taxi

commit cdd050295c3c6df780bdc65088959d908b2cf2a5
parent 9a60f6c4e39c09187710608a9e225b6024b34364
Author: Alex Auvolat <alex.auvolat@ens.fr>
Date:   Mon, 27 Apr 2015 18:27:58 -0400

A reasonnable neural network, only it doesn't work.

Diffstat:
M
model.py
 | 
28
+++++++++++++++++++---------

1 file changed, 19 insertions(+), 9 deletions(-)
diff --git a/model.py b/model.py
@@ -5,10 +5,13 @@ from argparse import ArgumentParser
 import numpy
 
 import theano
+from theano import printing
 from theano import tensor
 from theano.ifelse import ifelse
 
-from blocks.bricks import MLP, Rectifier, Linear
+from blocks.filter import VariableFilter
+
+from blocks.bricks import MLP, Rectifier, Linear, Sigmoid, Identity
 from blocks.bricks.lookup import LookupTable
 
 from blocks.initialization import IsotropicGaussian, Constant
@@ -18,7 +21,7 @@ from fuel.transformers import Batch
 from fuel.streams import DataStream
 from fuel.schemes import ConstantScheme
 
-from blocks.algorithms import GradientDescent, Scale
+from blocks.algorithms import GradientDescent, Scale, AdaDelta
 from blocks.graph import ComputationGraph
 from blocks.main_loop import MainLoop
 from blocks.extensions import Printing
@@ -43,7 +46,7 @@ dim_embed = 50
 dim_hidden = 200
 
 learning_rate = 0.01
-batch_size = 64
+batch_size = 32
 
 def main():
     # The input and the targets
@@ -65,25 +68,31 @@ def main():
     client_embed = client_embed_table.apply(x_client).flatten(ndim=2)
     stand_embed = stand_embed_table.apply(x_stand).flatten(ndim=2)
     inputs = tensor.concatenate([x_firstk, x_lastk,
-                                 client_embed, stand_embed],
+                                 client_embed.zeros_like(), stand_embed.zeros_like()],
                                 axis=1)
+    # inputs = theano.printing.Print("inputs")(inputs)
     hidden = hidden_layer.apply(inputs)
+    # hidden = theano.printing.Print("hidden")(hidden)
     outputs = output_layer.apply(hidden)
 
     # Calculate the cost
     # cost = (outputs - y).norm(2, axis=1).mean()
     # outputs = numpy.array([[ -8.621953, 41.162142]], dtype='float32') + 0 * outputs
-    cost = hdist.hdist(outputs, y).mean()
+    cost = (outputs - y).norm(2, axis=1).mean()
     cost.name = 'cost'
+    hcost = hdist.hdist(outputs, y).mean()
+    hcost.name = 'hcost'
 
     # Initialization
     client_embed_table.weights_init = IsotropicGaussian(0.001)
+    stand_embed_table.weights_init = IsotropicGaussian(0.001)
     hidden_layer.weights_init = IsotropicGaussian(0.01)
     hidden_layer.biases_init = Constant(0.001)
     output_layer.weights_init = IsotropicGaussian(0.001)
     output_layer.biases_init = Constant(0.001)
 
     client_embed_table.initialize()
+    stand_embed_table.initialize()
     hidden_layer.initialize()
     output_layer.initialize()
 
@@ -107,13 +116,14 @@ def main():
 
     # Training
     cg = ComputationGraph(cost)
+    params = VariableFilter(bricks=[Linear])(cg.parameters)
     algorithm = GradientDescent(
         cost=cost,
-        # step_rule=AdaDelta(decay_rate=0.5),
-        step_rule=Scale(learning_rate=learning_rate),
-        params=cg.parameters)
+        step_rule=AdaDelta(decay_rate=0.5),
+        # step_rule=Scale(learning_rate=learning_rate),
+        params=params)
 
-    extensions=[DataStreamMonitoring([cost], valid_stream,
+    extensions=[DataStreamMonitoring([cost, hcost], valid_stream,
                                      prefix='valid',
                                      every_n_batches=1000),
                 Printing(every_n_batches=1000),