diff --git a/paddle/gserver/layers/BatchNormBaseLayer.cpp b/paddle/gserver/layers/BatchNormBaseLayer.cpp index bc7d1c83a48aefeb4bc6d3baa32b78aba712e58d..925af31289d0c8ca534a30a16b14bfd2df90b013 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.cpp +++ b/paddle/gserver/layers/BatchNormBaseLayer.cpp @@ -41,6 +41,7 @@ bool BatchNormBaseLayer::init(const LayerMap& layerMap, useGlobalStats_ = config_.use_global_stats(); } movingAvgFraction_ = config_.moving_average_fraction(); + epsilon_ = config_.epsilon(); weight_.reset(new Weight(1, channels_, parameters_[0])); movingMean_.reset(new Weight(1, channels_, parameters_[1])); diff --git a/paddle/gserver/layers/BatchNormBaseLayer.h b/paddle/gserver/layers/BatchNormBaseLayer.h index e721d2d267a31cae46407673b8b1281e87055608..2ac3cd9d670d0fcf9c40ad2f117d5a72479663a3 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.h +++ b/paddle/gserver/layers/BatchNormBaseLayer.h @@ -94,6 +94,8 @@ protected: bool useGlobalStats_; // use to compute moving mean and variance. real movingAvgFraction_; + // Epsilon is a small random noise used in batch normalization for stability. + real epsilon_; }; } // namespace paddle diff --git a/paddle/gserver/layers/BatchNormalizationLayer.cpp b/paddle/gserver/layers/BatchNormalizationLayer.cpp index dacff25e5927daf9c991577a71be86b160228317..25ab5cd927792d18f78bc1fa33eee4029b427cc7 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.cpp +++ b/paddle/gserver/layers/BatchNormalizationLayer.cpp @@ -22,8 +22,6 @@ namespace paddle { REGISTER_LAYER(batch_norm, BatchNormalizationLayer); -const real BatchNormalizationLayer::EPS = 1E-5; - bool BatchNormalizationLayer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { /* Initialize the basic parent class */ @@ -53,7 +51,7 @@ void BatchNormalizationLayer::calMeanAndStd(const MatrixPtr& mat) { calMovingMeanAndVar(); - savedInvVar_->subScalar(-EPS); + savedInvVar_->subScalar(-epsilon_); savedInvVar_->sqrt2(*savedInvVar_); } @@ -74,7 +72,7 @@ void BatchNormalizationLayer::setMeanAndStd() { savedInvVar_->copyFrom(*(movingVar_->getW())); savedInvVar_->downClip(real(0.0)); - savedInvVar_->subScalar(-EPS); + savedInvVar_->subScalar(-epsilon_); savedInvVar_->sqrt2(*savedInvVar_); } diff --git a/paddle/gserver/layers/BatchNormalizationLayer.h b/paddle/gserver/layers/BatchNormalizationLayer.h index f6115801fc6b341c0718f8851617de43bdeeec09..1fdb5e2070259a14ab6f70957c9cf03f0699f734 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.h +++ b/paddle/gserver/layers/BatchNormalizationLayer.h @@ -39,9 +39,6 @@ public: void backward(const UpdateCallback& callback = nullptr) override; protected: - /// Epsilon value used in the batch normalization formula. - static const real EPS; - /// Load pre-calculated mean and std. void setMeanAndStd(); diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp index 49a9540c0b6e36b59ed786287ff5c4569b69a6a5..8390b55026c895b661cb514714ba92c05a7bf02e 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp +++ b/paddle/gserver/layers/CudnnBatchNormLayer.cpp @@ -21,8 +21,6 @@ namespace paddle { REGISTER_LAYER(cudnn_batch_norm, CudnnBatchNormLayer); -const double CudnnBatchNormLayer::EPS = 1E-5; - bool CudnnBatchNormLayer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { /* Initialize the basic parent class */ @@ -61,6 +59,9 @@ void CudnnBatchNormLayer::forward(PassType passType) { real* movingMean = movingMean_->getW()->getData(); real* movingVar = movingVar_->getW()->getData(); + // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. + eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); + if (!useGlobalStats_) { REGISTER_TIMER_INFO("CudnnBatchFwTimer", getName().c_str()); real* savedMean = savedMean_->getData(); @@ -75,7 +76,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { 1.0 - movingAvgFraction_, movingMean, movingVar, - EPS, + eps_, savedMean, savedInvVar); } else { @@ -90,7 +91,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { beta, movingMean, movingVar, - EPS); + eps_); } else { // There is a limitation in cudnn library. // When the batch size is larger than 1024 in cuDNN v5.1, @@ -101,7 +102,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { beta, movingMean, movingVar, - EPS, + eps_, batchSize, channels_, imageH_ * imageD_, @@ -128,6 +129,9 @@ void CudnnBatchNormLayer::backward(const UpdateCallback& callback) { real* savedMean = savedMean_->getData(); real* savedInvVar = savedInvVar_->getData(); + // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. + eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); + auto create = [](MatrixPtr& m, size_t h, size_t w, real** p) { Matrix::resizeOrCreate(m, h, w, false, true); m->zeroMem(); @@ -157,7 +161,7 @@ void CudnnBatchNormLayer::backward(const UpdateCallback& callback) { gamma, gammaGrad, betaGrad, - EPS, + eps_, savedMean, savedInvVar); diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.h b/paddle/gserver/layers/CudnnBatchNormLayer.h index 413efd4d3ecd734b343efbcf8328ac0592daddda..1a3f0c0cbf8a1540e77cef70c753c91298728484 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.h +++ b/paddle/gserver/layers/CudnnBatchNormLayer.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "BatchNormBaseLayer.h" #include "Layer.h" #include "paddle/utils/Stat.h" @@ -46,12 +47,9 @@ public: void backward(const UpdateCallback& callback = nullptr) override; protected: - /** - * Epsilon value used in the batch normalization formula. - * Minimum allowed value is CUDNN_BN_MIN_EPSILON defined in cudnn.h. - * Same epsilon value should be used in forward and backward functions. - */ - static const double EPS; + /// Epsilon value used in the batch normalization formula. + /// Same epsilon value should be used in forward and backward functions. + double eps_; /// Input/output tensor descriptor desc hl_tensor_descriptor ioDesc_; diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp b/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp index d66c361ae05e4a1089786e4620d2eb2218a8a29c..7faca0f8b7f54fa0a09e8fdab11064c8c26df375 100644 --- a/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp +++ b/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp @@ -21,8 +21,6 @@ namespace paddle { REGISTER_LAYER(mkldnn_batch_norm, MKLDNNBatchNormLayer); -const real MKLDNNBatchNormLayer::EPS = 1E-5; - bool MKLDNNBatchNormLayer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { if (!MKLDNNLayer::init(layerMap, parameterMap)) { @@ -50,6 +48,8 @@ bool MKLDNNBatchNormLayer::init(const LayerMap& layerMap, useGlobalStats_ = config_.use_global_stats(); } movingAvgFraction_ = config_.moving_average_fraction(); + epsilon_ = config_.epsilon(); + VLOG(MKLDNN_BASE) << "--- " << (useGlobalStats_ ? "use" : "do not use") << " --- global stats"; VLOG(MKLDNN_BASE) << "Moving average fraction: " << movingAvgFraction_; @@ -210,7 +210,7 @@ void MKLDNNBatchNormLayer::resetFwdPD( if (wgt) { flags_ = (flags_ | batch_normalization_flag::use_scale_shift); } - auto fwdDesc = bn_fwd::desc(pk, in->getMemoryDesc(), EPS, flags_); + auto fwdDesc = bn_fwd::desc(pk, in->getMemoryDesc(), epsilon_, flags_); pd.reset(new bn_fwd::primitive_desc(fwdDesc, engine_)); CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc()); if (wgt) { @@ -277,7 +277,7 @@ void MKLDNNBatchNormLayer::resetBwdPD( } CHECK_PRIMITIVE_DESC_EQ(out, in->getPrimitiveDesc()); auto md = in->getMemoryDesc(); - auto bwdDesc = bn_bwd::desc(prop_kind::backward, md, md, EPS, flags_); + auto bwdDesc = bn_bwd::desc(prop_kind::backward, md, md, epsilon_, flags_); pd.reset(new bn_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); CHECK(pd->weights_primitive_desc() == fwdPD_->weights_primitive_desc()); CHECK_PRIMITIVE_DESC_EQ(wgt, pd->diff_weights_primitive_desc()); diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.h b/paddle/gserver/layers/MKLDNNBatchNormLayer.h index 387c58f02298b0441cc3bbbc4879eed6d892164c..1cf33cb34fa9cd7c9b8487a0a4a0011fb129e311 100644 --- a/paddle/gserver/layers/MKLDNNBatchNormLayer.h +++ b/paddle/gserver/layers/MKLDNNBatchNormLayer.h @@ -32,7 +32,8 @@ protected: std::shared_ptr fwdPD_; // Epsilon value used in the batch normalization formula. - static const real EPS; + real epsilon_; + // weight and bias in paddle std::unique_ptr weight_; std::unique_ptr biases_; diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index 2c2cc6245932d4af56a68d6399ce31f008bf3748..e2f5592248fd0b6166c2d11af02cef7815673def 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -540,6 +540,10 @@ message LayerConfig { // for switch order layer optional ReshapeConfig reshape_conf = 59; + + // for batch normalization layer + // The small constant added to the variance to improve numeric stability. + optional double epsilon = 60 [ default = 0.00001 ]; } message EvaluatorConfig { diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index 0941f10cf1ef337ac0e0225aea250dcdd8a27614..5ba0e50c6ba0f84a3ea87d5a5199fef23a5b05ea 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -2412,6 +2412,7 @@ class BatchNormLayer(LayerBase): bias=True, img3D=False, use_global_stats=True, + epsilon=1e-5, moving_average_fraction=0.9, batch_norm_type=None, mean_var_names=None, @@ -2460,6 +2461,9 @@ class BatchNormLayer(LayerBase): self.config.use_global_stats = use_global_stats if moving_average_fraction is not None: self.config.moving_average_fraction = moving_average_fraction + if epsilon is not None: + assert epsilon >= 1e-5, "epsilon must be no less than 1e-5." + self.config.epsilon = epsilon input_layer = self.get_input_layer(0) image_conf = self.config.inputs[0].image_conf diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 6bd5ce4fe2f70946befb388986dff603bdae0b8e..8e127c9489ca5a4ed190e6d4e12ec4c9b28ad9cf 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -3118,6 +3118,7 @@ def batch_norm_layer(input, param_attr=None, layer_attr=None, batch_norm_type=None, + epsilon=1e-5, moving_average_fraction=0.9, use_global_stats=None, mean_var_names=None): @@ -3188,6 +3189,8 @@ def batch_norm_layer(input, will use the mean and variance of the current batch of test data. :type use_global_stats: bool | None. + :param epsilon: The small constant added to the variance to improve numeric stability. + :type epsilon: float. :param moving_average_fraction: Factor used in the moving average computation. :math:`runningMean = newMean*(1-factor) + runningMean*factor` :type moving_average_fraction: float. @@ -3205,6 +3208,7 @@ def batch_norm_layer(input, assert (batch_norm_type is None) or (batch_norm_type == "batch_norm") or \ (batch_norm_type == "mkldnn_batch_norm") or \ (batch_norm_type == "cudnn_batch_norm") + l = Layer( name=name, img3D=img3D, @@ -3214,6 +3218,7 @@ def batch_norm_layer(input, type=LayerType.BATCH_NORM_LAYER, batch_norm_type=batch_norm_type, bias=ParamAttr.to_bias(bias_attr), + epsilon=epsilon, moving_average_fraction=moving_average_fraction, use_global_stats=use_global_stats, mean_var_names=mean_var_names, diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr index b14121e82cb7d9516c4771fc896b9b3b9e01d1c8..3e0f957648879d4350d662b336c953273bac1378 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr @@ -65,6 +65,7 @@ layers { height: 227 width: 227 depth: 1 + epsilon: 1e-05 } layers { name: "__crmnorm_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr index c7a487a11231cba6182b654108773037bdb0ec35..a18a4652e14c0cfc4dbca87e67d31aa663ee756b 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr @@ -65,6 +65,7 @@ layers { height: 256 width: 256 depth: 1 + epsilon: 1e-05 } layers { name: "__crmnorm_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr index 832ed24a31dd2bedba9a4fce77d7a088d1796fdb..9b69ae4a3b3cbcc7c0c69a2d5b3728e2f0204f33 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr @@ -36,6 +36,7 @@ layers { height: 6 width: 20 depth: 3 + epsilon: 1e-05 } parameters { name: "___batch_norm_0__.w0"