COMPMID-417 - Fixed auto-config in NEConvolutionLayer and in CLConvolutionLayer
Change-Id: Ibfd772200348b326738bb3b8357f0abbb7a583d7
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/82943
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
diff --git a/arm_compute/core/Helpers.inl b/arm_compute/core/Helpers.inl
index d342a60..78e0c70 100644
--- a/arm_compute/core/Helpers.inl
+++ b/arm_compute/core/Helpers.inl
@@ -251,8 +251,8 @@
if(info.tensor_shape().total_size() == 0)
{
info.set_data_type(data_type);
- info.set_tensor_shape(shape);
info.set_num_channels(num_channels);
+ info.set_tensor_shape(shape);
info.set_fixed_point_position(fixed_point_position);
return true;
}
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 3092976..1d04f35 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -607,13 +607,13 @@
float _kappa;
};
-/** Convolution Layer Weights Information class */
+/** Convolution Layer Weights Information class. This class stores the necessary information to compute convolution layer when the weights are already reshaped */
class WeightsInfo
{
public:
/** Default constructor */
WeightsInfo()
- : _are_reshaped(false), _kernel_width(0), _kernel_height(0)
+ : _are_reshaped(false), _kernel_width(0), _kernel_height(0), _num_kernels(0)
{
}
/** Constructor
@@ -621,9 +621,10 @@
* @param[in] are_reshaped True if the weights have been reshaped
* @param[in] kernel_width Kernel width.
* @param[in] kernel_height Kernel height.
+ * @param[in] num_kernels Number of convolution kernels.
*/
- WeightsInfo(bool are_reshaped, unsigned int kernel_width, unsigned int kernel_height)
- : _are_reshaped(are_reshaped), _kernel_width(kernel_width), _kernel_height(kernel_height)
+ WeightsInfo(bool are_reshaped, unsigned int kernel_width, unsigned int kernel_height, unsigned int num_kernels)
+ : _are_reshaped(are_reshaped), _kernel_width(kernel_width), _kernel_height(kernel_height), _num_kernels(num_kernels)
{
}
/** Flag which specifies if the weights tensor has been reshaped.
@@ -634,6 +635,14 @@
{
return _are_reshaped;
};
+ /** Return the number of convolution kernels
+ *
+ * @return The number of convolution kernels
+ */
+ unsigned int num_kernels() const
+ {
+ return _num_kernels;
+ };
/** Return the width and height of the kernel
*
* @return The width and height of the kernel
@@ -647,6 +656,7 @@
const bool _are_reshaped;
const unsigned int _kernel_width;
const unsigned int _kernel_height;
+ const unsigned int _num_kernels;
};
/** IO formatting information class*/
diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index 4362cfd..ff94e9d 100644
--- a/src/runtime/CL/functions/CLConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp
@@ -97,8 +97,8 @@
void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, output);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
@@ -132,7 +132,6 @@
const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
conv_info);
- ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
// Check if its a "fully connected" convolution
_is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
@@ -143,7 +142,7 @@
// Reshape weights if needed
if(_are_weights_reshaped)
{
- mat_weights_cols = output->info()->dimension(2);
+ mat_weights_cols = weights_info.num_kernels();
const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4;
mat_weights_rows = (_has_bias ? 1 + quarter_reshaped_cols : quarter_reshaped_cols);
}
@@ -206,6 +205,8 @@
}
_output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
+ ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
+
// Allocate intermediate tensor
if(!_are_weights_reshaped)
{
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index f6481f1..1c87f60 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -96,8 +96,8 @@
void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, output);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
@@ -131,7 +131,6 @@
const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
conv_info);
- ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
// Check if its a "fully connected" convolution
_is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
@@ -142,7 +141,7 @@
// Reshape weights if needed
if(_are_weights_reshaped)
{
- mat_weights_cols = output->info()->dimension(2);
+ mat_weights_cols = weights_info.num_kernels();
const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4;
mat_weights_rows = (_has_bias ? 1 + quarter_reshaped_cols : quarter_reshaped_cols);
}
@@ -205,6 +204,8 @@
}
_output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
+ ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
+
// Allocate intermediate tensor
if(!_are_weights_reshaped)
{
diff --git a/tests/model_objects/AlexNet.h b/tests/model_objects/AlexNet.h
index dad6832..c9fd448 100644
--- a/tests/model_objects/AlexNet.h
+++ b/tests/model_objects/AlexNet.h
@@ -251,26 +251,26 @@
// Configure Layers
{
// Layer 1
- conv1->configure(&input, w[0].get(), b[0].get(), &conv1_out, PadStrideInfo(4, 4, 0, 0), WeightsInfo(_reshaped_weights, 11U, 11U));
+ conv1->configure(&input, w[0].get(), b[0].get(), &conv1_out, PadStrideInfo(4, 4, 0, 0), WeightsInfo(_reshaped_weights, 11U, 11U, 96U));
act1->configure(&conv1_out, &act1_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
norm1->configure(&act1_out, &norm1_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f));
pool1->configure(&norm1_out, &pool1_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)));
// Layer 2
- conv21->configure(pool11_out.get(), w21.get(), b21.get(), conv21_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U));
- conv22->configure(pool12_out.get(), w22.get(), b22.get(), conv22_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U));
+ conv21->configure(pool11_out.get(), w21.get(), b21.get(), conv21_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U, 128U));
+ conv22->configure(pool12_out.get(), w22.get(), b22.get(), conv22_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U, 128U));
act2->configure(&conv2_out, &act2_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
norm2->configure(&act2_out, &norm2_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f));
pool2->configure(&norm2_out, &pool2_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)));
// Layer 3
- conv3->configure(&pool2_out, w[2].get(), b[2].get(), &conv3_out, PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
+ conv3->configure(&pool2_out, w[2].get(), b[2].get(), &conv3_out, PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 384U));
act3->configure(&conv3_out, &act3_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
// Layer 4
- conv41->configure(act31_out.get(), w41.get(), b41.get(), conv41_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
- conv42->configure(act32_out.get(), w42.get(), b42.get(), conv42_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
+ conv41->configure(act31_out.get(), w41.get(), b41.get(), conv41_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 192U));
+ conv42->configure(act32_out.get(), w42.get(), b42.get(), conv42_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 192U));
act4->configure(&conv4_out, &act4_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
// Layer 5
- conv51->configure(act41_out.get(), w51.get(), b51.get(), conv51_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
- conv52->configure(act42_out.get(), w52.get(), b52.get(), conv52_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
+ conv51->configure(act41_out.get(), w51.get(), b51.get(), conv51_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 128U));
+ conv52->configure(act42_out.get(), w52.get(), b52.get(), conv52_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 128U));
act5->configure(&conv5_out, &act5_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
pool5->configure(&act5_out, &pool5_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)));
// Layer 6
diff --git a/tests/networks_new/AlexNetNetwork.h b/tests/networks_new/AlexNetNetwork.h
index 8c801f7..28bcfa1 100644
--- a/tests/networks_new/AlexNetNetwork.h
+++ b/tests/networks_new/AlexNetNetwork.h
@@ -212,27 +212,27 @@
// Configure Layers
// Layer 1
TensorType *b0 = _reshaped_weights ? nullptr : &b[0];
- conv1.configure(&input, &w[0], b0, &conv1_out, PadStrideInfo(4, 4, 0, 0), WeightsInfo(_reshaped_weights, 11U, 11U));
+ conv1.configure(&input, &w[0], b0, &conv1_out, PadStrideInfo(4, 4, 0, 0), WeightsInfo(_reshaped_weights, 11U, 11U, 96U));
act1.configure(&conv1_out, &act1_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
norm1.configure(&act1_out, &norm1_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f));
pool1.configure(&norm1_out, &pool1_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)));
// Layer 2
- conv21.configure(pool11_out.get(), w21.get(), b21.get(), conv21_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U));
- conv22.configure(pool12_out.get(), w22.get(), b22.get(), conv22_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U));
+ conv21.configure(pool11_out.get(), w21.get(), b21.get(), conv21_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U, 128U));
+ conv22.configure(pool12_out.get(), w22.get(), b22.get(), conv22_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U, 128U));
act2.configure(&conv2_out, &act2_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
norm2.configure(&act2_out, &norm2_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f));
pool2.configure(&norm2_out, &pool2_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)));
// Layer 3
TensorType *b2 = (_reshaped_weights && !_is_direct_conv) ? nullptr : &b[2];
- conv3.configure(&pool2_out, &w[2], b2, &conv3_out, PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
+ conv3.configure(&pool2_out, &w[2], b2, &conv3_out, PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 384U));
act3.configure(&conv3_out, &act3_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
// Layer 4
- conv41.configure(act31_out.get(), w41.get(), b41.get(), conv41_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
- conv42.configure(act32_out.get(), w42.get(), b42.get(), conv42_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
+ conv41.configure(act31_out.get(), w41.get(), b41.get(), conv41_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 192U));
+ conv42.configure(act32_out.get(), w42.get(), b42.get(), conv42_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 192U));
act4.configure(&conv4_out, &act4_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
// Layer 5
- conv51.configure(act41_out.get(), w51.get(), b51.get(), conv51_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
- conv52.configure(act42_out.get(), w52.get(), b52.get(), conv52_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U));
+ conv51.configure(act41_out.get(), w51.get(), b51.get(), conv51_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 128U));
+ conv52.configure(act42_out.get(), w52.get(), b52.get(), conv52_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 128U));
act5.configure(&conv5_out, &act5_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
pool5.configure(&act5_out, &pool5_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)));
// Layer 6