Add memory_net_bidir_mom

2 files changed, 60 insertions, 2 deletions

diff --git a/config/bidirectional_tgtcls_1_momentum.py b/config/bidirectional_tgtcls_1_momentum.py
index b286e0f..65ad021 100644
--- a/config/bidirectional_tgtcls_1_momentum.py
+++ b/config/bidirectional_tgtcls_1_momentum.py
@@ -27,12 +27,12 @@ embed_weights_init = IsotropicGaussian(0.01)
 weights_init = IsotropicGaussian(0.1) 
 biases_init = Constant(0.01)
 
-batch_size = 100
+batch_size = 300
 batch_sort_size = 20
 
 max_splits = 100
 
 # monitor_freq = 10000 # temporary, for finding good learning rate
 
-step_rule= Momentum(learning_rate=0.1, momentum=0.9)
+step_rule= Momentum(learning_rate=0.01, momentum=0.9)
 
diff --git a/config/memory_network_bidir_momentum.py b/config/memory_network_bidir_momentum.py
new file mode 100644
index 0000000..e5863ae
--- /dev/null
+++ b/config/memory_network_bidir_momentum.py
@@ -0,0 +1,58 @@
+from blocks.initialization import IsotropicGaussian, Constant
+from blocks.algorithms import Momentum
+
+from blocks.bricks import Tanh
+
+import data
+from model.memory_network_bidir import Model, Stream
+
+
+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),
+]
+
+embed_weights_init = IsotropicGaussian(0.001)
+
+
+class RNNConfig(object):
+    __slots__ = ('rec_state_dim', 'dim_embeddings', 'embed_weights_init',
+                 'dim_hidden', 'weights_init', 'biases_init')
+
+prefix_encoder = RNNConfig()
+prefix_encoder.dim_embeddings = dim_embeddings
+prefix_encoder.embed_weights_init = embed_weights_init
+prefix_encoder.rec_state_dim = 100
+prefix_encoder.dim_hidden = [100, 100]
+prefix_encoder.weights_init = IsotropicGaussian(0.01)
+prefix_encoder.biases_init = Constant(0.001)
+
+candidate_encoder = RNNConfig()
+candidate_encoder.dim_embeddings = dim_embeddings
+candidate_encoder.embed_weights_init = embed_weights_init
+candidate_encoder.rec_state_dim = 100
+candidate_encoder.dim_hidden = [100, 100]
+candidate_encoder.weights_init = IsotropicGaussian(0.01)
+candidate_encoder.biases_init = Constant(0.001)
+
+representation_size = 100
+representation_activation = Tanh
+
+normalize_representation = True
+
+
+batch_size = 32
+batch_sort_size = 20
+
+max_splits = 100
+num_cuts = 1000
+
+train_candidate_size = 1000
+valid_candidate_size = 1000
+test_candidate_size = 1000
+
+step_rule = Momentum(learning_rate=0.01, momentum=0.9)