misc changes to reduce binary size.
diff --git a/build.py b/build.py
index fce583c..331712d 100644
--- a/build.py
+++ b/build.py
@@ -41,6 +41,11 @@
# True.
USE_GLOG = False
+ # Whether to use RTTI or not. Note that this might not always work; to
+ # disable RTTI means that all your dependencies, most notably protobuf, have
+ # to be built without RTTI. If you don't know, leave USE_RTTI True.
+ USE_RTTI = False
+
# Manually specified defines.
DEFINES = []
diff --git a/build_android.py b/build_android.py
index c71fe27..f6ca328 100644
--- a/build_android.py
+++ b/build_android.py
@@ -31,6 +31,11 @@
# True.
USE_GLOG = False
+ # Whether to use RTTI or not. Note that this might not always work; to
+ # disable RTTI means that all your dependencies, most notably protobuf, have
+ # to be built without RTTI. If you don't know, leave USE_RTTI True.
+ USE_RTTI = False
+
# Manually specified defines.
DEFINES = []
@@ -46,9 +51,7 @@
CFLAGS = []
# Additional link flags you would like to add to the compilation.
- LINKFLAGS = [
- "-pie",
- ]
+ LINKFLAGS = []
###########################################################################
# (optional) CUDA. If you do not specify this, the GPU part of Caffe2 will
@@ -107,7 +110,7 @@
# build command, do it here.
ENVIRONMENTAL_VARS = {}
# Optimization flags: -O2 in default.
- OPTIMIZATION_FLAGS = ["-O2"]
+ OPTIMIZATION_FLAGS = ["-Os"]
# brew.py
diff --git a/caffe2/core/operator.cc b/caffe2/core/operator.cc
index 63206da..1acad9a 100644
--- a/caffe2/core/operator.cc
+++ b/caffe2/core/operator.cc
@@ -3,6 +3,7 @@
#include "caffe2/core/net.h"
#include "caffe2/core/operator.h"
+#include "caffe2/core/operator_gradient.h"
#include "caffe2/core/workspace.h"
#include "caffe2/proto/caffe2.pb.h"
@@ -146,4 +147,50 @@
DEFINE_REGISTRY(GradientRegistry, vector<OperatorDef>, const OperatorDef&);
+
+vector<OperatorDef>* CreateGradientDefsInternal(
+ const OperatorDef& def, GetGradientDefBaseVerbose* obj) {
+ CAFFE_DCHECK_NOTNULL(obj);
+ vector<OperatorDef>* grad_defs = obj->Create(def);
+ CAFFE_CHECK(grad_defs != nullptr);
+ // Copy device option if needed.
+ if (obj->CopyDeviceOption() && def.has_device_option()) {
+ for (OperatorDef& grad_def : *grad_defs) {
+ grad_def.mutable_device_option()->CopyFrom(def.device_option());
+ }
+ }
+ // Copy engine if needed.
+ if (obj->CopyEngine() && def.has_engine()) {
+ for (OperatorDef& grad_def : *grad_defs) {
+ grad_def.set_engine(def.engine());
+ }
+ }
+ // Copy arguments if needed.
+ if (obj->CopyArguments() && def.arg_size()) {
+ for (OperatorDef& grad_def : *grad_defs) {
+ grad_def.mutable_arg()->CopyFrom(def.arg());
+ }
+ }
+ for (const OperatorDef& grad_def : *grad_defs) {
+ CAFFE_VLOG(1) << "Gradient: " << grad_def.DebugString();
+ }
+ delete obj;
+ return grad_defs;
+}
+
+vector<OperatorDef>* ThrowTheTowelIfGradientIsCalled::Create(
+ const OperatorDef& def) {
+ CAFFE_LOG_FATAL << "You should not call the gradient of operator of type "
+ << def.type();
+ // Just to suppress compiler warnings
+ return new vector<OperatorDef>();
+}
+
+vector<OperatorDef>* GradientNotImplementedYet::Create(
+ const OperatorDef& def) {
+ CAFFE_LOG_FATAL << "Gradient for operator type "
+ << def.type() << " has not been implemented yet.";
+ return nullptr;
+}
+
} // namespace caffe2
diff --git a/caffe2/core/operator_gradient.h b/caffe2/core/operator_gradient.h
index cf0c90b..6c9e72d2 100644
--- a/caffe2/core/operator_gradient.h
+++ b/caffe2/core/operator_gradient.h
@@ -14,48 +14,19 @@
DECLARE_REGISTRY(GradientRegistry, vector<OperatorDef>, const OperatorDef&);
-template <class GetGradientDef>
-class GradientRegisterer {
- public:
- GradientRegisterer(const string& key) {
- GradientRegistry()->Register(
- key, GradientRegisterer<GetGradientDef>::Creator);
- }
- static vector<OperatorDef>* Creator(const OperatorDef& def) {
- CAFFE_VLOG(1) << "Creator: " << def.DebugString();
- vector<OperatorDef>* grad_defs = GetGradientDef::Create(def);
- CAFFE_CHECK(grad_defs != nullptr);
- // Copy device option if needed.
- if (GetGradientDef().CopyDeviceOption() && def.has_device_option()) {
- for (OperatorDef& grad_def : *grad_defs) {
- grad_def.mutable_device_option()->CopyFrom(def.device_option());
- }
- }
- // Copy engine if needed.
- if (GetGradientDef().CopyEngine() && def.has_engine()) {
- for (OperatorDef& grad_def : *grad_defs) {
- grad_def.set_engine(def.engine());
- }
- }
- // Copy arguments if needed.
- if (GetGradientDef().CopyArguments() && def.arg_size()) {
- for (OperatorDef& grad_def : *grad_defs) {
- grad_def.mutable_arg()->CopyFrom(def.arg());
- }
- }
- for (const OperatorDef& grad_def : *grad_defs) {
- CAFFE_VLOG(1) << "Gradient: " << grad_def.DebugString();
- }
- return grad_defs;
- }
-};
-
-template <bool copy_device_option, bool copy_engine, bool copy_args>
struct GetGradientDefBaseVerbose {
- constexpr bool CopyDeviceOption() const { return copy_device_option; }
- constexpr bool CopyEngine() const { return copy_engine; }
- constexpr bool CopyArguments() const { return copy_args; }
+ public:
+ GetGradientDefBaseVerbose(
+ const bool copy_device_option, const bool copy_engine,
+ const bool copy_args)
+ : copy_device_option_(copy_device_option), copy_engine_(copy_engine),
+ copy_args_(copy_args) {}
+ virtual ~GetGradientDefBaseVerbose() {}
+
+ bool CopyDeviceOption() const { return copy_device_option_; }
+ bool CopyEngine() const { return copy_engine_; }
+ bool CopyArguments() const { return copy_args_; }
inline static string I(const OperatorDef& def, const int i) {
return def.input(i);
}
@@ -68,32 +39,59 @@
inline static string GO(const OperatorDef& def, const int i) {
return GradientName(def.output(i));
}
+
+ virtual vector<OperatorDef>* Create(const OperatorDef& def) {
+ NOT_IMPLEMENTED;
+ return nullptr;
+ }
+
template <class... Args>
inline static vector<OperatorDef>* SingleGradientDef(Args ... args) {
return new vector<OperatorDef>{CreateOperatorDef(args...)};
}
+
+ bool copy_device_option_;
+ bool copy_engine_;
+ bool copy_args_;
};
-typedef struct GetGradientDefBaseVerbose<true, true, true> GetGradientDefBase;
+
+
+struct GetGradientDefBase : public GetGradientDefBaseVerbose {
+ public:
+ GetGradientDefBase() : GetGradientDefBaseVerbose(true, true, true) {}
+};
struct NoGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return new vector<OperatorDef>();
}
};
+// This is used when the operator definition is designed to not have a gradient.
+// Calling a gradient on this operator def will cause Caffe2 to throw the towel.
struct ThrowTheTowelIfGradientIsCalled : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
- CAFFE_LOG_FATAL << "You should not call the gradient of operator of type "
- << def.type();
- // Just to suppress compiler warnings
- return new vector<OperatorDef>();
- }
+ vector<OperatorDef>* Create(const OperatorDef& def) override;
};
+// This should only be used sparsely when the gradient does exist, but we have
+// not implemented it yet.
struct GradientNotImplementedYet : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
- CAFFE_LOG_FATAL << "Gradient for operator type "
- << def.type() << " has not been implemented yet.";
+ vector<OperatorDef>* Create(const OperatorDef& def) override;
+};
+
+vector<OperatorDef>* CreateGradientDefsInternal(
+ const OperatorDef& def, GetGradientDefBaseVerbose* obj);
+
+template <class GetGradientDef>
+class GradientRegisterer {
+ public:
+ GradientRegisterer(const string& key) {
+ GradientRegistry()->Register(
+ key, GradientRegisterer<GetGradientDef>::Creator);
+ }
+
+ static vector<OperatorDef>* Creator(const OperatorDef& def) {
+ return CreateGradientDefsInternal(def, new GetGradientDef());
}
};
@@ -111,13 +109,6 @@
GradientRegisterer<ThrowTheTowelIfGradientIsCalled> \
g_GradientRegisterer_##name(#name)
-// SHOULD_NOT_DO_GRADIENT means that the operator is not designed to have
-// gradient operators. If you attempt to call the gradient, a log fatal will
-// occur.
-#define SHOULD_NOT_DO_GRADIENT(name) \
- GradientRegisterer<ThrowTheTowelIfGradientIsCalled> \
- g_GradientRegisterer_##name(#name)
-
#define GRADIENT_NOT_IMPLEMENTED_YET(name) \
GradientRegisterer<GradientNotImplementedYet> \
g_GradientRegisterer_##name(#name)
diff --git a/caffe2/core/operator_test.cc b/caffe2/core/operator_test.cc
index be6baf9..8bd252e 100644
--- a/caffe2/core/operator_test.cc
+++ b/caffe2/core/operator_test.cc
@@ -303,7 +303,7 @@
}
struct GetFooGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return new vector<OperatorDef>{
CreateOperatorDef(
"FooGradient", "",
diff --git a/caffe2/core/typeid.h b/caffe2/core/typeid.h
index c64ecfe..ea9454f 100644
--- a/caffe2/core/typeid.h
+++ b/caffe2/core/typeid.h
@@ -2,7 +2,9 @@
#define CAFFE2_CORE_TYPEID_H_
#include <map>
+#ifdef __GXX_RTTI
#include <typeinfo>
+#endif
#include "caffe2/core/common.h"
@@ -79,7 +81,13 @@
* Returns the printable name of the type.
*/
template <typename T>
- static const char* Name() { return typeid(T).name(); }
+ static const char* Name() {
+#ifdef __GXX_RTTI
+ return typeid(T).name();
+#else // __GXX_RTTI
+ return "(RTTI disabled, cannot show name)";
+#endif
+ }
/**
* Returns a TypeMeta object that corresponds to the typename T.
*/
diff --git a/caffe2/operators/averagepool_op.cc b/caffe2/operators/averagepool_op.cc
index 2d94268..45860a7 100644
--- a/caffe2/operators/averagepool_op.cc
+++ b/caffe2/operators/averagepool_op.cc
@@ -193,7 +193,7 @@
AveragePoolGradientOp<float, CPUContext>);
struct GetAveragePoolGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"AveragePoolGradient", "",
vector<string>{I(def, 0), GO(def, 0)},
diff --git a/caffe2/operators/clip_op.cc b/caffe2/operators/clip_op.cc
index 3ee0105..33bef74 100644
--- a/caffe2/operators/clip_op.cc
+++ b/caffe2/operators/clip_op.cc
@@ -38,7 +38,7 @@
REGISTER_CPU_OPERATOR(ClipGradient, ClipGradientOp<float, CPUContext>);
struct GetClipGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"ClipGradient", "",
vector<string>{O(def, 0), GO(def, 0)},
diff --git a/caffe2/operators/conv_op.cc b/caffe2/operators/conv_op.cc
index 6cdadd2..2cfa874 100644
--- a/caffe2/operators/conv_op.cc
+++ b/caffe2/operators/conv_op.cc
@@ -7,7 +7,7 @@
REGISTER_CPU_OPERATOR(ConvGradient, ConvGradientOp<float, CPUContext>);
struct GetConvGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
CAFFE_CHECK_EQ(def.input_size(), 3);
return SingleGradientDef(
"ConvGradient", "",
diff --git a/caffe2/operators/conv_op_cudnn.cc b/caffe2/operators/conv_op_cudnn.cc
index 28bd798..da43830 100644
--- a/caffe2/operators/conv_op_cudnn.cc
+++ b/caffe2/operators/conv_op_cudnn.cc
@@ -380,7 +380,7 @@
REGISTER_CUDNN_OPERATOR(ConvFp16Gradient, CudnnConvGradientOp<float16>);
struct GetConvFp16Gradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
CAFFE_CHECK_EQ(def.input_size(), 3);
return SingleGradientDef(
"ConvFp16Gradient", "",
diff --git a/caffe2/operators/cross_entropy_op.cc b/caffe2/operators/cross_entropy_op.cc
index bec6d61..ec31767 100644
--- a/caffe2/operators/cross_entropy_op.cc
+++ b/caffe2/operators/cross_entropy_op.cc
@@ -58,7 +58,7 @@
LabelCrossEntropyGradientOp<float, CPUContext>);
struct GetLabelCrossEntropyGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"LabelCrossEntropyGradient", "",
vector<string>{I(def, 0), I(def, 1), GO(def, 0)},
diff --git a/caffe2/operators/depth_split_op.cc b/caffe2/operators/depth_split_op.cc
index e02c863..029551d 100644
--- a/caffe2/operators/depth_split_op.cc
+++ b/caffe2/operators/depth_split_op.cc
@@ -6,7 +6,7 @@
REGISTER_CPU_OPERATOR(DepthConcat, DepthConcatOp<CPUContext>);
struct GetDepthSplitGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
vector<string> grad_inputs;
for (const string& out : def.output()) {
grad_inputs.push_back(GradientName(out));
@@ -19,7 +19,7 @@
REGISTER_GRADIENT(DepthSplit, GetDepthSplitGradient);
struct GetDepthConcatGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
vector<string> grad_outputs;
for (const string& in : def.input()) {
grad_outputs.push_back(GradientName(in));
diff --git a/caffe2/operators/dropout_op.cc b/caffe2/operators/dropout_op.cc
index 54f252d..a447e98 100644
--- a/caffe2/operators/dropout_op.cc
+++ b/caffe2/operators/dropout_op.cc
@@ -61,7 +61,7 @@
REGISTER_CPU_OPERATOR(DropoutGrad, DropoutGradientOp<float, CPUContext>);
struct GetDropoutGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"DropoutGrad", "",
vector<string>{GO(def, 0), O(def, 1)},
diff --git a/caffe2/operators/fully_connected_op.cc b/caffe2/operators/fully_connected_op.cc
index 0300771..26c7438 100644
--- a/caffe2/operators/fully_connected_op.cc
+++ b/caffe2/operators/fully_connected_op.cc
@@ -7,7 +7,7 @@
REGISTER_CPU_OPERATOR(FCGradient, FullyConnectedGradientOp<float, CPUContext>);
struct GetFCGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
CAFFE_CHECK_EQ(def.input_size(), 3);
return SingleGradientDef(
"FCGradient", "",
diff --git a/caffe2/operators/half_float_ops.cu b/caffe2/operators/half_float_ops.cu
index a274084..4bdd43a 100644
--- a/caffe2/operators/half_float_ops.cu
+++ b/caffe2/operators/half_float_ops.cu
@@ -64,7 +64,7 @@
REGISTER_CUDA_OPERATOR(HalfToFloat, HalfToFloatCUDA);
struct GetFloatToHalfGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"HalfToFloat", "",
vector<string>{GO(def, 0)},
@@ -74,7 +74,7 @@
REGISTER_GRADIENT(FloatToHalf, GetFloatToHalfGradient);
struct GetHalfToFloatGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"FloatToHalf", "",
vector<string>{GO(def, 0)},
diff --git a/caffe2/operators/l2_distance_op.cc b/caffe2/operators/l2_distance_op.cc
index d3c0413..0f618ef 100644
--- a/caffe2/operators/l2_distance_op.cc
+++ b/caffe2/operators/l2_distance_op.cc
@@ -35,7 +35,7 @@
SquaredL2DistanceGradientOp<float, CPUContext>);
struct GetSquaredL2DistanceGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"SquaredL2DistanceGradient", "",
vector<string>{I(def, 0), I(def, 1), GO(def, 0)},
diff --git a/caffe2/operators/local_response_normalization_op.cc b/caffe2/operators/local_response_normalization_op.cc
index 59ac2fa..5da070a 100644
--- a/caffe2/operators/local_response_normalization_op.cc
+++ b/caffe2/operators/local_response_normalization_op.cc
@@ -232,7 +232,7 @@
REGISTER_CPU_OPERATOR(LRNGradient, LRNGradientOp<float, CPUContext>);
struct GetLRNGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"LRNGradient", "",
vector<string>{I(def, 0), O(def, 0), O(def, 1), GO(def, 0)},
diff --git a/caffe2/operators/loss_op.cc b/caffe2/operators/loss_op.cc
index 4d5e535..15c12e4 100644
--- a/caffe2/operators/loss_op.cc
+++ b/caffe2/operators/loss_op.cc
@@ -12,7 +12,7 @@
struct GetAveragedLossGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"AveragedLossGradient", "",
vector<string>{I(def, 0)},
@@ -22,7 +22,7 @@
REGISTER_GRADIENT(AveragedLoss, GetAveragedLossGradient);
struct GetWeightedSumLossGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"WeightedSumLossGradient", "",
vector<string>{I(def, 1)},
diff --git a/caffe2/operators/maxpool_op.cc b/caffe2/operators/maxpool_op.cc
index d3c744a..e78aa156 100644
--- a/caffe2/operators/maxpool_op.cc
+++ b/caffe2/operators/maxpool_op.cc
@@ -141,7 +141,7 @@
REGISTER_CPU_OPERATOR(MaxPoolGradient, MaxPoolGradientOp<float, CPUContext>);
struct GetMaxPoolGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"MaxPoolGradient", "",
vector<string>{I(def, 0), GO(def, 0), O(def, 1)},
diff --git a/caffe2/operators/order_switch_ops.cc b/caffe2/operators/order_switch_ops.cc
index 3fdf7b9..5b584fc 100644
--- a/caffe2/operators/order_switch_ops.cc
+++ b/caffe2/operators/order_switch_ops.cc
@@ -50,7 +50,7 @@
REGISTER_CPU_OPERATOR(NCHW2NHWC, NCHW2NHWCOp<float, CPUContext>);
struct GetNHWC2NCHWGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"NCHW2NHWC", "",
vector<string>{GO(def, 0)},
@@ -60,7 +60,7 @@
REGISTER_GRADIENT(NHWC2NCHW, GetNHWC2NCHWGradient);
struct GetNCHW2NHWCGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"NHWC2NCHW", "",
vector<string>{GO(def, 0)},
diff --git a/caffe2/operators/relu_op.cc b/caffe2/operators/relu_op.cc
index 857819c..42e577a1 100644
--- a/caffe2/operators/relu_op.cc
+++ b/caffe2/operators/relu_op.cc
@@ -38,7 +38,7 @@
REGISTER_CPU_OPERATOR(ReluGradient, ReluGradientOp<float, CPUContext>);
struct GetReluGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"ReluGradient", "",
vector<string>{O(def, 0), GO(def, 0)},
diff --git a/caffe2/operators/sigmoid_op.cc b/caffe2/operators/sigmoid_op.cc
index 323cc34..dfb831b 100644
--- a/caffe2/operators/sigmoid_op.cc
+++ b/caffe2/operators/sigmoid_op.cc
@@ -42,7 +42,7 @@
SigmoidGradientCPUFunctor<float> >);
struct GetSigmoidGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"SigmoidGradient", "",
vector<string>{O(def, 0), GO(def, 0)},
diff --git a/caffe2/operators/softmax_op.cc b/caffe2/operators/softmax_op.cc
index 1735f31..f5356a1 100644
--- a/caffe2/operators/softmax_op.cc
+++ b/caffe2/operators/softmax_op.cc
@@ -90,7 +90,7 @@
REGISTER_CPU_OPERATOR(SoftmaxGradient, SoftmaxGradientOp<float, CPUContext>);
struct GetSoftmaxGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"SoftmaxGradient", "",
vector<string>{O(def, 0), GO(def, 0)},
diff --git a/caffe2/operators/spatial_batch_norm_op.cc b/caffe2/operators/spatial_batch_norm_op.cc
index 3bb1806..ef06a1a 100644
--- a/caffe2/operators/spatial_batch_norm_op.cc
+++ b/caffe2/operators/spatial_batch_norm_op.cc
@@ -8,7 +8,7 @@
// is a bit more complex than usual gradient operators.
namespace {
struct GetSpatialBNGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
// Check if we are in training or testing mode.
bool is_test = false;
if (HasArgument(def, "is_test")) {
diff --git a/caffe2/operators/tanh_op.cc b/caffe2/operators/tanh_op.cc
index bfbf50a..91bf13e 100644
--- a/caffe2/operators/tanh_op.cc
+++ b/caffe2/operators/tanh_op.cc
@@ -42,7 +42,7 @@
TanhGradientCPUFunctor<float> >);
struct GetTanhGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"TanhGradient", "",
std::vector<string>{O(def, 0), GO(def, 0)},
diff --git a/caffe2/operators/utility_ops.cc b/caffe2/operators/utility_ops.cc
index ab6ec97..47cb7c7 100644
--- a/caffe2/operators/utility_ops.cc
+++ b/caffe2/operators/utility_ops.cc
@@ -19,7 +19,7 @@
SHOULD_NOT_DO_GRADIENT(PrintInt);
struct GetFlattenGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"ReshapeLike", "",
vector<string>{GO(def, 0), I(def, 0)},
@@ -29,7 +29,7 @@
REGISTER_GRADIENT(Flatten, GetFlattenGradient);
struct GetAliasGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
return SingleGradientDef(
"Alias", "",
vector<string>{GO(def, 0)},
@@ -41,7 +41,7 @@
SHOULD_NOT_DO_GRADIENT(ReshapeLike);
struct GetSplitGradient : public GetGradientDefBase {
- static vector<OperatorDef>* Create(const OperatorDef& def) {
+ vector<OperatorDef>* Create(const OperatorDef& def) override {
vector<string> grad_input;
for (const string out : def.output()) {
grad_input.push_back(GradientName(out));
diff --git a/pycaffe2/convnet_benchmarks.py b/pycaffe2/convnet_benchmarks.py
index 8bec892..48b7a70 100644
--- a/pycaffe2/convnet_benchmarks.py
+++ b/pycaffe2/convnet_benchmarks.py
@@ -259,7 +259,7 @@
if (not args.batch_size or not args.model or not args.order or not args.cudnn_ws):
parser.print_help()
- workspace.GlobalInit(['caffe2', '--caffe2_log_level=0'])
+ workspace.GlobalInit(['caffe2', '--caffe2_log_level=-1'])
model_map = {'AlexNet': AlexNet, 'OverFeat': OverFeat, 'VGGA': VGGA, 'Inception': Inception}
Benchmark(model_map[args.model], args.order, args.batch_size, args.cudnn_ws,
args.forward_only, args.iterations)
