From 30faf44a08edcc2075362c4633f6b1d291944cd3 Mon Sep 17 00:00:00 2001
From: Alex Auvolat <alex@adnab.me>
Date: Tue, 29 Mar 2016 12:27:35 +0200
Subject: This HPC stuff doesn't work very well.

---
 config/hpc-gru-0.py  | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++
 config/hpc-gru-1.py  | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++
 config/hpc-lstm-0.py | 43 ++++++++++++++++++++++++++++++++++++++++++
 config/hpc-lstm-2.py | 46 +++++++++++++++++++++++++++++++++++++++++++++
 config/hpc-lstm-3.py | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 246 insertions(+)
 create mode 100644 config/hpc-gru-0.py
 create mode 100644 config/hpc-gru-1.py
 create mode 100644 config/hpc-lstm-0.py
 create mode 100644 config/hpc-lstm-2.py
 create mode 100644 config/hpc-lstm-3.py

(limited to 'config')

diff --git a/config/hpc-gru-0.py b/config/hpc-gru-0.py
new file mode 100644
index 0000000..ab58a86
--- /dev/null
+++ b/config/hpc-gru-0.py
@@ -0,0 +1,52 @@
+import numpy
+from numpy.random import RandomState
+
+from blocks.algorithms import AdaDelta, Momentum, RMSProp, CompositeRule, BasicMomentum
+from blocks.bricks import Tanh, Rectifier
+from blocks.initialization import IsotropicGaussian, Constant
+
+from model.hpc_gru import Model
+
+dataset = 'data/logcompil-2016-03-07.txt'
+
+io_dim = 256
+repr_dim = 64
+embedding_matrix = (RandomState(42).binomial(1, 10./repr_dim, ((io_dim, repr_dim)))
+                   -RandomState(123).binomial(1, 10./repr_dim, ((io_dim, repr_dim))))
+
+# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
+# divided in chunks of lengh 'seq_div_size'
+num_seqs = 100
+seq_len = 2000
+seq_div_size = 100
+
+hidden_dims = [128, 384, 1024]
+cost_factors = [1., 1., 1.]
+hidden_q = [0.5, 0.5, 0.5]
+activation_function = Tanh()
+
+out_hidden = [512]
+out_hidden_act = [Tanh]
+
+weight_noise = 0
+
+step_rule = AdaDelta()
+#step_rule = CompositeRule([RMSProp(learning_rate=0.01),
+#						   BasicMomentum(momentum=0.9)])
+#step_rule = Momentum(learning_rate=.1, momentum=0.9)
+
+weights_init = IsotropicGaussian(0.1)
+biases_init = Constant(0.)
+
+# parameter saving freq (number of batches)
+monitor_freq = 100
+save_freq = 100
+
+# used for sample generation and IRC mode
+sample_temperature = 0.5 #0.7
+
+# do we want to generate samples at times during training?
+sample_len = 1000
+sample_freq = 100
+sample_init = '\nalex\ttu crois?\n'
+
diff --git a/config/hpc-gru-1.py b/config/hpc-gru-1.py
new file mode 100644
index 0000000..b59b025
--- /dev/null
+++ b/config/hpc-gru-1.py
@@ -0,0 +1,52 @@
+import numpy
+from numpy.random import RandomState
+
+from blocks.algorithms import AdaDelta, Momentum, RMSProp, CompositeRule, BasicMomentum, Adam
+from blocks.bricks import Tanh, Rectifier
+from blocks.initialization import IsotropicGaussian, Constant
+
+from model.hpc_gru import Model
+
+dataset = 'data/logcompil-2016-03-07.txt'
+
+io_dim = 256
+repr_dim = 128
+embedding_matrix = (RandomState(42).binomial(1, 0.1, ((io_dim, repr_dim)))
+                   -RandomState(123).binomial(1, 0.1, ((io_dim, repr_dim))))
+
+# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
+# divided in chunks of lengh 'seq_div_size'
+num_seqs = 20
+seq_len = 5000
+seq_div_size = 50
+
+hidden_dims = [128, 192, 256, 512]
+cost_factors = [1., 1., 1., 1.]
+hidden_q = [0.5, 0.5, 0.5, 0.5]
+activation_function = Tanh()
+
+out_hidden = [512]
+out_hidden_act = [Rectifier]
+
+weight_noise = 0.05
+
+step_rule = Adam()
+#step_rule = CompositeRule([RMSProp(learning_rate=0.01),
+#						   BasicMomentum(momentum=0.9)])
+#step_rule = Momentum(learning_rate=.1, momentum=0.9)
+
+weights_init = IsotropicGaussian(0.1)
+biases_init = Constant(0.01)
+
+# parameter saving freq (number of batches)
+monitor_freq = 500
+save_freq = monitor_freq
+
+# used for sample generation and IRC mode
+sample_temperature = 0.5 #0.7
+
+# do we want to generate samples at times during training?
+sample_len = 1000
+sample_freq = monitor_freq
+sample_init = '\nalex\ttu crois?\n'
+
diff --git a/config/hpc-lstm-0.py b/config/hpc-lstm-0.py
new file mode 100644
index 0000000..afb6471
--- /dev/null
+++ b/config/hpc-lstm-0.py
@@ -0,0 +1,43 @@
+import numpy
+from numpy.random import RandomState
+
+from blocks.algorithms import AdaDelta, Momentum
+from blocks.bricks import Tanh, Rectifier
+
+from model.hpc_lstm import Model
+
+dataset = 'data/logcompil-2016-03-07.txt'
+
+io_dim = 256
+repr_dim = 256
+embedding_matrix = numpy.eye(io_dim)
+
+# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
+# divided in chunks of lengh 'seq_div_size'
+num_seqs = 100
+seq_len = 2000
+seq_div_size = 100
+
+hidden_dims = [128, 128, 256, 512]
+cost_factors = [1., 1., 1., 1.]
+hidden_q = [0.1, 0.15, 0.22, 0.33]
+activation_function = Tanh()
+
+out_hidden = [512]
+out_hidden_act = [Rectifier]
+
+step_rule = AdaDelta()
+#step_rule = Momentum(learning_rate=0.0001, momentum=0.99)
+
+# parameter saving freq (number of batches)
+monitor_freq = 10
+save_freq = 100
+
+# used for sample generation and IRC mode
+sample_temperature = 0.7 #0.5
+
+# do we want to generate samples at times during training?
+sample_len = 1000
+sample_freq = 100
+sample_init = '\nalex\ttu crois?\n'
+
diff --git a/config/hpc-lstm-2.py b/config/hpc-lstm-2.py
new file mode 100644
index 0000000..aaed80e
--- /dev/null
+++ b/config/hpc-lstm-2.py
@@ -0,0 +1,46 @@
+import numpy
+from numpy.random import RandomState
+
+from blocks.algorithms import AdaDelta, Momentum, RMSProp, CompositeRule, BasicMomentum
+from blocks.bricks import Tanh, Rectifier
+
+from model.hpc_lstm import Model
+
+dataset = 'data/logcompil-2016-03-07.txt'
+
+io_dim = 256
+repr_dim = 64
+embedding_matrix = (RandomState(42).binomial(1, 10./repr_dim, ((io_dim, repr_dim)))
+                   -RandomState(123).binomial(1, 10./repr_dim, ((io_dim, repr_dim))))
+
+# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
+# divided in chunks of lengh 'seq_div_size'
+num_seqs = 100
+seq_len = 2000
+seq_div_size = 100
+
+hidden_dims = [64, 256, 1024]
+cost_factors = [1., 1., 1.]
+hidden_q = [0.5, 0.5, 0.5]
+activation_function = Tanh()
+
+out_hidden = [512]
+out_hidden_act = [Rectifier]
+
+step_rule = AdaDelta()
+#step_rule = CompositeRule([RMSProp(learning_rate=0.01),
+#						   BasicMomentum(momentum=0.9)])
+#step_rule = Momentum(learning_rate=.1, momentum=0.9)
+
+# parameter saving freq (number of batches)
+monitor_freq = 100
+save_freq = 100
+
+# used for sample generation and IRC mode
+sample_temperature = 0.7 #0.5
+
+# do we want to generate samples at times during training?
+sample_len = 1000
+sample_freq = 100
+sample_init = '\nalex\ttu crois?\n'
+
diff --git a/config/hpc-lstm-3.py b/config/hpc-lstm-3.py
new file mode 100644
index 0000000..fa0f77e
--- /dev/null
+++ b/config/hpc-lstm-3.py
@@ -0,0 +1,53 @@
+
+import numpy
+from numpy.random import RandomState
+
+from blocks.algorithms import AdaDelta, Momentum, RMSProp, CompositeRule, BasicMomentum, Adam
+from blocks.bricks import Tanh, Rectifier
+from blocks.initialization import IsotropicGaussian, Constant
+
+from model.hpc_lstm import Model
+
+dataset = 'data/logcompil-2016-03-07.txt'
+
+io_dim = 256
+repr_dim = 128
+embedding_matrix = (RandomState(42).binomial(1, 0.1, ((io_dim, repr_dim)))
+                   -RandomState(123).binomial(1, 0.1, ((io_dim, repr_dim))))
+
+# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
+# divided in chunks of lengh 'seq_div_size'
+num_seqs = 20
+seq_len = 5000
+seq_div_size = 50
+
+hidden_dims = [128, 192, 256, 512]
+cost_factors = [1., 1., 1., 1.]
+hidden_q = [0.5, 0.5, 0.5, 0.5]
+activation_function = Tanh()
+
+out_hidden = [512]
+out_hidden_act = [Rectifier]
+
+weight_noise = 0.05
+
+step_rule = Adam()
+#step_rule = CompositeRule([RMSProp(learning_rate=0.01),
+#						   BasicMomentum(momentum=0.9)])
+#step_rule = Momentum(learning_rate=.1, momentum=0.9)
+
+weights_init = IsotropicGaussian(0.1)
+biases_init = Constant(0.01)
+
+# parameter saving freq (number of batches)
+monitor_freq = 500
+save_freq = monitor_freq
+
+# used for sample generation and IRC mode
+sample_temperature = 0.5 #0.7
+
+# do we want to generate samples at times during training?
+sample_len = 1000
+sample_freq = monitor_freq
+sample_init = '\nalex\ttu crois?\n'
+
-- 
cgit v1.2.3