Add support for multiple data types

[hcho3]

Addresses #95 and #111.

@trivialfis Your RTTI implementation in XGBoost JSON served as a great inspiration for this work. Thank you.

cc @canonizer @teju85

[codecov-io]

Codecov Report

Merging #130 into master will decrease coverage by 6.35%.
The diff coverage is 6.03%.

@@             Coverage Diff              @@
##             master     #130      +/-   ##
============================================
- Coverage     47.82%   41.47%   -6.36%     
============================================
  Files            81       67      -14     
  Lines          5733     5155     -578     
  Branches         37        0      -37     
============================================
- Hits           2742     2138     -604     
- Misses         2964     3017      +53     
+ Partials         27        0      -27

Impacted Files	Coverage Δ	Complexity Δ
include/treelite/base.h	0% <ø> (ø)	0 <0> (ø)	⬇️
include/treelite/frontend.h	0% <ø> (ø)	0 <0> (ø)	⬇️
include/treelite/adt_value_type.h	0% <0%> (ø)	0 <0> (?)
src/annotator.cc	1.11% <0%> (-0.08%)	0 <0> (ø)
src/frontend/lightgbm.cc	0.42% <0%> (ø)	0 <0> (ø)	⬇️
src/compiler/ast/ast.h	0% <0%> (ø)	0 <0> (ø)	⬇️
include/treelite/tree.h	7.5% <0%> (ø)	0 <0> (ø)	⬇️
src/compiler/ast/build.cc	2% <0%> (ø)	0 <0> (ø)	⬇️
src/frontend/protobuf.cc	0.41% <0%> (-0.05%)	0 <0> (ø)
src/frontend/builder.cc	0.46% <0%> (ø)	0 <0> (ø)	⬇️
... and 35 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update fca7387...7a3ec2b. Read the comment docs.

[hcho3]

Note: code coverage is broken.

[hcho3]

How to use the Value type:

Value value;
switch (value.Type()) {
 case treelite::ADT::ValueImpl::ValueKind::kInt32:
  int32_t int32val = treelite::ADT::get<treelite::ADT::Int32Value>(value);
  // ...
  break;
 case treelite::ADT::ValueImpl::ValueKind::kFloat32:
  float float32val = treelite::ADT::get<treelite::ADT::Float32Value>(value);
  // ...
  break;
 case treelite::ADT::ValueImpl::ValueKind::kFloat64:
  double float64val = treelite::ADT::get<treelite::ADT::Float64Value>(value);
  // ...
  break;
 default:
  LOG(FATAL) << "Unknown value variant";
}

[hcho3]

Need to also change generated lib to expose function interface with a different type.

[hcho3]

Now there are at least 4 different possible signatures for the prediction function:

int predict(union Entry* data, int pred_margin);
double predict(union Entry* data, int pred_margin);
size_t predict(union Entry* data, int pred_margin, float* result);
size_t predict(union Entry* data, int pred_margin, double* result);

This is quickly getting out of control. We should introduce type-erased interface to reduce complexity.

[CJxD]

Now there are at least 4 different possible signatures for the prediction function:

int predict(union Entry* data, int pred_margin);
double predict(union Entry* data, int pred_margin);
size_t predict(union Entry* data, int pred_margin, float* result);
size_t predict(union Entry* data, int pred_margin, double* result);

This is quickly getting out of control. We should introduce type-erased interface to reduce complexity.

Indeed. I strongly recommended an 8-bit datatype in here too. For the sake of reducing model size by as much as 75% in compiled C code, I always run my C source through a post processor to convert all floats to uint8_t.

I was about to create a pull request that could intelligently determine the smallest data type that could successfully separate the branches of a tree, but this pull request has the potential.

[shashankg7]

Hi,

Any plans to release this version anytime soon?

[simon19891101]

Hi there, may I know if this is merged? Just to say I'm still having the wrong predictions using the latest master #222. Thanks!

[hcho3]

@simon19891101 Take a look at #203