Airport embedding in keras

airport-embedding.py

# %matplotlib inline
import pandas as pd
import numpy as np
import keras
import tensorflow as tf
import os

from keras.models import Model
from keras.layers import Input, Embedding, merge
from keras.layers.core import Flatten, Dense, Dropout, Lambda
from keras.optimizers import Adam
from keras.regularizers import l2

# https://www.transtats.bts.gov/DL_SelectFields.asp?Table_ID=236
# download the flight data from the above url and unzip/rename.
# select the following fields: Origin, Dest, Actual Elapsed time, distance

flights = pd.read_csv('flights_1.csv')
flights = flights[np.isfinite(flights['ACTUAL_ELAPSED_TIME'])] #remove empty and nan values for elapsed time.
# flights = flights.drop('ACTUAL_ELAPSED_TIME', 1)
flights = flights.drop('Unnamed: 4', 1)
# flights = flights.drop_duplicates()
# flights = flights[np.isfinite(flights['DISTANCE'])] #remove empty and nan values for elapsed time.
flights = flights.drop('DISTANCE', 1)
flights.head()

n_airports = flights.ORIGIN.nunique()
n_airports

airports = flights.ORIGIN.unique()
airports.sort()
airports_d = airports
# airports

airports2idx = {o:i for i,o in enumerate(airports)}
# airports2idx

#convert all airports to index values in the flights data.
flights.ORIGIN = flights.ORIGIN.apply(lambda x: airports2idx[x])
flights.DEST = flights.DEST.apply(lambda x: airports2idx[x])
# flights

# seperate data into train and test values.
msk = np.random.rand(len(flights)) < 0.8
trn = flights[msk]
val = flights[~msk]

def embedding_input(name, n_in, n_out, reg):
    inp = Input(shape=(1,), dtype='int64', name=name)
    return inp, Embedding(n_in, n_out, input_length=1, W_regularizer=l2(reg))(inp)


n_factors = 5
origin_airports, o = embedding_input('airports', n_airports, n_factors, 1e-4)
dest_airports, d = embedding_input('airports_d', n_airports, n_factors, 1e-4)

x = merge([o, d], mode='concat')
x = Flatten()(x)
# x = Dropout(0.3)(x)
x = Dense(70, activation='relu')(x)
# x = Dropout(0.7)(x)
x = Dense(1)(x)
nn = Model([origin_airports, dest_airports], x)
nn.compile(Adam(0.001), loss='mse')

nn.fit([trn.ORIGIN, trn.DEST], trn.ACTUAL_ELAPSED_TIME, batch_size=64, nb_epoch=8, 
          validation_data=([val.ORIGIN, val.DEST], val.ACTUAL_ELAPSED_TIME))
nn.optimizer.lr=0.1
nn.fit([trn.ORIGIN, trn.DEST], trn.ACTUAL_ELAPSED_TIME, batch_size=64, nb_epoch=20, 
          validation_data=([val.ORIGIN, val.DEST], val.ACTUAL_ELAPSED_TIME))

saver = tf.train.Saver()

LOG_DIR = '/home/surya/tensorboard_log'
os.makedirs(LOG_DIR, exist_ok=True)
saver.save(keras.backend.get_session(), os.path.join(LOG_DIR, "model.ckpt"))
#start the tensorboard with `tensorboard --logdir=~/tensorboard_log` and explore the embedding tab (look for 294x2)

def estimate_time(src_airport, dest_airport):
    return nn.predict([np.array([airports2idx[src_airport]]),np.array([airports2idx[dest_airport]])])

estimate_time('ABE', 'ATL')