handle floating point values more accurate

.github/workflows/main.yml

@@ -3,7 +3,7 @@ name: GitHub Actions
 on:
   push:
     branches:
-      - master
+      - floats_improvement
   pull_request:
     branches:
       - master
@@ -17,8 +17,6 @@ jobs:
       matrix:
         python:
           - 3.6
-          - 3.7
-          - 3.8
     steps:
       - name: Checkout repository
         uses: actions/checkout@v1
@@ -46,12 +44,20 @@ jobs:
       matrix:
         python:
           - 3.6
-          - 3.7
-          - 3.8
         lang:
-          - "c_lang or python or java or go_lang or javascript or php or haskell or ruby"
-          - "c_sharp or visual_basic or f_sharp"
-          - "r_lang or dart"
+          - "c_lang"
+          - "python"
+          - "java"
+          - "go_lang"
+          - "javascript"
+          - "php"
+          - "haskell"
+          - "ruby"
+          - "c_sharp"
+          - "visual_basic"
+          - "f_sharp"
+          - "r_lang"
+          - "dart"
           - "powershell"
     steps:
       - name: Checkout repository

README.md

@@ -132,3 +132,7 @@ A: If this error occurs while generating code using an ensemble model, try to re
 **Q: Generation fails with `ImportError: No module named <module_name_here>` error while transpiling model from a serialized model object.**
 
 A: This error indicates that pickle protocol cannot deserialize model object. For unpickling serialized model objects, it is required that their classes must be defined in the top level of an importable module in the unpickling environment. So installation of package which provided model's class definition should solve the problem.
+
+**Q: Generated by m2cgen code provides different results for some inputs compared to original Python model from which the code were obtained.**
+
+A: Some models force input data to be particular type during prediction phase in their native Python libraries. Currently, m2cgen works only with ``float64`` (``double``) data type. You can try to cast your input data to another type manually and check results again. Also, some small differences can happen due to specific implementation of floating-point arithmetic in a target language.

m2cgen/assemblers/boosting.py

@@ -151,7 +151,7 @@ def __init__(self, model):
 
     def _assemble_tree(self, tree):
         if "leaf" in tree:
-            return ast.NumVal(tree["leaf"])
+            return ast.NumVal(tree["leaf"], dtype=np.float32)
 
         threshold = ast.NumVal(tree["split_condition"], dtype=np.float32)
         split = tree["split"]

m2cgen/assemblers/linear.py

@@ -1,7 +1,7 @@
 import numpy as np
 
 from m2cgen import ast
-from m2cgen.assemblers import utils
+from m2cgen.assemblers import fallback_expressions, utils
 from m2cgen.assemblers.base import ModelAssembler
 
 
@@ -95,14 +95,7 @@ def _get_supported_inversed_funs(self):
         raise NotImplementedError
 
     def _logit_inversed(self, ast_to_transform):
-        return utils.div(
-            ast.NumVal(1.0),
-            utils.add(
-                ast.NumVal(1.0),
-                ast.ExpExpr(
-                    utils.sub(
-                        ast.NumVal(0.0),
-                        ast_to_transform))))
+        return fallback_expressions.sigmoid(ast_to_transform)
 
     def _power_inversed(self, ast_to_transform):
         power = self._get_power()
@@ -146,16 +139,15 @@ def _cloglog_inversed(self, ast_to_transform):
 
     def _negativebinomial_inversed(self, ast_to_transform):
         alpha = self._get_alpha()
+        res = utils.sub(
+            ast.NumVal(1.0),
+            ast.ExpExpr(
+                utils.sub(
+                    ast.NumVal(0.0),
+                    ast_to_transform)))
         return utils.div(
             ast.NumVal(-1.0),
-            utils.mul(
-                ast.NumVal(alpha),
-                utils.sub(
-                    ast.NumVal(1.0),
-                    ast.ExpExpr(
-                        utils.sub(
-                            ast.NumVal(0.0),
-                            ast_to_transform)))))
+            utils.mul(ast.NumVal(alpha), res) if alpha != 1.0 else res)
 
     def _get_power(self):
         raise NotImplementedError

m2cgen/assemblers/tree.py

@@ -1,5 +1,3 @@
-import numpy as np
-
 from m2cgen import ast
 from m2cgen.assemblers import utils
 from m2cgen.assemblers.base import ModelAssembler
@@ -49,11 +47,5 @@ def _assemble_leaf(self, node_id):
 
     def _assemble_cond(self, node_id):
         feature_idx = self._tree.feature[node_id]
-        threshold = self._tree.threshold[node_id]
-
-        # sklearn's trees internally work with float32 numbers, so in order
-        # to have consistent results across all supported languages, we convert
-        # all thresholds into float32.
-        threshold_num_val = ast.NumVal(threshold, dtype=np.float32)
-
+        threshold_num_val = ast.NumVal(self._tree.threshold[node_id])
         return utils.lte(ast.FeatureRef(feature_idx), threshold_num_val)

m2cgen/interpreters/code_generator.py

@@ -1,6 +1,10 @@
 from io import StringIO
 from weakref import finalize
 
+import numpy as np
+
+from m2cgen.interpreters.utils import format_float
+
 
 class CodeTemplate:
 
@@ -11,12 +15,14 @@ def __str__(self):
         return self.str_template
 
     def __call__(self, *args, **kwargs):
-        # Force calling str() representation
-        # because without it numpy gives the same output
-        # for different float types
+
+        def _is_float(value):
+            return isinstance(value, (float, np.floating))
+
         return self.str_template.format(
-            *[str(i) for i in args],
-            **{k: str(v) for k, v in kwargs.items()})
+            *[format_float(i) if _is_float(i) else i for i in args],
+            **{k: format_float(v) if _is_float(v) else v
+               for k, v in kwargs.items()})
 
 
 class BaseCodeGenerator:

m2cgen/interpreters/utils.py

@@ -1,5 +1,7 @@
 import re
 
+import numpy as np
+
 from collections import namedtuple
 from functools import lru_cache
 from math import ceil, log
@@ -22,3 +24,7 @@ def _get_handler_name(expr_tpe):
 
 def _normalize_expr_name(name):
     return re.sub("(?!^)([A-Z]+)", r"_\1", name).lower()
+
+
+def format_float(value):
+    return np.format_float_positional(value, unique=True, trim="0")

tests/assemblers/test_lightgbm.py

@@ -26,15 +26,15 @@ def test_binary_classification():
                                 ast.FeatureRef(23),
                                 ast.NumVal(868.2000000000002),
                                 ast.CompOpType.GT),
-                            ast.NumVal(0.25986931215073095),
-                            ast.NumVal(0.6237178414050242)),
+                            ast.NumVal(0.26400127816506497),
+                            ast.NumVal(0.633133056485969)),
                         ast.IfExpr(
                             ast.CompExpr(
-                                ast.FeatureRef(7),
-                                ast.NumVal(0.05142),
+                                ast.FeatureRef(22),
+                                ast.NumVal(105.95000000000002),
                                 ast.CompOpType.GT),
-                            ast.NumVal(-0.1909605544006228),
-                            ast.NumVal(0.1293965108676673)),
+                            ast.NumVal(-0.18744882409486507),
+                            ast.NumVal(0.13458899352064668)),
                         ast.BinNumOpType.ADD),
                     ast.BinNumOpType.SUB)),
             ast.BinNumOpType.ADD),
@@ -85,17 +85,17 @@ def test_regression():
         ast.IfExpr(
             ast.CompExpr(
                 ast.FeatureRef(5),
-                ast.NumVal(6.918),
+                ast.NumVal(6.837500000000001),
                 ast.CompOpType.GT),
-            ast.NumVal(24.011454621684155),
-            ast.NumVal(22.289277544391084)),
+            ast.NumVal(23.961356387224317),
+            ast.NumVal(22.32858336612959)),
         ast.IfExpr(
             ast.CompExpr(
                 ast.FeatureRef(12),
-                ast.NumVal(9.63),
+                ast.NumVal(9.725000000000003),
                 ast.CompOpType.GT),
-            ast.NumVal(-0.49461212269771115),
-            ast.NumVal(0.7174324413014594)),
+            ast.NumVal(-0.5031712645462916),
+            ast.NumVal(0.6885501354513913)),
         ast.BinNumOpType.ADD)
 
     assert utils.cmp_exprs(actual, expected)
@@ -114,18 +114,18 @@ def test_regression_random_forest():
         ast.BinNumExpr(
             ast.IfExpr(
                 ast.CompExpr(
-                    ast.FeatureRef(5),
-                    ast.NumVal(6.954000000000001),
+                    ast.FeatureRef(12),
+                    ast.NumVal(5.200000000000001),
                     ast.CompOpType.GT),
-                ast.NumVal(37.24347877367631),
-                ast.NumVal(19.936999995530854)),
+                ast.NumVal(20.206688945020474),
+                ast.NumVal(38.30000037757679)),
             ast.IfExpr(
                 ast.CompExpr(
                     ast.FeatureRef(5),
-                    ast.NumVal(6.971500000000001),
+                    ast.NumVal(6.837500000000001),
                     ast.CompOpType.GT),
-                ast.NumVal(38.48600037864964),
-                ast.NumVal(20.183783757300255)),
+                ast.NumVal(36.40634951405711),
+                ast.NumVal(19.57067132709245)),
             ast.BinNumOpType.ADD),
         ast.NumVal(0.5),
         ast.BinNumOpType.MUL)
@@ -154,15 +154,15 @@ def test_simple_sigmoid_output_transform():
                                 ast.FeatureRef(12),
                                 ast.NumVal(19.23),
                                 ast.CompOpType.GT),
-                            ast.NumVal(4.0026305187),
-                            ast.NumVal(4.0880438137)),
+                            ast.NumVal(4.0050691250),
+                            ast.NumVal(4.0914737728)),
                         ast.IfExpr(
                             ast.CompExpr(
                                 ast.FeatureRef(12),
-                                ast.NumVal(14.895),
+                                ast.NumVal(15.065),
                                 ast.CompOpType.GT),
-                            ast.NumVal(-0.0412703078),
-                            ast.NumVal(0.0208393767)),
+                            ast.NumVal(-0.0420531079),
+                            ast.NumVal(0.0202891577)),
                         ast.BinNumOpType.ADD),
                     ast.BinNumOpType.SUB)),
             ast.BinNumOpType.ADD),
@@ -188,15 +188,15 @@ def test_log1p_exp_output_transform():
                         ast.FeatureRef(12),
                         ast.NumVal(19.23),
                         ast.CompOpType.GT),
-                    ast.NumVal(0.6623502468),
-                    ast.NumVal(0.6683497987)),
+                    ast.NumVal(0.6623996001),
+                    ast.NumVal(0.6684477608)),
                 ast.IfExpr(
                     ast.CompExpr(
                         ast.FeatureRef(12),
-                        ast.NumVal(15.145),
+                        ast.NumVal(15.065),
                         ast.CompOpType.GT),
-                    ast.NumVal(0.1405181490),
-                    ast.NumVal(0.1453602134)),
+                    ast.NumVal(0.1405782705),
+                    ast.NumVal(0.1453764991)),
                 ast.BinNumOpType.ADD)))
 
     assert utils.cmp_exprs(actual, expected)
@@ -216,17 +216,17 @@ def test_maybe_sqr_output_transform():
             ast.IfExpr(
                 ast.CompExpr(
                     ast.FeatureRef(12),
-                    ast.NumVal(9.905),
+                    ast.NumVal(11.655),
                     ast.CompOpType.GT),
-                ast.NumVal(4.5658116817),
-                ast.NumVal(4.6620790482)),
+                ast.NumVal(4.5671830654),
+                ast.NumVal(4.6516575813)),
             ast.IfExpr(
                 ast.CompExpr(
                     ast.FeatureRef(12),
-                    ast.NumVal(9.77),
+                    ast.NumVal(9.725),
                     ast.CompOpType.GT),
-                ast.NumVal(-0.0340889740),
-                ast.NumVal(0.0543687153)),
+                ast.NumVal(-0.0348178434),
+                ast.NumVal(0.0549301624)),
             ast.BinNumOpType.ADD),
         to_reuse=True)
 
@@ -251,17 +251,17 @@ def test_exp_output_transform():
             ast.IfExpr(
                 ast.CompExpr(
                     ast.FeatureRef(5),
-                    ast.NumVal(6.918),
+                    ast.NumVal(6.8375),
                     ast.CompOpType.GT),
-                ast.NumVal(3.1480683932),
-                ast.NumVal(3.1101554907)),
+                ast.NumVal(3.1481886430),
+                ast.NumVal(3.1123367238)),
             ast.IfExpr(
                 ast.CompExpr(
                     ast.FeatureRef(12),
-                    ast.NumVal(9.63),
+                    ast.NumVal(9.725),
                     ast.CompOpType.GT),
-                ast.NumVal(-0.0111969636),
-                ast.NumVal(0.0160298303)),
+                ast.NumVal(-0.0113689739),
+                ast.NumVal(0.0153551274)),
             ast.BinNumOpType.ADD))
 
     assert utils.cmp_exprs(actual, expected)
@@ -289,8 +289,8 @@ def test_bin_class_sigmoid_output_transform():
                                 ast.FeatureRef(23),
                                 ast.NumVal(868.2),
                                 ast.CompOpType.GT),
-                            ast.NumVal(0.5197386243),
-                            ast.NumVal(1.2474356828)),
+                            ast.NumVal(0.5280025563),
+                            ast.NumVal(1.2662661130)),
                         ast.BinNumOpType.MUL),
                     ast.BinNumOpType.SUB)),
             ast.BinNumOpType.ADD),