Initial commit

4 files changed, 112 insertions, 0 deletions

diff --git a/config/__init__.py b/config/__init__.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/config/__init__.py
diff --git a/config/deep_bidir_lstm_2x128.py b/config/deep_bidir_lstm_2x128.py
new file mode 100644
index 0000000..f07f43f
--- /dev/null
+++ b/config/deep_bidir_lstm_2x128.py
@@ -0,0 +1,37 @@
+from blocks.algorithms import BasicMomentum, AdaDelta, RMSProp, Adam, CompositeRule, StepClipping
+from blocks.initialization import IsotropicGaussian, Constant
+from blocks.bricks import Tanh
+
+from model.deep_bidir_lstm import Model
+
+
+batch_size = 32
+sort_batch_count = 20
+
+shuffle_questions = True
+shuffle_entities = True
+
+concat_ctx_and_question = True
+concat_question_before = True		## should not matter for bidirectionnal network
+
+embed_size = 200
+
+lstm_size = [128, 128]
+skip_connections = True
+
+n_entities = 550
+out_mlp_hidden = []
+out_mlp_activations = []
+
+step_rule = CompositeRule([RMSProp(decay_rate=0.95, learning_rate=5e-5),
+                           BasicMomentum(momentum=0.9)])
+
+dropout = 0.1
+w_noise = 0.05
+
+valid_freq = 1000
+save_freq = 1000
+print_freq = 100
+
+weights_init = IsotropicGaussian(0.01)
+biases_init = Constant(0.)
diff --git a/config/deepmind_attentive_reader.py b/config/deepmind_attentive_reader.py
new file mode 100644
index 0000000..84a6cf0
--- /dev/null
+++ b/config/deepmind_attentive_reader.py
@@ -0,0 +1,42 @@
+from blocks.bricks import Tanh
+from blocks.algorithms import BasicMomentum, AdaDelta, RMSProp, Adam, CompositeRule, StepClipping, Momentum
+from blocks.initialization import IsotropicGaussian, Constant
+
+from model.attentive_reader import Model
+
+
+batch_size = 32
+sort_batch_count = 20
+
+shuffle_questions = True
+
+concat_ctx_and_question = False
+
+n_entities = 550
+embed_size = 200
+
+ctx_lstm_size = [256]
+ctx_skip_connections = True
+
+question_lstm_size = [256]
+question_skip_connections = True
+
+attention_mlp_hidden = [100]
+attention_mlp_activations = [Tanh()]
+
+out_mlp_hidden = []
+out_mlp_activations = []
+
+step_rule = CompositeRule([RMSProp(decay_rate=0.95, learning_rate=5e-5),
+                           BasicMomentum(momentum=0.9)])
+
+dropout = 0.2
+w_noise = 0.
+
+valid_freq = 1000
+save_freq = 1000
+print_freq = 100
+
+weights_init = IsotropicGaussian(0.01)
+biases_init = Constant(0.)
+
diff --git a/config/deepmind_deep_lstm.py b/config/deepmind_deep_lstm.py
new file mode 100644
index 0000000..10b5c9b
--- /dev/null
+++ b/config/deepmind_deep_lstm.py
@@ -0,0 +1,33 @@
+from blocks.algorithms import BasicMomentum, AdaDelta, RMSProp, Adam, CompositeRule, StepClipping
+from blocks.initialization import IsotropicGaussian, Constant
+
+from model.deep_lstm import Model
+
+
+batch_size = 32
+sort_batch_count = 20
+
+shuffle_questions = True
+
+concat_ctx_and_question = True
+concat_question_before = True
+
+embed_size = 200
+
+lstm_size = [256, 256]
+skip_connections = True
+
+out_mlp_hidden = []
+out_mlp_activations = []
+
+step_rule = CompositeRule([RMSProp(decay_rate=0.95, learning_rate=1e-4),
+                           BasicMomentum(momentum=0.9)])
+
+dropout = 0.1
+
+valid_freq = 1000
+save_freq = 1000
+print_freq = 100
+
+weights_init = IsotropicGaussian(0.01)
+biases_init = Constant(0.)