FLasso: time_of_week_feature for Kölapp, Triage, Klar and Läkare kris…

…toferB/SP#89

models/FLasso/time_of_week_feature.py

@@ -0,0 +1,53 @@
+# -*- coding: utf-8 -*-
+
+import pandas as pd
+import tensorflow as tf
+import numpy as np
+
+
+pdframe = pd.read_csv("FLasso2017-08-14T15:29:28.093+02:00.csv")
+
+WEEK_IN_SECONDS = 3600 * 24 * 7
+BIN_WIDTH = 600 # in seconds
+N_BINS = WEEK_IN_SECONDS / BIN_WIDTH
+
+def _incidence_bins(event_title):
+    incidence_bins = {i: [] for i in range(N_BINS)}
+    epochsecond = np.array(pdframe["epochseconds"].get_values(), dtype=np.int32)
+    next_hour_incidence = np.array(pdframe["NextHour" + event_title].get_values(), dtype=np.float32)
+    for i in range(len(epochsecond)):
+        second_in_week = epochsecond[i] % WEEK_IN_SECONDS
+        bin_number = second_in_week / BIN_WIDTH
+        incidence_bins[bin_number] = incidence_bins[bin_number] + [next_hour_incidence[i]]
+    return incidence_bins
+
+def to_time_of_week_feature_np(epochsecond, event_title):
+    incidence_bins = _incidence_bins(event_title)
+    incidence_mean_bins = np.array([np.mean(incidence_bins[i]) for i in range(N_BINS)], dtype=np.float32)
+    second_in_week = epochsecond % WEEK_IN_SECONDS
+    return incidence_mean_bins[second_in_week / BIN_WIDTH]
+
+def to_time_of_week_feature(epochsecond, event_title):
+    incidence_bins = _incidence_bins(event_title)
+    incidence_mean_bins = tf.constant([np.mean(incidence_bins[i]) for i in range(N_BINS)], dtype=tf.float32)
+    second_in_week = epochsecond % WEEK_IN_SECONDS
+    return tf.map_fn(lambda s: incidence_mean_bins[s / BIN_WIDTH], second_in_week, dtype=tf.float32)
+
+if __name__ == "__main__":
+    sess = tf.Session()
+
+    pdframe = pd.read_csv("FLasso2017-08-14T15:29:28.093+02:00.csv")
+
+    for event_title in ["Kölapp", "Triage", "Klar", "Läkare"]:
+        print "time-of-week mse for NextHour" + event_title + ":"
+
+        epochsecond = tf.constant(pdframe["epochseconds"].get_values(), dtype=tf.int32)
+        next_hour_incidence = tf.constant(pdframe["NextHour" + event_title].get_values(), dtype=tf.float32)
+        pred = to_time_of_week_feature(epochsecond, event_title)
+        mse_minutes = tf.losses.mean_squared_error(next_hour_incidence, pred)
+        print sess.run(mse_minutes)
+
+        epochsecond_np = np.array(pdframe["epochseconds"].get_values(), dtype=np.int32)
+        next_hour_incidence = np.array(pdframe["NextHour" + event_title].get_values(), dtype=np.float32)
+        pred_np = to_time_of_week_feature_np(epochsecond_np, event_title)
+        print np.mean((pred_np - next_hour_incidence) ** 2)