Fix: remove old values from feature values

bt_monitor_server.py

@@ -1,16 +1,16 @@
 #!/usr/bin/env python3
 
+import os
 import aiomqtt
 import json
 import asyncio
 from time import time
 from pathlib import Path
-from collections import namedtuple, defaultdict
+from collections import namedtuple, defaultdict, deque
 from typing import Dict, Optional, List
 from Crypto.Cipher import AES
 import pandas as pd
 import numpy as np
-from copy import copy
 import logging
 from sklearn import svm
 from sklearn.model_selection import cross_val_score
@@ -103,10 +103,20 @@ class KnownRoomCsvLogger:
             file=self.csv_file_handle,)
 
 
-FAR_AWAY_FEATURE_VALUE = 1
+class RunningFeatureVector:
+    FAR_AWAY_FEATURE_VALUE = 1 
+    MIN_TIME_UNTIL_PREDICTION = 40 # wait until every reachable tracker detected the device
+    TIME_TO_DELETE_IF_NOT_SEEN = 30 # if device wasn't spotted for this time period, the measure is set to inf
 
+    def __init__(self, trackers: List[str]):
+        self.trackers = trackers
+        self.feature_vecs_per_device = defaultdict(lambda: [self.FAR_AWAY_FEATURE_VALUE] * len(trackers))
+        self.last_measurements = deque()
+        self.tracker_name_to_idx = {name: i for i, name in enumerate(trackers)}
+        self.start_time = None
 
-def get_feature_value(rssi, tx_power):
+    @staticmethod
+    def _get_feature_value(rssi, tx_power):
         """Transforms rssi and tx power into a value between 0 and 1, where 0 is close and 1 is far away"""
         MIN_RSSI = -90
         MAX_TRANSFORMED_RSSI = 40
@@ -115,34 +125,48 @@ def get_feature_value(rssi, tx_power):
             v = 0
         return v / (-MIN_RSSI)
 
+    def add_measurement(self, new_measurement: BtleDeviceMeasurement):
+        if self.start_time is None:
+            self.start_time = new_measurement.time
 
-def training_data_from_df(df: pd.DataFrame, device: str):
+        self.last_measurements.append(new_measurement)
+        while len(self.last_measurements) > 0 and new_measurement.time - self.last_measurements[0].time > self.TIME_TO_DELETE_IF_NOT_SEEN:
+            self.last_measurements.popleft()
+
+        feature_vec = [self.FAR_AWAY_FEATURE_VALUE] * len(self.trackers)
+        for m in self.last_measurements: 
+            if m.device == new_measurement.device:
+                tracker_idx = self.tracker_name_to_idx[m.tracker]
+                feature_vec[tracker_idx] = self._get_feature_value(m.rssi, m.tx_power)
+        return feature_vec if new_measurement.time - self.start_time > self.MIN_TIME_UNTIL_PREDICTION else None
+    
+
+
+def training_data_from_df(df: pd.DataFrame, device_to_train: str):
     """Returns a feature matrix (num_measurement, num_trackers) and a label vector (both numeric) to be used in scikit learn"""
-    idx_to_tracker = dict(enumerate(df["tracker"].cat.categories))
-    tracker_to_idx = {v: k for k, v in idx_to_tracker.items()}
+    trackers = list(df["tracker"].cat.categories)
     idx_to_room = dict(enumerate(df["known_room"].cat.categories))
     room_to_idx = {v: k for k, v in idx_to_room.items()}
 
     last_known_room = None
-    start_time = None
-    current_feature = [FAR_AWAY_FEATURE_VALUE] * len(idx_to_tracker)
 
     features = []
     labels = []
     
+    feature_accumulator = RunningFeatureVector(trackers)
+
     # Feature vectors - rssi column for each room
     for i, row in df.iterrows():
         time, device, tracker, rssi, tx_power, known_room = row
-        if device != device:
+        m = BtleDeviceMeasurement(time, device, tracker, rssi, tx_power)
+        if device != device_to_train:
             continue
         if last_known_room != known_room:
-            start_time = time
+            feature_accumulator = RunningFeatureVector(trackers) # reset for new room
             last_known_room = known_room
-
-        tracker_idx = tracker_to_idx[tracker]
-        current_feature[tracker_idx] = get_feature_value(rssi, tx_power)
-        if time - start_time > 20:  # Wait 20secs to have measurements from all trackers
-            features.append(copy(current_feature))
+        feature_vec = feature_accumulator.add_measurement(m)
+        if feature_vec is not None:
+            features.append(feature_vec)
             labels.append(room_to_idx[known_room])
 
     return np.array(features), np.array(labels)
@@ -187,9 +211,8 @@ async def async_main(
     device_decoder: DeviceDecoder,
     training_data_logger: KnownRoomCsvLogger,
 ):
-    feature_vecs_per_device = defaultdict(lambda: [FAR_AWAY_FEATURE_VALUE] * len(trackers))
     current_rooms = defaultdict(lambda: "unknown")
-    tracker_name_to_idx = {name: i for i, name in enumerate(trackers)}
+    feature_accumulator = RunningFeatureVector(trackers)    
     async with aiomqtt.Client(
         hostname=mqtt_info.server, username=mqtt_info.username, password=mqtt_info.password
     ) as client:
@@ -216,22 +239,24 @@ async def async_main(
                     if m is not None:
                         logging.debug(f"Decoded Measurement {m}")
                         training_data_logger.report_measure(m)
-                        tracker_idx = tracker_name_to_idx[m.tracker]
-                        feature_vecs_per_device[m.device][tracker_idx] = get_feature_value(m.rssi, m.tx_power)
-                        if classifier is not None:
-                            room = classifier(m.device, feature_vecs_per_device[m.device])
+                        feature_vec =feature_accumulator.add_measurement(m)
+                        if feature_vec:
+                            feature_str={tracker : value for tracker, value in zip(trackers, feature_vec)} 
+                            logging.debug(f"Features: {feature_str}")
+                        if feature_vec is not None and classifier is not None:
+                            room = classifier(m.device, feature_vec)
                             if room != current_rooms[m.device]:
                                 logging.info(f"{m.device} moved room {current_rooms[m.device]} to {room}")
                                 current_rooms[m.device] = room
                             await client.publish(f"my_btmonitor/ml/{m.device}", room.encode())
 
 
-def get_classification_func(training_df: pd.DataFrame, log_classifier_scores=False):
+def get_classification_func(training_df: pd.DataFrame, log_classifier_scores=True):
     devices_to_track = list(training_df["device"].unique())
     classifiers = {}
     rooms = list(training_df["known_room"].dtype.categories)
     for device_to_track in devices_to_track:
-        features, labels = training_data_from_df(training_df, devices_to_track)
+        features, labels = training_data_from_df(training_df, device_to_track)
         clf = svm.SVC(kernel="rbf")
         logging.info(f"Computing cross validation score for {device_to_track}")
         if log_classifier_scores:
@@ -258,8 +283,8 @@ if __name__ == "__main__":
         "aa67542b82c0e05d65c27fb7e313aba5": "martins_apple_watch",
         "840e3892644c1ebd1594a9069c14ce0d": "martins_iphone",
     }
-
-    data_file = Path("training_data.csv")
+    script_path = os.path.dirname(os.path.realpath(__file__))
+    data_file = Path(script_path) / Path("training_data.csv")
     training_df = load_measurements_from_csv(data_file)
     classification_func = get_classification_func(training_df)
     training_data_logger = KnownRoomCsvLogger(data_file)
@@ -267,3 +292,4 @@ if __name__ == "__main__":
     trackers = list(training_df["tracker"].cat.categories)
     devices = list(training_df['device'].cat.categories)
     asyncio.run(async_main(mqtt_info, trackers, devices, classification_func, device_decoder, training_data_logger))
+