1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
|
import os
import cPickle
from blocks.initialization import IsotropicGaussian, Constant
from blocks.algorithms import Momentum
import data
from model.dest_mlp_tgtcls import Model, Stream
n_begin_end_pts = 5 # how many points we consider at the beginning and end of the known trajectory
with open(os.path.join(data.path, 'arrival-clusters.pkl')) as f: tgtcls = cPickle.load(f)
dim_embeddings = [
('origin_call', data.origin_call_train_size, 10),
('origin_stand', data.stands_size, 10),
('week_of_year', 52, 10),
('day_of_week', 7, 10),
('qhour_of_day', 24 * 4, 10),
('day_type', 3, 10),
('taxi_id', 448, 10),
]
dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
dim_hidden = [100, 100]
dim_output = tgtcls.shape[0]
embed_weights_init = IsotropicGaussian(0.01)
mlp_weights_init = IsotropicGaussian(0.1)
mlp_biases_init = Constant(0.01)
step_rule = Momentum(learning_rate=0.01, momentum=0.9)
batch_size = 200
max_splits = 100
|
