From 9a779f7328a712a20dd393bdf32c6a84bf9fbe52 Mon Sep 17 00:00:00 2001
From: Alex Auvolat <alex.auvolat@ens.fr>
Date: Thu, 21 May 2015 10:46:05 -0400
Subject: Model changes

---
 .../joint_simple_mlp_tgtcls_111_cswdtx_bigger.py   | 56 ++++++++++++++++++++
 ..._simple_mlp_tgtcls_111_cswdtx_bigger_dropout.py | 59 ++++++++++++++++++++++
 config/joint_simple_mlp_tgtcls_1_cswdtx_bigger.py  | 55 ++++++++++++++++++++
 model/joint_simple_mlp_tgtcls.py                   |  2 +
 train.py                                           |  6 +--
 5 files changed, 175 insertions(+), 3 deletions(-)
 create mode 100644 config/joint_simple_mlp_tgtcls_111_cswdtx_bigger.py
 create mode 100644 config/joint_simple_mlp_tgtcls_111_cswdtx_bigger_dropout.py
 create mode 100644 config/joint_simple_mlp_tgtcls_1_cswdtx_bigger.py

diff --git a/config/joint_simple_mlp_tgtcls_111_cswdtx_bigger.py b/config/joint_simple_mlp_tgtcls_111_cswdtx_bigger.py
new file mode 100644
index 0000000..8adb6e7
--- /dev/null
+++ b/config/joint_simple_mlp_tgtcls_111_cswdtx_bigger.py
@@ -0,0 +1,56 @@
+import cPickle
+
+import model.joint_simple_mlp_tgtcls as model
+
+from blocks.initialization import IsotropicGaussian, Constant
+
+import data
+
+n_begin_end_pts = 10     # how many points we consider at the beginning and end of the known trajectory
+n_end_pts = 10
+
+n_valid = 1000
+
+with open("%s/arrival-clusters.pkl" % data.path) as f:
+    dest_tgtcls = cPickle.load(f)
+
+# generate target classes for time prediction as a Fibonacci sequence
+time_tgtcls = [1, 2]
+for i in range(21):
+    time_tgtcls.append(time_tgtcls[-1] + time_tgtcls[-2])
+
+dim_embeddings = [
+    ('origin_call', data.origin_call_size+1, 15),
+    ('origin_stand', data.stands_size+1, 10),
+    ('week_of_year', 52, 10),
+    ('day_of_week', 7, 10),
+    ('qhour_of_day', 24 * 4, 10),
+    ('day_type', 3, 10),
+    ('taxi_id', 448, 10),
+]
+
+# Common network part
+dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+dim_hidden = [1000]
+
+# Destination prediction part
+dim_hidden_dest = [400]
+dim_output_dest = dest_tgtcls.shape[0]
+
+# Time prediction part
+dim_hidden_time = [400]
+dim_output_time = len(time_tgtcls)
+
+# Cost ratio between distance cost and time cost
+time_cost_factor = 4
+
+embed_weights_init = IsotropicGaussian(0.01)
+mlp_weights_init = IsotropicGaussian(0.1)
+mlp_biases_init = Constant(0.01)
+
+learning_rate = 0.000001
+momentum = 0.99
+batch_size = 200
+
+valid_set = 'cuts/test_times_0'
+
diff --git a/config/joint_simple_mlp_tgtcls_111_cswdtx_bigger_dropout.py b/config/joint_simple_mlp_tgtcls_111_cswdtx_bigger_dropout.py
new file mode 100644
index 0000000..02c8bd8
--- /dev/null
+++ b/config/joint_simple_mlp_tgtcls_111_cswdtx_bigger_dropout.py
@@ -0,0 +1,59 @@
+import cPickle
+
+import model.joint_simple_mlp_tgtcls as model
+
+from blocks.initialization import IsotropicGaussian, Constant
+
+import data
+
+n_begin_end_pts = 10     # how many points we consider at the beginning and end of the known trajectory
+n_end_pts = 10
+
+n_valid = 1000
+
+with open("%s/arrival-clusters.pkl" % data.path) as f:
+    dest_tgtcls = cPickle.load(f)
+
+# generate target classes for time prediction as a Fibonacci sequence
+time_tgtcls = [1, 2]
+for i in range(21):
+    time_tgtcls.append(time_tgtcls[-1] + time_tgtcls[-2])
+
+dim_embeddings = [
+    ('origin_call', data.origin_call_size+1, 15),
+    ('origin_stand', data.stands_size+1, 10),
+    ('week_of_year', 52, 10),
+    ('day_of_week', 7, 10),
+    ('qhour_of_day', 24 * 4, 10),
+    ('day_type', 3, 10),
+    ('taxi_id', 448, 10),
+]
+
+# Common network part
+dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+dim_hidden = [5000]
+
+# Destination prediction part
+dim_hidden_dest = [1000]
+dim_output_dest = dest_tgtcls.shape[0]
+
+# Time prediction part
+dim_hidden_time = [500]
+dim_output_time = len(time_tgtcls)
+
+# Cost ratio between distance cost and time cost
+time_cost_factor = 4
+
+embed_weights_init = IsotropicGaussian(0.01)
+mlp_weights_init = IsotropicGaussian(0.1)
+mlp_biases_init = Constant(0.01)
+
+# apply_dropout = True
+# dropout_p = 0.5
+
+learning_rate = 0.001
+momentum = 0.9
+batch_size = 200
+
+valid_set = 'cuts/test_times_0'
+
diff --git a/config/joint_simple_mlp_tgtcls_1_cswdtx_bigger.py b/config/joint_simple_mlp_tgtcls_1_cswdtx_bigger.py
new file mode 100644
index 0000000..995f858
--- /dev/null
+++ b/config/joint_simple_mlp_tgtcls_1_cswdtx_bigger.py
@@ -0,0 +1,55 @@
+import cPickle
+
+import model.joint_simple_mlp_tgtcls as model
+
+from blocks.initialization import IsotropicGaussian, Constant
+
+import data
+
+n_begin_end_pts = 7     # how many points we consider at the beginning and end of the known trajectory
+n_end_pts = 7
+
+n_valid = 1000
+
+with open("%s/arrival-clusters.pkl" % data.path) as f:
+    dest_tgtcls = cPickle.load(f)
+
+# generate target classes for time prediction as a Fibonacci sequence
+time_tgtcls = [1, 2]
+for i in range(21):
+    time_tgtcls.append(time_tgtcls[-1] + time_tgtcls[-2])
+
+dim_embeddings = [
+    ('origin_call', data.origin_call_size+1, 15),
+    ('origin_stand', data.stands_size+1, 10),
+    ('week_of_year', 52, 10),
+    ('day_of_week', 7, 10),
+    ('qhour_of_day', 24 * 4, 10),
+    ('day_type', 3, 10),
+    ('taxi_id', 448, 10),
+]
+
+# Common network part
+dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
+dim_hidden = [5000]
+
+# Destination prediction part
+dim_hidden_dest = []
+dim_output_dest = dest_tgtcls.shape[0]
+
+# Time prediction part
+dim_hidden_time = []
+dim_output_time = len(time_tgtcls)
+
+# Cost ratio between distance cost and time cost
+time_cost_factor = 4
+
+embed_weights_init = IsotropicGaussian(0.01)
+mlp_weights_init = IsotropicGaussian(0.1)
+mlp_biases_init = Constant(0.01)
+
+learning_rate = 0.0001
+momentum = 0.99
+batch_size = 200
+
+valid_set = 'cuts/test_times_0'
diff --git a/model/joint_simple_mlp_tgtcls.py b/model/joint_simple_mlp_tgtcls.py
index 834afbf..0aaf554 100644
--- a/model/joint_simple_mlp_tgtcls.py
+++ b/model/joint_simple_mlp_tgtcls.py
@@ -58,10 +58,12 @@ class Model(object):
         hidden = common_mlp.apply(inputs)
 
         dest_cls_probas = dest_mlp.apply(hidden)
+        # dest_cls_probas = theano.printing.Print("dest_cls_probas")(dest_cls_probas)
         dest_outputs = tensor.dot(dest_cls_probas, dest_classes)
         dest_outputs.name = 'dest_outputs'
 
         time_cls_probas = time_mlp.apply(hidden)
+        # time_cls_probas = theano.printing.Print("time_cls_probas")(time_cls_probas)
         time_outputs = tensor.dot(time_cls_probas, time_classes) + x_input_time
         time_outputs.name = 'time_outputs'
 
diff --git a/train.py b/train.py
index 96dd798..677ed45 100755
--- a/train.py
+++ b/train.py
@@ -73,7 +73,7 @@ def setup_test_stream(req_vars):
     test = transformers.TaxiAddFirstLastLen(config.n_begin_end_pts, test)
     test = transformers.Select(test, tuple(req_vars))
 
-    test_stream = Batch(test, iteration_scheme=ConstantScheme(1000))
+    test_stream = Batch(test, iteration_scheme=ConstantScheme(1))
 
     return test_stream
 
@@ -100,8 +100,8 @@ def main():
         cost=cost,
         step_rule=CompositeRule([
                 RemoveNotFinite(),
-                #AdaDelta(decay_rate=0.95),
-                Momentum(learning_rate=config.learning_rate, momentum=config.momentum),
+                AdaDelta(),
+                #Momentum(learning_rate=config.learning_rate, momentum=config.momentum),
         ]),
         params=params)
     
-- 
cgit v1.2.3