path: root/config/dest_simple_mlp_tgtcls_0_cs.py

import os
import cPickle

from blocks.initialization import IsotropicGaussian, Constant

import data
from model.dest_simple_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)
]

dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
dim_hidden = []
dim_output = tgtcls.shape[0]

embed_weights_init = IsotropicGaussian(0.001)
mlp_weights_init = IsotropicGaussian(0.01)
mlp_biases_init = Constant(0.001)

learning_rate = 0.0001
momentum = 0.99
batch_size = 32

valid_set = 'cuts/test_times_0'
max_splits = 100