = static_cast(weight2->h_tensor().mutable_data()); - std::copy_n(combile_bias_t.data(), combile_bias_t.numel(), bias_data); - engine_->AddOpAttr(op_name, "weight_2", *weight2); + GraphGlobalMem::Global().template new_block(shape2); + auto *variance_data = + static_cast(weight2->h_tensor().mutable_data()); + std::copy_n(variance_t.data(), variance_t.numel(), variance_data); + engine_->AddOpAttr(bn_op_name, "weight_2", *weight2); + + Shape shape3(std::vector({1, 1, 1, 1})); + auto *weight3 = + GraphGlobalMem::Global().template new_block(shape3); + auto *alpha_data = static_cast(weight3->h_tensor().mutable_data()); + float weight3_data[] = {1}; + std::copy(std::begin(weight3_data), std::end(weight3_data), alpha_data); + engine_->AddOpAttr(bn_op_name, "weight_3", *weight3); + + Shape scale_shape(fill_shape(4, framework::vectorize2int(scale_t.dims()))); + auto *scale = + GraphGlobalMem::Global().template new_block(scale_shape); + auto *scale_data = static_cast(scale->h_tensor().mutable_data()); + std::copy_n(scale_t.data(), scale_t.numel(), scale_data); + + Shape bias_shape(fill_shape(4, framework::vectorize2int(bias_t.dims()))); + auto *bias = + GraphGlobalMem::Global().template new_block(bias_shape); + auto *bias_data = static_cast(bias->h_tensor().mutable_data()); + std::copy_n(bias_t.data(), bias_t.numel(), bias_data); + + engine_->AddOp(scale_op_name, "Scale", {bn_output}, {output}); + engine_->AddOpAttr(scale_op_name, "axis", 1); + engine_->AddOpAttr(scale_op_name, "num_axes", 1); + engine_->AddOpAttr(scale_op_name, "bias_term", true); + engine_->AddOpAttr(scale_op_name, "weight_1", *scale); + engine_->AddOpAttr(scale_op_name, "weight_2", *bias); } } // namespace anakin diff --git a/paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc b/paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc index 9504b2bf89..c6eebf5d0c 100644 --- a/paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc +++ b/paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc @@ -54,7 +54,6 @@ TEST(batch_norm_op, test) { float eps = 1e-5f; desc.SetAttr("epsilon", eps); desc.SetAttr("is_test", true); - // desc.SetAttr("momentum", 0.8f); validator.SetOp(*desc.Proto()); -- GitLab