commit 1e8da55c32746e7bf898717c032144b056256d3c
parent ccc92aaa1eb24fdf00afb254f56982b1bac406d1
Author: Alex Auvolat <alex.auvolat@ens.fr>
Date: Tue, 14 Jul 2015 15:43:13 -0400
Add memory network config files (not sure how usefull they are)
Diffstat:
2 files changed, 114 insertions(+), 0 deletions(-)
diff --git a/config/memory_network_2.py b/config/memory_network_2.py
@@ -0,0 +1,57 @@
+from blocks import roles
+from blocks.bricks import Rectifier, Tanh, Logistic
+from blocks.filter import VariableFilter
+from blocks.initialization import IsotropicGaussian, Constant
+
+import data
+from model.memory_network import Model, Stream
+
+
+n_begin_end_pts = 10 # how many points we consider at the beginning and end of the known trajectory
+
+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),
+]
+
+
+class MLPConfig(object):
+ __slots__ = ('dim_input', 'dim_hidden', 'dim_output', 'weights_init', 'biases_init')
+
+prefix_encoder = MLPConfig()
+prefix_encoder.dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+prefix_encoder.dim_hidden = [1000, 1000]
+prefix_encoder.weights_init = IsotropicGaussian(0.01)
+prefix_encoder.biases_init = Constant(0.001)
+
+candidate_encoder = MLPConfig()
+candidate_encoder.dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+candidate_encoder.dim_hidden = [1000, 1000]
+candidate_encoder.weights_init = IsotropicGaussian(0.01)
+candidate_encoder.biases_init = Constant(0.001)
+
+representation_size = 1000
+representation_activation = Tanh
+normalize_representation = True
+
+embed_weights_init = IsotropicGaussian(0.001)
+
+dropout = 0.5
+dropout_inputs = VariableFilter(bricks=[Rectifier], name='output')
+
+noise = 0.01
+noise_inputs = VariableFilter(roles=[roles.PARAMETER])
+
+batch_size = 512
+
+valid_set = 'cuts/test_times_0'
+max_splits = 1
+num_cuts = 1000
+
+train_candidate_size = 10000
+valid_candidate_size = 20000
+
diff --git a/config/memory_network_3.py b/config/memory_network_3.py
@@ -0,0 +1,57 @@
+from blocks import roles
+from blocks.bricks import Rectifier, Tanh, Logistic
+from blocks.filter import VariableFilter
+from blocks.initialization import IsotropicGaussian, Constant
+
+import data
+from model.memory_network import Model, Stream
+
+
+n_begin_end_pts = 10 # how many points we consider at the beginning and end of the known trajectory
+
+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),
+]
+
+
+class MLPConfig(object):
+ __slots__ = ('dim_input', 'dim_hidden', 'dim_output', 'weights_init', 'biases_init')
+
+prefix_encoder = MLPConfig()
+prefix_encoder.dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+prefix_encoder.dim_hidden = [200, 200, 200]
+prefix_encoder.weights_init = IsotropicGaussian(0.01)
+prefix_encoder.biases_init = Constant(0.001)
+
+candidate_encoder = MLPConfig()
+candidate_encoder.dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+candidate_encoder.dim_hidden = [200, 200, 200]
+candidate_encoder.weights_init = IsotropicGaussian(0.01)
+candidate_encoder.biases_init = Constant(0.001)
+
+representation_size = 500
+representation_activation = Tanh
+normalize_representation = True
+
+embed_weights_init = IsotropicGaussian(0.001)
+
+dropout = 0.5
+dropout_inputs = VariableFilter(bricks=[Rectifier], name='output')
+
+noise = 0.01
+noise_inputs = VariableFilter(roles=[roles.PARAMETER])
+
+batch_size = 512
+
+valid_set = 'cuts/test_times_0'
+max_splits = 1
+num_cuts = 1000
+
+train_candidate_size = 10000
+valid_candidate_size = 20000
+
