memory_network_adeb.py (1454B)
1 from blocks.initialization import IsotropicGaussian, Constant 2 from blocks.algorithms import AdaDelta, CompositeRule, GradientDescent, RemoveNotFinite, StepRule, Momentum 3 4 import data 5 from model.memory_network import Model, Stream 6 7 8 n_begin_end_pts = 5 # how many points we consider at the beginning and end of the known trajectory 9 10 dim_embeddings = [ 11 ('origin_call', data.origin_call_train_size, 10), 12 ('origin_stand', data.stands_size, 10), 13 ('week_of_year', 52, 10), 14 ('day_of_week', 7, 10), 15 ('qhour_of_day', 24 * 4, 10), 16 ('day_type', 3, 10), 17 ] 18 19 20 class MLPConfig(object): 21 __slots__ = ('dim_input', 'dim_hidden', 'dim_output', 'weights_init', 'biases_init') 22 23 prefix_encoder = MLPConfig() 24 prefix_encoder.dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings) 25 prefix_encoder.dim_hidden = [100, 100] 26 prefix_encoder.weights_init = IsotropicGaussian(0.001) 27 prefix_encoder.biases_init = Constant(0.0001) 28 29 candidate_encoder = MLPConfig() 30 candidate_encoder.dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings) 31 candidate_encoder.dim_hidden = [100, 100] 32 candidate_encoder.weights_init = IsotropicGaussian(0.001) 33 candidate_encoder.biases_init = Constant(0.0001) 34 35 36 embed_weights_init = IsotropicGaussian(0.001) 37 38 step_rule = Momentum(learning_rate=0.001, momentum=0.9) 39 batch_size = 32 40 41 max_splits = 1 42 num_cuts = 1000 43 44 train_candidate_size = 1000 45 valid_candidate_size = 10000 46 47 load_model = False