trees: Use int16 features instead of float

Means that evaluation does not need floating point
Should be considerably smaller and faster,
especially on platforms without FPU

Author: jonnor

examples/xor_trees/xor_model.csv

@@ -5,17 +5,17 @@ l,1
 r,0
 r,5
 r,9
-n,1,0.197349,1,2
-n,0,0.466316,-1,-2
-n,1,0.256702,1,2
-n,0,0.464752,-1,-2
-n,0,0.476659,-2,-1
-n,1,0.897163,1,-2
-n,0,0.504671,1,2
-n,1,0.501874,-1,-2
-n,1,0.498237,-2,-1
-n,0,0.11973,1,2
-n,1,0.394697,-1,-2
-n,1,0.261566,1,2
-n,0,0.473531,-1,-2
-n,0,0.421164,-2,-1
+n,1,6466.0,1,2
+n,0,15279.5,-1,-2
+n,1,8411.0,1,2
+n,0,15228.0,-1,-2
+n,0,15618.5,-2,-1
+n,1,29397.0,1,-2
+n,0,16536.0,1,2
+n,1,16444.5,-1,-2
+n,1,16325.5,-2,-1
+n,0,3922.5,1,2
+n,1,12932.5,-1,-2
+n,1,8570.5,1,2
+n,0,15516.0,-1,-2
+n,0,13800.0,-2,-1

examples/xor_trees/xor_train.py

@@ -1,6 +1,7 @@
 # python/host code
 
 import emlearn
+from emlearn.preprocessing import Quantizer
 import numpy
 from sklearn.ensemble import RandomForestClassifier
 from sklearn.metrics import get_scorer
@@ -10,6 +11,7 @@ def make_xor(lower=0.0, upper=1.0, threshold=0.5, samples=100, seed=42):
     rng = numpy.random.RandomState(seed)
     X = rng.uniform(lower, upper, size=(samples, 2))
     y = numpy.logical_xor(X[:, 0] > threshold, X[:, 1] > threshold)
+    X = Quantizer(max_value=1.0).fit_transform(X) # convert to int16
     return X, y
 
 X, y = make_xor()

src/emltrees/trees.c

@@ -125,8 +125,8 @@ static mp_obj_t builder_addnode(size_t n_args, const mp_obj_t *args) {
 
     const int16_t left = mp_obj_get_int(args[1]);
     const int16_t right = mp_obj_get_int(args[2]);
-    const mp_int_t feature = mp_obj_get_int(args[3]);
-    const float value = mp_obj_get_float_to_f(args[4]);
+    const int feature = mp_obj_get_int(args[3]);
+    const int16_t value = mp_obj_get_int(args[4]);
 
     if (feature > 127 || feature < -1) {
         mp_raise_ValueError(MP_ERROR_TEXT("feature out of bounds"));
@@ -191,7 +191,7 @@ static mp_obj_t builder_addleaf(mp_obj_t self_obj, mp_obj_t leaf_obj) {
 static MP_DEFINE_CONST_FUN_OBJ_2(builder_addleaf_obj, builder_addleaf);
 
 
-// Takes a float array
+// Takes a array of input data
 static mp_obj_t builder_predict(mp_obj_t self_obj, mp_obj_t features_obj) {
 
     mp_obj_trees_builder_t *o = MP_OBJ_TO_PTR(self_obj);
@@ -200,11 +200,11 @@ static mp_obj_t builder_predict(mp_obj_t self_obj, mp_obj_t features_obj) {
     // Extract buffer pointer and verify typecode
     mp_buffer_info_t bufinfo;
     mp_get_buffer_raise(features_obj, &bufinfo, MP_BUFFER_RW);
-    if (bufinfo.typecode != 'f') {
-        mp_raise_ValueError(MP_ERROR_TEXT("expecting float array"));
+    if (bufinfo.typecode != 'h') {
+        mp_raise_ValueError(MP_ERROR_TEXT("expecting int16 (h) array"));
     }
 
-    float *features = bufinfo.buf;
+    const int16_t *features = bufinfo.buf;
     const int n_features = bufinfo.len / sizeof(*features);
 
 #if EMLEARN_MICROPYTHON_DEBUG

src/emltrees/trees.py

@@ -16,7 +16,7 @@ def load_model(builder, f):
             builder.addroot(root)
         elif kind == 'n':
             feature = int(tok[1])
-            value = float(tok[2])
+            value = int(float(tok[2]))
             left = int(tok[3])
             right = int(tok[4])
             builder.addnode(left, right, feature, value)

tests/test_trees.py

@@ -36,16 +36,17 @@ def test_trees_xor():
         emltrees.load_model(model, f)
 
     # run it
+    s = 32767 # max int16
     examples = [
         # input, expected output
-        ( [0.0, 0.0], 0 ),
-        ( [1.0, 1.0], 0 ),
-        ( [0.0, 1.0], 1 ),
-        ( [1.0, 0.0], 1 ),
+        ( [0, 0], 0 ),
+        ( [1*s, 1*s], 0 ),
+        ( [0, 1*s], 1 ),
+        ( [1*s, 0], 1 ),
     ]
 
     for (ex, expect) in examples:
-        f = array.array('f', ex)
+        f = array.array('h', ex)
         result = model.predict(f)
         assert result == expect, (ex, expect, result)