Refactor models, clean the code and separate training from testing.

1 files changed, 55 insertions, 93 deletions

diff --git a/model/joint_simple_mlp_tgtcls.py b/model/joint_simple_mlp_tgtcls.py
index 834afbf..d6d4e49 100644
--- a/model/joint_simple_mlp_tgtcls.py
+++ b/model/joint_simple_mlp_tgtcls.py
@@ -1,109 +1,71 @@
-from blocks.bricks import MLP, Rectifier, Linear, Sigmoid, Identity, Softmax
-from blocks.bricks.lookup import LookupTable
-
-from blocks.filter import VariableFilter
-from blocks.graph import ComputationGraph, apply_dropout
-
 import numpy
 import theano
 from theano import tensor
+from blocks import roles
+from blocks.bricks import application, MLP, Rectifier, Softmax
 
-import data
 import error
+from model.mlp import FFMLP, Stream
 
-class Model(object):
-    def __init__(self, config):
-        # The input and the targets
-        x_firstk_latitude = (tensor.matrix('first_k_latitude') - data.train_gps_mean[0]) / data.train_gps_std[0]
-        x_firstk_longitude = (tensor.matrix('first_k_longitude') - data.train_gps_mean[1]) / data.train_gps_std[1]
 
-        x_lastk_latitude = (tensor.matrix('last_k_latitude') - data.train_gps_mean[0]) / data.train_gps_std[0]
-        x_lastk_longitude = (tensor.matrix('last_k_longitude') - data.train_gps_mean[1]) / data.train_gps_std[1]
+class Model(FFMLP):
+    def __init__(self, config, **kwargs):
+        super(Model, self).__init__(config, **kwargs)
+        
+        self.dest_mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden_dest] + [Softmax()],
+                       dims=[config.dim_hidden[-1]] + config.dim_hidden_dest + [config.dim_output_dest],
+                       name='dest_mlp')
+        self.time_mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden_time] + [Softmax()],
+                       dims=[config.dim_hidden[-1]] + config.dim_hidden_time + [config.dim_output_time],
+                       name='time_mlp')
 
-        x_input_time = tensor.lvector('input_time')
+        self.dest_classes = theano.shared(numpy.array(config.dest_tgtcls, dtype=theano.config.floatX), name='dest_classes')
+        self.time_classes = theano.shared(numpy.array(config.time_tgtcls, dtype=theano.config.floatX), name='time_classes')
 
-        input_list = [x_firstk_latitude, x_firstk_longitude, x_lastk_latitude, x_lastk_longitude]
-        embed_tables = []
+        self.inputs.append('input_time')
+        self.children.extend([self.dest_mlp, self.time_mlp])
 
-        self.require_inputs = ['first_k_latitude', 'first_k_longitude', 'last_k_latitude', 'last_k_longitude', 'input_time']
+    def _push_initialization_config(self):
+        super(Model, self)._push_initialization_config()
+        for mlp in [self.dest_mlp, self.time_mlp]:
+            mlp.weights_init = self.config.mlp_weights_init
+            mlp.biases_init = self.config.mlp_biases_init
 
-        for (varname, num, dim) in config.dim_embeddings:
-            self.require_inputs.append(varname)
-            vardata = tensor.lvector(varname)
-            tbl = LookupTable(length=num, dim=dim, name='%s_lookup'%varname)
-            embed_tables.append(tbl)
-            input_list.append(tbl.apply(vardata))
+    @application(outputs=['destination', 'duration'])
+    def predict(self, **kwargs):
+        hidden = super(Model, self).predict(**kwargs)
 
-        y_dest = tensor.concatenate((tensor.vector('destination_latitude')[:, None],
-                                     tensor.vector('destination_longitude')[:, None]), axis=1)
-        y_time = tensor.lvector('travel_time')
+        dest_cls_probas = self.dest_mlp.apply(hidden)
+        dest_outputs = tensor.dot(dest_cls_probas, self.dest_classes)
 
-        # Define the model
-        common_mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden],
-                           dims=[config.dim_input] + config.dim_hidden)
+        time_cls_probas = self.time_mlp.apply(hidden)
+        time_outputs = kwargs['input_time'] + tensor.dot(time_cls_probas, self.time_classes)
 
-        dest_mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden_dest] + [Softmax()],
-                       dims=[config.dim_hidden[-1]] + config.dim_hidden_dest + [config.dim_output_dest],
-                       name='dest_mlp')
-        dest_classes = theano.shared(numpy.array(config.dest_tgtcls, dtype=theano.config.floatX), name='dest_classes')
+        self.add_auxiliary_variable(dest_cls_probas, name='destination classes ponderations')
+        self.add_auxiliary_variable(time_cls_probas, name='time classes ponderations')
 
-        time_mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden_time] + [Softmax()],
-                       dims=[config.dim_hidden[-1]] + config.dim_hidden_time + [config.dim_output_time],
-                       name='time_mlp')
-        time_classes = theano.shared(numpy.array(config.time_tgtcls, dtype=theano.config.floatX), name='time_classes')
-
-        # Create the Theano variables
-        inputs = tensor.concatenate(input_list, axis=1)
-        # inputs = theano.printing.Print("inputs")(inputs)
-        hidden = common_mlp.apply(inputs)
-
-        dest_cls_probas = dest_mlp.apply(hidden)
-        dest_outputs = tensor.dot(dest_cls_probas, dest_classes)
-        dest_outputs.name = 'dest_outputs'
-
-        time_cls_probas = time_mlp.apply(hidden)
-        time_outputs = tensor.dot(time_cls_probas, time_classes) + x_input_time
-        time_outputs.name = 'time_outputs'
-
-        # Calculate the cost
-        dest_cost = error.erdist(dest_outputs, y_dest).mean()
-        dest_cost.name = 'dest_cost'
-        dest_hcost = error.hdist(dest_outputs, y_dest).mean()
-        dest_hcost.name = 'dest_hcost'
-
-        time_cost = error.rmsle(time_outputs.flatten(), y_time.flatten())
-        time_cost.name = 'time_cost'
-        time_scost = config.time_cost_factor * time_cost
-        time_scost.name = 'time_scost'
-
-        cost = dest_cost + time_scost
-
-        if hasattr(config, 'dropout_p'):
-            cg = ComputationGraph(cost)
-            dropout_inputs = VariableFilter(
-                    bricks=[b for b in list(common_mlp.children) +
-                                       list(dest_mlp.children) +
-                                       list(time_mlp.children)
-                              if isinstance(b, Rectifier)],
-                    name='output')(cg)
-            cg = apply_dropout(cg, dropout_inputs, config.dropout_p)
-            cost = cg.outputs[0]
-
-        cost.name = 'cost'
-
-        # Initialization
-        for tbl in embed_tables:
-            tbl.weights_init = config.embed_weights_init
-            tbl.initialize()
-
-        for mlp in [common_mlp, dest_mlp, time_mlp]:
-            mlp.weights_init = config.mlp_weights_init
-            mlp.biases_init = config.mlp_biases_init
-            mlp.initialize()
-
-        self.cost = cost
-        self.monitor = [cost, dest_cost, dest_hcost, time_cost, time_scost]
-        self.outputs = tensor.concatenate([dest_outputs, time_outputs[:, None]], axis=1)
-        self.outputs.name = 'outputs'
-        self.pred_vars = ['destination_longitude', 'destination_latitude', 'travel_time']
+        return (dest_outputs, time_outputs)
+
+    @predict.property('inputs')
+    def predict_inputs(self):
+        return self.inputs
+
+    @application(outputs=['cost'])
+    def cost(self, **kwargs):
+        (destination_hat, time_hat) = self.predict(**kwargs)
+
+        destination = tensor.concatenate((kwargs['destination_latitude'][:, None],
+                                          kwargs['destination_longitude'][:, None]), axis=1)
+        time = kwargs['travel_time']
+
+        destination_cost = error.erdist(destination_hat, destination).mean()
+        time_cost = error.rmsle(time_hat.flatten(), time.flatten())
+
+        self.add_auxiliary_variable(destination_cost, [roles.COST], 'destination_cost')
+        self.add_auxiliary_variable(time_cost, [roles.COST], 'time_cost')
+
+        return destination_cost + self.config.time_cost_factor * time_cost
 
+    @cost.property('inputs')
+    def cost_inputs(self):
+        return self.inputs + ['destination_latitude', 'destination_longitude', 'travel_time']