add XGBClassifier

MANIFEST

@@ -5,6 +5,7 @@ micromlgen/__init__.py
 micromlgen/convolution.py
 micromlgen/decisiontree.py
 micromlgen/gaussiannb.py
+micromlgen/linear_regression.py
 micromlgen/logisticregression.py
 micromlgen/micromlgen.py
 micromlgen/pca.py
@@ -16,6 +17,7 @@ micromlgen/sefr.py
 micromlgen/svm.py
 micromlgen/utils.py
 micromlgen/wifiindoorpositioning.py
+micromlgen/xgboost.py
 micromlgen/templates/_skeleton.jinja
 micromlgen/templates/classmap.jinja
 micromlgen/templates/dot.jinja
@@ -27,6 +29,7 @@ micromlgen/templates/decisiontree/decisiontree.jinja
 micromlgen/templates/decisiontree/tree.jinja
 micromlgen/templates/gaussiannb/gaussiannb.jinja
 micromlgen/templates/gaussiannb/vote.jinja
+micromlgen/templates/linearregression/linearregression.jinja
 micromlgen/templates/logisticregression/logisticregression.jinja
 micromlgen/templates/logisticregression/vote.arduino.jinja
 micromlgen/templates/logisticregression/vote.attiny.jinja
@@ -49,3 +52,5 @@ micromlgen/templates/svm/kernel/arduino.jinja
 micromlgen/templates/svm/kernel/attiny.jinja
 micromlgen/templates/svm/kernel/kernel.jinja
 micromlgen/templates/wifiindoorpositioning/wifiindoorpositioning.jinja
+micromlgen/templates/xgboost/tree.jinja
+micromlgen/templates/xgboost/xgboost.jinja

micromlgen/micromlgen.py

@@ -9,7 +9,7 @@
 from micromlgen.pca import is_pca, port_pca
 from micromlgen.principalfft import is_principalfft, port_principalfft
 from micromlgen.linear_regression import is_linear_regression, port_linear_regression
-
+from micromlgen.xgboost import is_xgboost, port_xgboost
 
 def port(
         clf,
@@ -40,4 +40,6 @@ def port(
         return port_principalfft(**locals(), **kwargs)
     elif is_linear_regression(clf):
         return port_linear_regression(**locals(), **kwargs)
+    elif is_xgboost(clf):
+        return port_xgboost(**locals(), **kwargs)
     raise TypeError('clf MUST be one of %s' % ', '.join(platforms.ALLOWED_CLASSIFIERS))

micromlgen/platforms.py

@@ -16,5 +16,6 @@
     'GaussianNB',
     'LogisticRegression',
     'PCA',
-    'PrincipalFFT'
+    'PrincipalFFT',
+    'LinearRegression'
 ]

micromlgen/templates/xgboost/tree.jinja

@@ -0,0 +1,14 @@
+{% if tree['left'][i] != tree['right'][i] %}
+    if (x[{{ tree['features'][i] }}] <= {{ tree['thresholds'][i] }}) {
+        {% with i = tree['left'][i] %}
+            {% include 'xgboost/tree.jinja' %}
+        {% endwith %}
+    }
+    else {
+        {% with i = tree['right'][i] %}
+            {% include 'xgboost/tree.jinja' %}
+        {% endwith %}
+    }
+{% else %}
+    votes[{{ class_idx }}] += {{ tree['thresholds'][i] }};
+{% endif %}

micromlgen/templates/xgboost/xgboost.jinja

@@ -0,0 +1,14 @@
+{% extends '_skeleton.jinja' %}
+
+{% block predict %}
+    float votes[{{ n_classes }}] = { 0.0f };
+
+    {% for k, tree in f.enumerate(trees) %}
+        {% with i = 0, class_idx = k % n_classes %}
+            // tree #{{ k + 1 }}
+            {% include 'xgboost/tree.jinja' %}
+        {% endwith %}
+    {% endfor %}
+
+    {% include 'vote.jinja' %}
+{% endblock %}

micromlgen/xgboost.py

@@ -0,0 +1,42 @@
+from micromlgen.utils import jinja, check_type
+from tempfile import NamedTemporaryFile
+import json
+
+
+def format_tree(tree):
+    """
+    Format xgboost tree like a sklearn DecisionTree
+    :param tree:
+    :return:
+    """
+    split_indices = tree['split_indices']
+    split_conditions = tree['split_conditions']
+    left_children = tree['left_children']
+    right_children = tree['right_children']
+    return {
+        'left': left_children,
+        'right': right_children,
+        'features': split_indices,
+        'thresholds': split_conditions
+    }
+
+
+def is_xgboost(clf):
+    """Test if classifier can be ported"""
+    return check_type(clf, 'XGBClassifier')
+
+
+def port_xgboost(clf, **kwargs):
+    """Port a XGBoost classifier"""
+    with NamedTemporaryFile('w+', suffix='.json', encoding='utf-8') as tmp:
+        clf.save_model(tmp.name)
+        tmp.seek(0)
+        decoded = json.load(tmp)
+        trees = [format_tree(tree) for tree in decoded['learner']['gradient_booster']['model']['trees']]
+        print(trees)
+        return jinja('xgboost/xgboost.jinja', {
+            'n_classes': int(decoded['learner']['learner_model_param']['num_class']),
+            'trees': trees,
+        }, {
+            'classname': 'XGBClassifier'
+        }, **kwargs)

setup.py

@@ -7,13 +7,13 @@
 setup(
   name = 'micromlgen',
   packages = ['micromlgen'],
-  version = '1.1.10',
+  version = '1.1.11',
   license='MIT',
   description = 'Generate C code for microcontrollers from Python\'s sklearn classifiers',
   author = 'Simone Salerno',
   author_email = '[email protected]',
   url = 'https://github.com/eloquentarduino/micromlgen',
-  download_url = 'https://github.com/eloquentarduino/micromlgen/archive/v_1110.tar.gz',
+  download_url = 'https://github.com/eloquentarduino/micromlgen/archive/v_1111.tar.gz',
   keywords = [
     'ML',
     'microcontrollers',