Skip to content

Make eltwise product gradient stabler (fixes issues with WITHIN_CHANNEL LRN in cifar_full example) #981

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Aug 26, 2014
Merged

Make eltwise product gradient stabler (fixes issues with WITHIN_CHANNEL LRN in cifar_full example) #981

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
910db97
Add "stable_prod_grad" option (on by default) to ELTWISE layer to
jeffdonahue Aug 25, 2014
0063028
remove warning about LRN layers in CPU mode
jeffdonahue Aug 25, 2014
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 0 additions & 5 deletions examples/cifar10/cifar10_full_solver.prototxt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,9 +23,4 @@ max_iter: 60000
snapshot: 10000
snapshot_prefix: "cifar10_full"
# solver mode: CPU or GPU
# Note: there seems to be a bug with CPU computation in the pooling layers,
# and changing to solver_mode: CPU may result in NaNs on this example.
# If you want to train a variant of this architecture on the
# CPU, try changing the pooling regions from WITHIN_CHANNEL to ACROSS_CHANNELS
# in both cifar_full_train.prototxt and cifar_full_test.prototxt.
solver_mode: GPU
2 changes: 2 additions & 0 deletions include/caffe/common_layers.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,8 @@ class EltwiseLayer : public Layer<Dtype> {

EltwiseParameter_EltwiseOp op_;
vector<Dtype> coeffs_;

bool stable_prod_grad_;
};

/* FlattenLayer
Expand Down
17 changes: 16 additions & 1 deletion src/caffe/layers/eltwise_layer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@ void EltwiseLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
coeffs_[i] = this->layer_param().eltwise_param().coeff(i);
}
}
stable_prod_grad_ = this->layer_param_.eltwise_param().stable_prod_grad();
}

template <typename Dtype>
Expand Down Expand Up @@ -73,7 +74,21 @@ void EltwiseLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
Dtype* bottom_diff = (*bottom)[i]->mutable_cpu_diff();
switch (op_) {
case EltwiseParameter_EltwiseOp_PROD:
caffe_div(count, top_data, bottom_data, bottom_diff);
if (stable_prod_grad_) {
bool initialized = false;
for (int j = 0; j < bottom->size(); ++j) {
if (i == j) { continue; }
if (!initialized) {
caffe_copy(count, (*bottom)[j]->cpu_data(), bottom_diff);
initialized = true;
} else {
caffe_mul(count, (*bottom)[j]->cpu_data(), bottom_diff,
bottom_diff);
}
}
} else {
caffe_div(count, top_data, bottom_data, bottom_diff);
}
caffe_mul(count, bottom_diff, top_diff, bottom_diff);
break;
case EltwiseParameter_EltwiseOp_SUM:
Expand Down
16 changes: 15 additions & 1 deletion src/caffe/layers/eltwise_layer.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,21 @@ void EltwiseLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
Dtype* bottom_diff = (*bottom)[i]->mutable_gpu_diff();
switch (op_) {
case EltwiseParameter_EltwiseOp_PROD:
caffe_gpu_div(count, top_data, bottom_data, bottom_diff);
if (stable_prod_grad_) {
bool initialized = false;
for (int j = 0; j < bottom->size(); ++j) {
if (i == j) { continue; }
if (!initialized) {
caffe_copy(count, (*bottom)[j]->gpu_data(), bottom_diff);
initialized = true;
} else {
caffe_gpu_mul(count, (*bottom)[j]->gpu_data(), bottom_diff,
bottom_diff);
}
}
} else {
caffe_gpu_div(count, top_data, bottom_data, bottom_diff);
}
caffe_gpu_mul(count, bottom_diff, top_diff, bottom_diff);
break;
case EltwiseParameter_EltwiseOp_SUM:
Expand Down
4 changes: 4 additions & 0 deletions src/caffe/proto/caffe.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -420,6 +420,10 @@ message EltwiseParameter {
}
optional EltwiseOp operation = 1 [default = SUM]; // element-wise operation
repeated float coeff = 2; // blob-wise coefficient for SUM operation

// Whether to use an asymptotically slower (for >2 inputs) but stabler method
// of computing the gradient for the PROD operation. (No effect for SUM op.)
optional bool stable_prod_grad = 3 [default = true];
}

// Message that stores parameters used by ThresholdLayer
Expand Down
15 changes: 14 additions & 1 deletion src/caffe/test/test_eltwise_layer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,11 +124,24 @@ TYPED_TEST(EltwiseLayerTest, TestSumCoeff) {
}
}

TYPED_TEST(EltwiseLayerTest, TestProdGradient) {
TYPED_TEST(EltwiseLayerTest, TestStableProdGradient) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
EltwiseParameter* eltwise_param = layer_param.mutable_eltwise_param();
eltwise_param->set_operation(EltwiseParameter_EltwiseOp_PROD);
eltwise_param->set_stable_prod_grad(true);
EltwiseLayer<Dtype> layer(layer_param);
GradientChecker<Dtype> checker(1e-2, 1e-3);
checker.CheckGradientEltwise(&layer, &(this->blob_bottom_vec_),
&(this->blob_top_vec_));
}

TYPED_TEST(EltwiseLayerTest, TestUnstableProdGradient) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
EltwiseParameter* eltwise_param = layer_param.mutable_eltwise_param();
eltwise_param->set_operation(EltwiseParameter_EltwiseOp_PROD);
eltwise_param->set_stable_prod_grad(false);
EltwiseLayer<Dtype> layer(layer_param);
GradientChecker<Dtype> checker(1e-2, 1e-3);
checker.CheckGradientEltwise(&layer, &(this->blob_bottom_vec_),
Expand Down