Merge pull request emlearn#26 from emlearn/trees-proba

trees: Switch to output prediction probabilities

Author: jonnor

examples/har_trees/har_live.py

@@ -15,6 +15,16 @@ def mean(arr):
     m = sum(arr) / float(len(arr))
     return m
 
+def argmax(arr):
+    idx_max = 0
+    value_max = arr[0]
+    for i in range(1, len(arr)):
+        if arr[i] > value_max:
+            value_max = arr[i]
+            idx_max = i
+
+    return idx_max
+
 def copy_array_into(source, target):
     assert len(source) == len(target)
     for i in range(len(target)):
@@ -63,6 +73,7 @@ def main():
     features_typecode = timebased.DATA_TYPECODE
     n_features = timebased.N_FEATURES
     features = array.array(features_typecode, (0 for _ in range(n_features)))
+    out = array.array('f', range(model.outputs()))
 
     while True:
 
@@ -87,7 +98,8 @@ def main():
 
                 # Cun classifier
                 #print(features)
-                result = model.predict(features)
+                model.predict(features, out)
+                result = argmax(out)
                 activity = class_index_to_name[result]
 
                 d = time.ticks_diff(time.ticks_ms(), start)

examples/har_trees/har_run.py

@@ -6,6 +6,16 @@
 import emlearn_trees
 import timebased
 
+def argmax(arr):
+    idx_max = 0
+    value_max = arr[0]
+    for i in range(1, len(arr)):
+        if arr[i] > value_max:
+            value_max = arr[i]
+            idx_max = i
+
+    return idx_max
+
 def har_load_test_data(path,
         skip_samples=0, limit_samples=None):
 
@@ -63,6 +73,7 @@ def main():
     with open(model_path, 'r') as f:
         emlearn_trees.load_model(model, f)
 
+    out = array.array('f', range(model.outputs()))
 
     errors = 0
     total = 0
@@ -72,7 +83,8 @@ def main():
 
         assert len(labels) == 1
         label = labels[0]
-        result = model.predict(features)
+        model.predict(features, out)
+        result = argmax(out)
         if result != label:
             errors += 1
         total += 1

src/emlearn_trees/trees.c

@@ -197,8 +197,25 @@ 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);
 
 
+// Return the shape of the output
+static mp_obj_t builder_get_outputs(mp_obj_t self_obj) {
+
+    mp_obj_trees_builder_t *o = MP_OBJ_TO_PTR(self_obj);
+    EmlTreesBuilder *self = &o->builder;
+
+    const int n_classes = self->trees.n_classes;
+    if (n_classes == 0) {
+        mp_raise_ValueError(MP_ERROR_TEXT("model not loaded"));
+    }
+
+    return mp_obj_new_int(n_classes);
+}
+static MP_DEFINE_CONST_FUN_OBJ_1(builder_get_outputs_obj, builder_get_outputs);
+
+
+
 // Takes a array of input data
-static mp_obj_t builder_predict(mp_obj_t self_obj, mp_obj_t features_obj) {
+static mp_obj_t builder_predict(mp_obj_t self_obj, mp_obj_t features_obj, mp_obj_t output_obj) {
 
     mp_obj_trees_builder_t *o = MP_OBJ_TO_PTR(self_obj);
     EmlTreesBuilder *self = &o->builder;    
@@ -212,28 +229,45 @@ static mp_obj_t builder_predict(mp_obj_t self_obj, mp_obj_t features_obj) {
 
     const int16_t *features = bufinfo.buf;
     const int n_features = bufinfo.len / sizeof(*features);
+    const int n_outputs = self->trees.n_classes;
 
 #if EMLEARN_MICROPYTHON_DEBUG
     mp_printf(&mp_plat_print,
-        "emltrees-predict n_features=%d n_classes=%d leaves=%d nodes=%d trees=%d length=%d \n",
+        "emltrees-predict n_features=%d n_classes=%d leaves=%d nodes=%d trees=%d length=%d outputs=%d \n",
         self->trees.n_features, self->trees.n_classes,
         self->trees.n_leaves, self->trees.n_nodes, self->trees.n_trees,
         n_features
     );
 #endif
 
+    if (n_features == 0 || n_outputs == 0) {        
+        mp_raise_ValueError(MP_ERROR_TEXT("model not loaded"));
+    }
+
+    // Extract output
+    mp_get_buffer_raise(output_obj, &bufinfo, MP_BUFFER_RW);
+    if (bufinfo.typecode != 'f') {
+        mp_raise_ValueError(MP_ERROR_TEXT("expecting float output array"));
+    }
+    float *output_buffer = bufinfo.buf;
+    const int output_length = bufinfo.len / sizeof(*output_buffer);
+
+
     // call model
-    const int result = eml_trees_predict(&self->trees, features, n_features);
-    if (result < 0) {
-        mp_raise_msg(&mp_type_RuntimeError, MP_ERROR_TEXT("eml_trees_predict error"));
+    // NOTE: also handles checking of input and output lengths
+    const EmlError err = \
+        eml_trees_predict_proba(&self->trees, features, n_features, output_buffer, output_length);
+
+    if (err != EmlOk) {
+        mp_raise_msg(&mp_type_RuntimeError, MP_ERROR_TEXT("eml_trees_predict_proba error"));
     }
 
-    return mp_obj_new_int(result);
+    return mp_const_none;
 }
-static MP_DEFINE_CONST_FUN_OBJ_2(builder_predict_obj, builder_predict);
+static MP_DEFINE_CONST_FUN_OBJ_3(builder_predict_obj, builder_predict);
 
 
-mp_map_elem_t trees_locals_dict_table[6];
+mp_map_elem_t trees_locals_dict_table[7];
 static MP_DEFINE_CONST_DICT(trees_locals_dict, trees_locals_dict_table);
 
 // This is the entry point and is called when the module is imported
@@ -253,8 +287,9 @@ mp_obj_t mpy_init(mp_obj_fun_bc_t *self, size_t n_args, size_t n_kw, mp_obj_t *a
     trees_locals_dict_table[3] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_addleaf), MP_OBJ_FROM_PTR(&builder_addleaf_obj) };
     trees_locals_dict_table[4] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR___del__), MP_OBJ_FROM_PTR(&builder_del_obj) };
     trees_locals_dict_table[5] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_setdata), MP_OBJ_FROM_PTR(&builder_setdata_obj) };
+    trees_locals_dict_table[6] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_outputs), MP_OBJ_FROM_PTR(&builder_get_outputs_obj) };
 
-    MP_OBJ_TYPE_SET_SLOT(&trees_builder_type, locals_dict, (void*)&trees_locals_dict, 6);
+    MP_OBJ_TYPE_SET_SLOT(&trees_builder_type, locals_dict, (void*)&trees_locals_dict, 7);
 
     // This must be last, it restores the globals dict
     MP_DYNRUNTIME_INIT_EXIT

tests/test_cnn.py

@@ -73,7 +73,7 @@ def test_cnn_mnist():
             if out == class_no:
                 correct += 1            
 
-    assert correct >= 6, correct
+    assert correct >= 9, correct
 
 
 test_cnn_create()

tests/test_trees.py

@@ -4,6 +4,16 @@
 import array
 import gc
 
+def argmax(arr):
+    idx_max = 0
+    value_max = arr[0]
+    for i in range(1, len(arr)):
+        if arr[i] > value_max:
+            value_max = arr[i]
+            idx_max = i
+
+    return idx_max
+
 def test_trees_del():
     """
     Deleting the model should free all the memory
@@ -45,9 +55,12 @@ def test_trees_xor():
         ( [1*s, 0], 1 ),
     ]
 
+    out = array.array('f', range(model.outputs()))
+
     for (ex, expect) in examples:
         f = array.array('h', ex)
-        result = model.predict(f)
+        model.predict(f, out)
+        result = argmax(out)
         assert result == expect, (ex, expect, result)
 
 test_trees_del()