Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add warnings for large labels when using GPU histogram algorithms (#2834)

Browse Source
This commit is contained in:
Rory Mitchell 2017-10-26 17:31:10 +13:00 committed by GitHub
parent 13e7a2cff0
commit d9d5293cdb
5 changed files with 51 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
doc/gpu/index.md
View File

@ -12,13 +12,13 @@ Specify the 'tree_method' parameter as one of the following algorithms.
### Algorithms
```eval_rst
+--------------+-----------------------------------------------------------------------------------------------------------------------------------------------+
+--------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| tree_method | Description |
+==============+===============================================================================================================================================+
+==============+=================================================================================================================================================================================================================+
| gpu_exact | The standard XGBoost tree construction algorithm. Performs exact search for splits. Slower and uses considerably more memory than 'gpu_hist' |
+--------------+-----------------------------------------------------------------------------------------------------------------------------------------------+
| gpu_hist | Equivalent to the XGBoost fast histogram algorithm. Faster and uses considerably less memory. Splits may be less accurate. |
+--------------+-----------------------------------------------------------------------------------------------------------------------------------------------+
+--------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| gpu_hist | Equivalent to the XGBoost fast histogram algorithm. Much faster and uses considerably less memory. NOTE: Cannot be used with labels larger in magnitude than 2^16 due to it's histogram aggregation algorithm. |
+--------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
```
### Supported parameters

12
include/xgboost/base.h
View File

@ -171,19 +171,19 @@ class bst_gpair_internal {
template<>
inline XGBOOST_DEVICE float bst_gpair_internal<int64_t>::GetGrad() const {
return grad_ * 1e-5f;
return grad_ * 1e-4f;
}
template<>
inline XGBOOST_DEVICE float bst_gpair_internal<int64_t>::GetHess() const {
return hess_ * 1e-5f;
return hess_ * 1e-4f;
}
template<>
inline XGBOOST_DEVICE void bst_gpair_internal<int64_t>::SetGrad(float g) {
grad_ = static_cast<int64_t>(std::round(g * 1e5));
grad_ = static_cast<int64_t>(std::round(g * 1e4));
}
template<>
inline XGBOOST_DEVICE void bst_gpair_internal<int64_t>::SetHess(float h) {
hess_ = static_cast<int64_t>(std::round(h * 1e5));
hess_ = static_cast<int64_t>(std::round(h * 1e4));
}
} // namespace detail
@ -194,10 +194,10 @@ typedef detail::bst_gpair_internal<float> bst_gpair;
/*! \brief High precision gradient statistics pair */
typedef detail::bst_gpair_internal<double> bst_gpair_precise;
/*! \brief High precision gradient statistics pair with integer backed
/*! \brief High precision gradient statistics pair with integer backed
* storage. Operators are associative where floating point versions are not
* associative. */
typedef detail::bst_gpair_internal<int64_t> bst_gpair_integer;
typedef detail::bst_gpair_internal<int64_t> bst_gpair_integer;
/*! \brief small eps gap for minimum split decision. */
const bst_float rt_eps = 1e-6f;

33
src/tree/updater_gpu_common.cuh
View File

@ -15,6 +15,27 @@
namespace xgboost {
namespace tree {
/**
* \fn void CheckGradientMax(const dh::dvec<bst_gpair>& gpair)
*
* \brief Check maximum gradient value is below 2^16. This is to prevent
* overflow when using integer gradient summation.
*/
inline void CheckGradientMax(const dh::dvec<bst_gpair>& gpair) {
auto dptr = thrust::device_ptr<const float>(
reinterpret_cast<const float*>(gpair.data()));
float abs_max = thrust::reduce(dptr, dptr + (gpair.size() * 2), 0.f,
[=] __device__(float a, float b) {
a = abs(a);
b = abs(b);
return max(a, b);
});
CHECK_LT(abs_max, std::pow(2.0f, 16.0f))
<< "Labels are too large for this algorithm. Rescale to less than 2^16.";
}
struct GPUTrainingParam {
// minimum amount of hessian(weight) allowed in a child
float min_child_weight;
@ -64,7 +85,7 @@ struct DeviceSplitCandidate {
: loss_chg(-FLT_MAX), dir(LeftDir), fvalue(0), findex(-1) {}
template <typename param_t>
__host__ __device__ void Update(const DeviceSplitCandidate &other,
__host__ __device__ void Update(const DeviceSplitCandidate& other,
const param_t& param) {
if (other.loss_chg > loss_chg &&
other.left_sum.GetHess() >= param.min_child_weight &&
@ -170,8 +191,10 @@ struct SumCallbackOp {
};
template <typename gpair_t>
__device__ inline float device_calc_loss_chg(
const GPUTrainingParam& param, const gpair_t& left, const gpair_t& parent_sum, const float& parent_gain) {
__device__ inline float device_calc_loss_chg(const GPUTrainingParam& param,
const gpair_t& left,
const gpair_t& parent_sum,
const float& parent_gain) {
gpair_t right = parent_sum - left;
float left_gain = CalcGain(param, left.GetGrad(), left.GetHess());
float right_gain = CalcGain(param, right.GetGrad(), right.GetHess());
@ -187,8 +210,8 @@ __device__ float inline loss_chg_missing(const gpair_t& scan,
bool& missing_left_out) { // NOLINT
float missing_left_loss =
device_calc_loss_chg(param, scan + missing, parent_sum, parent_gain);
float missing_right_loss = device_calc_loss_chg(
param, scan, parent_sum, parent_gain);
float missing_right_loss =
device_calc_loss_chg(param, scan, parent_sum, parent_gain);
if (missing_left_loss >= missing_right_loss) {
missing_left_out = true;

3
src/tree/updater_gpu_hist.cu
View File

@ -537,6 +537,9 @@ class GPUHistMaker : public TreeUpdater {
device_gpair[d_idx].copy(gpair.begin() + device_row_segments[d_idx],
gpair.begin() + device_row_segments[d_idx + 1]);
// Check gradients are within acceptable size range
CheckGradientMax(device_gpair[d_idx]);
subsample_gpair(&device_gpair[d_idx], param.subsample,
device_row_segments[d_idx]);

2
src/tree/updater_gpu_hist_experimental.cu
View File

@ -334,6 +334,8 @@ struct DeviceShard {
ridx_segments.front() = std::make_pair(0, ridx.size());
this->gpair.copy(host_gpair.begin() + row_start_idx,
host_gpair.begin() + row_end_idx);
// Check gradients are within acceptable size range
CheckGradientMax(gpair);
hist.Reset();
}