logreg: Remove predict_class

Consumers should use the continious values

Author: jonnor

src/emlearn_logreg/Makefile

@@ -1,5 +1,5 @@
 # Location of top-level MicroPython directory
-MPY_DIR = ../../micropython
+MPY_DIR ?= ../../micropython
 
 # Architecture to build for (x86, x64, armv6m, armv7m, xtensa, xtensawin)
 ARCH = x64

src/emlearn_logreg/logreg.c

@@ -137,30 +137,6 @@ static mp_obj_t logreg_model_predict(mp_obj_t self_obj, mp_obj_t features_obj) {
 }
 static MP_DEFINE_CONST_FUN_OBJ_2(logreg_model_predict_obj, logreg_model_predict);
 
-static mp_obj_t logreg_model_predict_class(mp_obj_t self_obj, mp_obj_t features_obj) {
-    mp_obj_logreg_model_t *o = MP_OBJ_TO_PTR(self_obj);
-    logreg_model_t *self = &o->model;
-
-    mp_buffer_info_t bufinfo;
-    mp_get_buffer_raise(features_obj, &bufinfo, MP_BUFFER_READ);
-    if (bufinfo.typecode != 'f') {
-        mp_raise_ValueError(MP_ERROR_TEXT("expecting float32 array"));
-    }
-    const float *features = bufinfo.buf;
-    const int n_features = bufinfo.len / sizeof(float);
-
-    if (n_features != self->n_features) {
-        mp_raise_ValueError(MP_ERROR_TEXT("Feature count mismatch"));
-    }
-
-    const float threshold = 0.5f;
-
-    uint8_t label = logreg_predict_proba(self, features) >= threshold ? 1 : 0;
-
-    return mp_obj_new_int(label);
-}
-static MP_DEFINE_CONST_FUN_OBJ_2(logreg_model_predict_class_obj, logreg_model_predict_class);
-
 // Get model weights
 static mp_obj_t logreg_model_get_weights(mp_obj_t self_obj, mp_obj_t out_obj) {
     mp_obj_logreg_model_t *o = MP_OBJ_TO_PTR(self_obj);
@@ -279,17 +255,16 @@ mp_obj_t mpy_init(mp_obj_fun_bc_t *self, size_t n_args, size_t n_kw, mp_obj_t *a
     logreg_model_type.name = MP_QSTR_logreg;
 
     logreg_model_locals_dict_table[0] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_predict), MP_OBJ_FROM_PTR(&logreg_model_predict_obj) };
-    logreg_model_locals_dict_table[1] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_predict_class), MP_OBJ_FROM_PTR(&logreg_model_predict_class_obj) };
-    logreg_model_locals_dict_table[2] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_step), MP_OBJ_FROM_PTR(&logreg_model_step_obj) };
-    logreg_model_locals_dict_table[3] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR___del__), MP_OBJ_FROM_PTR(&logreg_model_del_obj) };
-    logreg_model_locals_dict_table[4] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_get_weights), MP_OBJ_FROM_PTR(&logreg_model_get_weights_obj) };
-    logreg_model_locals_dict_table[5] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_set_weights), MP_OBJ_FROM_PTR(&logreg_model_set_weights_obj) };
-    logreg_model_locals_dict_table[6] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_get_bias), MP_OBJ_FROM_PTR(&logreg_model_get_bias_obj) };
-    logreg_model_locals_dict_table[7] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_set_bias), MP_OBJ_FROM_PTR(&logreg_model_set_bias_obj) };
-    logreg_model_locals_dict_table[8] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_get_n_features), MP_OBJ_FROM_PTR(&logreg_model_get_n_features_obj) };
-    logreg_model_locals_dict_table[9] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_score_logloss), MP_OBJ_FROM_PTR(&logreg_model_score_logloss_obj) };
-
-    MP_OBJ_TYPE_SET_SLOT(&logreg_model_type, locals_dict, (void *)&logreg_model_locals_dict, 10);
+    logreg_model_locals_dict_table[1] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_step), MP_OBJ_FROM_PTR(&logreg_model_step_obj) };
+    logreg_model_locals_dict_table[2] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR___del__), MP_OBJ_FROM_PTR(&logreg_model_del_obj) };
+    logreg_model_locals_dict_table[3] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_get_weights), MP_OBJ_FROM_PTR(&logreg_model_get_weights_obj) };
+    logreg_model_locals_dict_table[4] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_set_weights), MP_OBJ_FROM_PTR(&logreg_model_set_weights_obj) };
+    logreg_model_locals_dict_table[5] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_get_bias), MP_OBJ_FROM_PTR(&logreg_model_get_bias_obj) };
+    logreg_model_locals_dict_table[6] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_set_bias), MP_OBJ_FROM_PTR(&logreg_model_set_bias_obj) };
+    logreg_model_locals_dict_table[7] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_get_n_features), MP_OBJ_FROM_PTR(&logreg_model_get_n_features_obj) };
+    logreg_model_locals_dict_table[8] = (mp_map_elem_t){ MP_OBJ_NEW_QSTR(MP_QSTR_score_logloss), MP_OBJ_FROM_PTR(&logreg_model_score_logloss_obj) };
+
+    MP_OBJ_TYPE_SET_SLOT(&logreg_model_type, locals_dict, (void *)&logreg_model_locals_dict, 9);
 
     MP_DYNRUNTIME_INIT_EXIT
 }

tests/test_logreg.py

@@ -83,6 +83,10 @@ def read_weights(model):
     return buf
 
 
+def read_bias(model):
+    return model.get_bias()
+
+
 def assert_raises_value_error(func, message='Expected ValueError'):
     try:
         func()
@@ -105,7 +109,7 @@ def test_logreg_train_and_predict():
     assert zero_pred < 0.2, zero_pred
 
 
-def test_logreg_predict_class_and_weight_io():
+def test_logreg_weight_io_and_probabilities():
     model = emlearn_logreg.new(2, 0.1, 0.0, 0.0)
 
     manual_weights = array.array('f', [2.5, -1.5])
@@ -120,8 +124,8 @@ def test_logreg_predict_class_and_weight_io():
     bias = model.get_bias()
     assert abs(bias + 0.5) < 1e-6, bias
 
-    assert model.predict_class(array.array('f', [2.0, 0.0])) == 1
-    assert model.predict_class(array.array('f', [0.0, 1.0])) == 0
+    assert model.predict(array.array('f', [2.0, 0.0])) > 0.5
+    assert model.predict(array.array('f', [0.0, 1.0])) < 0.5
 
 
 def test_logreg_train_minibatch_reduces_loss():
@@ -228,9 +232,44 @@ def test_logreg_train_requires_targets():
     assert_raises_value_error(lambda: emlearn_logreg.train(model, X, y))
 
 
+def test_logreg_warm_start_sets_new_weights_and_bias():
+    X, y = make_dataset()
+    model = emlearn_logreg.new(2, 0.3, 0.0, 0.0)
+
+    emlearn_logreg.train(model, X, y, max_iterations=20, check_interval=5)
+    trained_weights = read_weights(model)
+    trained_bias = read_bias(model)
+
+    manual_model = emlearn_logreg.new(2, 0.3, 0.0, 0.0)
+    manual_model.set_weights(trained_weights)
+    manual_model.set_bias(trained_bias)
+
+    sample = array.array('f', [1.0, 1.0])
+    pred_trained = model.predict(sample)
+    pred_manual = manual_model.predict(sample)
+    assert abs(pred_trained - pred_manual) < 1e-6
+
+
+def test_logreg_threshold_adjustment_behaviour():
+    model = emlearn_logreg.new(2, 0.1, 0.0, 0.0)
+    weights = array.array('f', [5.0, -5.0])
+    model.set_weights(weights)
+    model.set_bias(-1.0)
+
+    features = array.array('f', [0.2, 0.1])
+    proba = model.predict(features)
+    assert 0.0 < proba < 1.0
+
+    default_label = 1 if proba >= 0.5 else 0
+    custom_threshold = 0.3
+    custom_label = 1 if proba >= custom_threshold else 0
+
+    assert custom_label >= default_label
+
+
 if __name__ == '__main__':
     test_logreg_train_and_predict()
-    test_logreg_predict_class_and_weight_io()
+    test_logreg_weight_io_and_probabilities()
     test_logreg_train_minibatch_reduces_loss()
     test_logreg_l2_penalty_shrinks_weights()
     test_logreg_l1_penalty_promotes_sparsity()