Skip to content
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

Jump to bottom

refactor: [sparse_weights] get for predict #4651

Merged
merged 25 commits into from
Nov 17, 2023
Merged

refactor: [sparse_weights] get for predict #4651

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
25 commits
Select commit Hold shift + click to select a range
261d47e
refactor: [sparse_weights] use std::map and get for predict
bassmang Oct 13, 2023
7c42c29
small fixes
bassmang Oct 13, 2023
ab4aab1
clang
bassmang Oct 13, 2023
1828768
add feature on learn path
bassmang Oct 13, 2023
1e577d4
update -q:: for sparse tests
bassmang Oct 13, 2023
fdbe482
add get to array_parameters
bassmang Oct 13, 2023
d55390a
brackets for interactions update
bassmang Oct 13, 2023
58cb44d
fix interaction tests
bassmang Oct 13, 2023
0205ef3
comments
bassmang Oct 16, 2023
e83e7d9
revert std::map
bassmang Oct 16, 2023
007606c
Merge remote-tracking branch 'upstream/master' into gd_get
bassmang Oct 16, 2023
24a0b63
clang
bassmang Oct 16, 2023
384fb17
Merge branch 'master' into gd_get
bassmang Oct 26, 2023
22f2435
cpp11 standard
bassmang Oct 30, 2023
b0e7ec7
downgrade gcc
bassmang Oct 30, 2023
bed6b57
cpp11
bassmang Oct 30, 2023
da82d06
just remove latest for ubuntu
bassmang Oct 30, 2023
bdf30ac
add include
bassmang Oct 30, 2023
3663357
revert settings
bassmang Oct 30, 2023
2636c7d
fix sparse random and test
bassmang Nov 1, 2023
cb46dbc
lint
bassmang Nov 1, 2023
7e244fa
revert vendored change
bassmang Nov 1, 2023
6040976
Merge remote-tracking branch 'upstream/master' into gd_get
bassmang Nov 1, 2023
fe0e22b
tests
bassmang Nov 2, 2023
aff1526
Merge branch 'master' into gd_get
bassmang Nov 17, 2023
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
4 changes: 2 additions & 2 deletions test/core.vwtest.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6021,7 +6021,7 @@
{
"id": 465,
"desc": "cb_explore_adf with epsilon-greedy exploration using --sparse_weights and saving model",
"vw_command": "--cb_explore_adf --epsilon 0.1 -d train-sets/cb_test.ldf --noconstant --sparse_weights -f standard_sparse_model.vw",
"vw_command": "--cb_explore_adf --epsilon 0.1 -d train-sets/cb_test.ldf --noconstant --sparse_weights -f standard_sparse_model.vw -q::",
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

there's a total of 3 tests using --sparse_weights, might be the opportunity to beef it up - maybe unit tests?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

apart from that, would be cool to add some sparse benchmarks to master and see if they are affected by this change

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

benchmarks added, tests can be added in a separate PR

"diff_files": {
"stderr": "train-sets/ref/sparse_save_check.stderr",
"stdout": "train-sets/ref/sparse_save_check.stdout"
Expand All @@ -6033,7 +6033,7 @@
{
"id": 466,
"desc": "cb_explore_adf with epsilon-greedy exploration using --sparse_weights and loading model",
"vw_command": "--cb_explore_adf --epsilon 0.1 -d train-sets/cb_test.ldf --noconstant --sparse_weights -i standard_sparse_model.vw",
"vw_command": "--cb_explore_adf --epsilon 0.1 -d train-sets/cb_test.ldf --noconstant --sparse_weights -i standard_sparse_model.vw -q::",
"diff_files": {
"stderr": "train-sets/ref/sparse_load_check.stderr",
"stdout": "train-sets/ref/sparse_load_check.stdout"
Expand Down
9 changes: 5 additions & 4 deletions test/train-sets/ref/sparse_load_check.stderr
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
creating quadratic features for pairs: ::
using no cache
Reading datafile = train-sets/cb_test.ldf
num sources = 1
Expand All @@ -6,17 +7,17 @@ learning rate = 0.5
initial_t = 3
power_t = 0.5
cb_type = mtr
Enabled learners: gd, scorer-identity, csoaa_ldf-rank, cb_adf, cb_explore_adf_greedy, shared_feature_merger
Enabled learners: gd, generate_interactions, scorer-identity, csoaa_ldf-rank, cb_adf, cb_explore_adf_greedy, shared_feature_merger
Input label = CB
Output pred = ACTION_PROBS
average since example example current current current
loss last counter weight label predict features
0.066667 0.066667 1 1.0 0:1:0.5 1:0.48 15
0.033333 0.000000 2 2.0 1:0:0.5 1:0.95 6
0.066667 0.066667 1 1.0 0:1:0.5 1:0.48 60
0.033333 0.000000 2 2.0 1:0:0.5 1:0.95 18

finished run
number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.033333
total feature number = 27
total feature number = 96
3 changes: 3 additions & 0 deletions test/train-sets/ref/sparse_load_check.stdout
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,3 @@
[warning] model file has set of {-q, --cubic, --interactions} settings stored, but they'll be OVERRIDDEN by set of {-q, --cubic, --interactions} settings from command line.
[warning] Any duplicate namespace interactions will be removed
You can use --leave_duplicate_interactions to disable this behaviour.
9 changes: 5 additions & 4 deletions test/train-sets/ref/sparse_save_check.stderr
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
creating quadratic features for pairs: ::
final_regressor = standard_sparse_model.vw
using no cache
Reading datafile = train-sets/cb_test.ldf
Expand All @@ -7,17 +8,17 @@ learning rate = 0.5
initial_t = 0
power_t = 0.5
cb_type = mtr
Enabled learners: gd, scorer-identity, csoaa_ldf-rank, cb_adf, cb_explore_adf_greedy, shared_feature_merger
Enabled learners: gd, generate_interactions, scorer-identity, csoaa_ldf-rank, cb_adf, cb_explore_adf_greedy, shared_feature_merger
Input label = CB
Output pred = ACTION_PROBS
average since example example current current current
loss last counter weight label predict features
0.666667 0.666667 1 1.0 0:1:0.5 0:0.33 15
0.333333 0.000000 2 2.0 1:0:0.5 1:0.95 6
0.666667 0.666667 1 1.0 0:1:0.5 0:0.33 60
0.333333 0.000000 2 2.0 1:0:0.5 1:0.95 18

finished run
number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
total feature number = 27
total feature number = 96
2 changes: 2 additions & 0 deletions test/train-sets/ref/sparse_save_check.stdout
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
[warning] Any duplicate namespace interactions will be removed
You can use --leave_duplicate_interactions to disable this behaviour.
6 changes: 6 additions & 0 deletions vowpalwabbit/core/include/vw/core/array_parameters.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,12 @@ class parameters
else { return dense_weights[i]; }
}

inline VW::weight& get(size_t i)
bassmang marked this conversation as resolved.
Show resolved Hide resolved
{
if (sparse) { return sparse_weights.get(i); }
else { return dense_weights.get(i); }
}

template <typename Lambda>
void set_default(Lambda&& default_func)
{
Expand Down
3 changes: 3 additions & 0 deletions vowpalwabbit/core/include/vw/core/array_parameters_dense.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,6 +122,9 @@ class dense_parameters
inline const VW::weight& operator[](size_t i) const { return _begin.get()[i & _weight_mask]; }
inline VW::weight& operator[](size_t i) { return _begin.get()[i & _weight_mask]; }

inline const VW::weight& get(size_t i) const { return _begin.get()[i & _weight_mask]; }
inline VW::weight& get(size_t i) { return _begin.get()[i & _weight_mask]; }

VW_ATTR(nodiscard) static dense_parameters shallow_copy(const dense_parameters& input);
VW_ATTR(nodiscard) static dense_parameters deep_copy(const dense_parameters& input);

Expand Down
9 changes: 6 additions & 3 deletions vowpalwabbit/core/include/vw/core/array_parameters_sparse.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,8 +10,8 @@

#include <cstddef>
#include <functional>
#include <map>
#include <memory>
#include <unordered_map>

namespace VW
{
Expand All @@ -20,7 +20,7 @@ class sparse_parameters;
namespace details
{

using weight_map = std::unordered_map<uint64_t, std::shared_ptr<VW::weight>>;
using weight_map = std::map<uint64_t, std::shared_ptr<VW::weight>>;
bassmang marked this conversation as resolved.
Show resolved Hide resolved

template <typename T>
class sparse_iterator
Expand Down Expand Up @@ -82,9 +82,11 @@ class sparse_parameters
const_iterator cend() const { return const_iterator(_map.end()); }

inline VW::weight& operator[](size_t i) { return *(get_or_default_and_get(i)); }

inline const VW::weight& operator[](size_t i) const { return *(get_or_default_and_get(i)); }

inline VW::weight& get(size_t i) { return *(get_impl(i)); };
inline const VW::weight& get(size_t i) const { return *(get_impl(i)); };

inline VW::weight& strided_index(size_t index) { return operator[](index << _stride_shift); }
inline const VW::weight& strided_index(size_t index) const { return operator[](index << _stride_shift); }

Expand Down Expand Up @@ -119,6 +121,7 @@ class sparse_parameters
// It is marked const so it can be used from both const and non const operator[]
// The map itself is mutable to facilitate this
VW::weight* get_or_default_and_get(size_t i) const;
VW::weight* get_impl(size_t i) const;
};
} // namespace VW
using sparse_parameters VW_DEPRECATED("sparse_parameters moved into VW namespace") = VW::sparse_parameters;
4 changes: 2 additions & 2 deletions vowpalwabbit/core/include/vw/core/gd_predict.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@ inline void foreach_feature(WeightsT& weights, const VW::features& fs, DataT& da
{
for (const auto& f : fs)
{
VW::weight& w = weights[(f.index() + offset)];
VW::weight& w = weights[f.index() + offset];
FuncT(dat, mult * f.value(), w);
}
}
Expand All @@ -46,7 +46,7 @@ template <class DataT, void (*FuncT)(DataT&, float, float), class WeightsT>
inline void foreach_feature(
const WeightsT& weights, const VW::features& fs, DataT& dat, uint64_t offset = 0, float mult = 1.)
{
for (const auto& f : fs) { FuncT(dat, mult * f.value(), weights[static_cast<size_t>(f.index() + offset)]); }
for (const auto& f : fs) { FuncT(dat, mult * f.value(), weights.get(static_cast<size_t>(f.index() + offset))); }
}

template <class DataT, class WeightOrIndexT, void (*FuncT)(DataT&, float, WeightOrIndexT),
Expand Down
2 changes: 1 addition & 1 deletion vowpalwabbit/core/include/vw/core/interactions_predict.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@ inline void call_func_t(DataT& dat, WeightsT& weights, const float ft_value, con
template <class DataT, void (*FuncT)(DataT&, const float, float), class WeightsT>
inline void call_func_t(DataT& dat, const WeightsT& weights, const float ft_value, const uint64_t ft_idx)
{
FuncT(dat, ft_value, weights[static_cast<size_t>(ft_idx)]);
FuncT(dat, ft_value, weights.get(static_cast<size_t>(ft_idx)));
}

template <class DataT, void (*FuncT)(DataT&, float, uint64_t), class WeightsT>
Expand Down
15 changes: 15 additions & 0 deletions vowpalwabbit/core/src/array_parameters_sparse.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,21 @@ VW::weight* VW::sparse_parameters::get_or_default_and_get(size_t i) const
return iter->second.get();
}

VW::weight* VW::sparse_parameters::get_impl(size_t i) const
{
static VW::weight default_value = 0.0f;

uint64_t index = i & _weight_mask;
auto iter = _map.find(index);
rajan-chari marked this conversation as resolved.
Show resolved Hide resolved
if (iter == _map.end())
{
if (_default_func != nullptr) { _default_func(&default_value, index); }
return &default_value;
}

return iter->second.get();
}

VW::sparse_parameters::sparse_parameters(size_t length, uint32_t stride_shift)
: _weight_mask((length << stride_shift) - 1), _stride_shift(stride_shift), _default_func(nullptr)
{
Expand Down
1 change: 1 addition & 0 deletions vowpalwabbit/core/src/reductions/cb/cb_explore_adf_rnd.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,7 @@ class lazy_gaussian
{
public:
inline float operator[](uint64_t index) const { return VW::details::merand48_boxmuller(index); }
inline float get(uint64_t index) const { return VW::details::merand48_boxmuller(index); }
bassmang marked this conversation as resolved.
Show resolved Hide resolved
};

inline void vec_add_with_norm(std::pair<float, float>& p, float fx, float fw)
Expand Down
Loading