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

7 files changed, 281 insertions, 326 deletions

diff --git a/model/__init__.py b/model/__init__.py
index e69de29..5c051f4 100644
--- a/model/__init__.py
+++ b/model/__init__.py
@@ -0,0 +1,36 @@
+from blocks.bricks import application, Initializable
+from blocks.bricks.lookup import LookupTable
+
+
+class ContextEmbedder(Initializable):
+    def __init__(self, config, **kwargs):
+        super(ContextEmbedder, self).__init__(**kwargs)
+        self.dim_embeddings = config.dim_embeddings
+        self.embed_weights_init = config.embed_weights_init
+
+        self.inputs = [ name for (name, _, _) in self.dim_embeddings ]
+        self.outputs = [ '%s_embedded' % name for name in self.inputs ]
+
+        self.lookups = { name: LookupTable(name='%s_lookup' % name) for name in self.inputs }
+        self.children = self.lookups.values()
+
+    def _push_allocation_config(self):
+        for (name, num, dim) in self.dim_embeddings:
+            self.lookups[name].length = num
+            self.lookups[name].dim = dim
+
+    def _push_initialization_config(self):
+        for name in self.inputs:
+            self.lookups[name].weights_init = self.embed_weights_init
+
+    @application
+    def apply(self, **kwargs):
+        return tuple(self.lookups[name].apply(kwargs[name]) for name in self.inputs)
+
+    @apply.property('inputs')
+    def apply_inputs(self):
+        return self.inputs
+
+    @apply.property('outputs')
+    def apply_outputs(self):
+        return self.outputs
diff --git a/model/dest_simple_mlp.py b/model/dest_simple_mlp.py
index a9e97cb..78d7131 100644
--- a/model/dest_simple_mlp.py
+++ b/model/dest_simple_mlp.py
@@ -1,71 +1,32 @@
-from blocks.bricks import MLP, Rectifier, Linear, Sigmoid, Identity
-from blocks.bricks.lookup import LookupTable
-
 from theano import tensor
+from blocks.bricks import application, Identity
 
 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]
-
-        input_list = [x_firstk_latitude, x_firstk_longitude, x_lastk_latitude, x_lastk_longitude]
-        embed_tables = []
-
-        self.require_inputs = ['first_k_latitude', 'first_k_longitude', 'last_k_latitude', 'last_k_longitude']
-
-        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))
-
-        y = tensor.concatenate((tensor.vector('destination_latitude')[:, None],
-                                tensor.vector('destination_longitude')[:, None]), axis=1)
-
-        # Define the model
-        mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden] + [Identity()],
-                           dims=[config.dim_input] + config.dim_hidden + [config.dim_output])
-
-        # Create the Theano variables
-        inputs = tensor.concatenate(input_list, axis=1)
-        # inputs = theano.printing.Print("inputs")(inputs)
-        outputs = mlp.apply(inputs)
-
-        # Normalize & Center
-        # outputs = theano.printing.Print("normal_outputs")(outputs)
-        outputs = data.train_gps_std * outputs + data.train_gps_mean
-
-        # outputs = theano.printing.Print("outputs")(outputs)
-        # y = theano.printing.Print("y")(y)
 
-        outputs.name = 'outputs'
+class Model(FFMLP):
+    def __init__(self, config, **kwargs):
+        super(Model, self).__init__(config, output_layer=Identity, **kwargs)
 
-        # Calculate the cost
-        cost = error.erdist(outputs, y).mean()
-        cost.name = 'cost'
-        hcost = error.hdist(outputs, y).mean()
-        hcost.name = 'hcost'
+    @application(outputs=['destination'])
+    def predict(self, **kwargs):
+        outputs = super(Model, self).predict(**kwargs)
+        return data.train_gps_std * outputs + data.train_gps_mean
 
-        # Initialization
-        for tbl in embed_tables:
-            tbl.weights_init = config.embed_weights_init
-        mlp.weights_init = config.mlp_weights_init
-        mlp.biases_init = config.mlp_biases_init
+    @predict.property('inputs')
+    def predict_inputs(self):
+        return self.inputs
 
-        for tbl in embed_tables:
-            tbl.initialize()
-        mlp.initialize()
+    @application(outputs=['cost'])
+    def cost(self, **kwargs):
+        y_hat = self.predict(**kwargs)
+        y = tensor.concatenate((kwargs['destination_latitude'][:, None],
+                                kwargs['destination_longitude'][:, None]), axis=1)
 
-        self.cost = cost
-        self.monitor = [cost, hcost]
-        self.outputs = outputs
-        self.pred_vars = ['destination_latitude', 'destination_longitude']
+        return error.erdist(y_hat, y).mean()
 
+    @cost.property('inputs')
+    def cost_inputs(self):
+        return self.inputs + ['destination_latitude', 'destination_longitude']
diff --git a/model/dest_simple_mlp_tgtcls.py b/model/dest_simple_mlp_tgtcls.py
index 1381d4c..2d65097 100644
--- a/model/dest_simple_mlp_tgtcls.py
+++ b/model/dest_simple_mlp_tgtcls.py
@@ -1,73 +1,34 @@
-import numpy 
-
+import numpy
 import theano
 from theano import tensor
+from blocks.bricks import application, Softmax
 
-from blocks.bricks import MLP, Rectifier, Linear, Sigmoid, Identity, Softmax
-from blocks.bricks.lookup import LookupTable
-
-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]
-
-        input_list = [x_firstk_latitude, x_firstk_longitude, x_lastk_latitude, x_lastk_longitude]
-        embed_tables = []
-
-        self.require_inputs = ['first_k_latitude', 'first_k_longitude', 'last_k_latitude', 'last_k_longitude']
-
-        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))
-
-        y = tensor.concatenate((tensor.vector('destination_latitude')[:, None],
-                                tensor.vector('destination_longitude')[:, None]), axis=1)
-
-        # Define the model
-        mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden] + [Softmax()],
-                           dims=[config.dim_input] + config.dim_hidden + [config.dim_output])
-        classes = theano.shared(numpy.array(config.tgtcls, dtype=theano.config.floatX), name='classes')
-
-        # Create the Theano variables
-        inputs = tensor.concatenate(input_list, axis=1)
-
-        # inputs = theano.printing.Print("inputs")(inputs)
-        cls_probas = mlp.apply(inputs)
-        outputs = tensor.dot(cls_probas, classes)
-
-        # outputs = theano.printing.Print("outputs")(outputs)
-        # y = theano.printing.Print("y")(y)
 
-        outputs.name = 'outputs'
+class Model(FFMLP):
+    def __init__(self, config, **kwargs):
+        super(Model, self, output_layer=Softmax).__init__(config, **kwargs)
+        self.classes = theano.shared(numpy.array(config.tgtcls, dtype=theano.config.floatX), name='classes')
 
-        # Calculate the cost
-        cost = error.erdist(outputs, y).mean()
-        cost.name = 'cost'
-        hcost = error.hdist(outputs, y).mean()
-        hcost.name = 'hcost'
+    @application(outputs=['destination'])
+    def predict(self, **kwargs):
+        cls_probas = super(Model, self).predict(**kwargs)
+        return tensor.dot(cls_probas, self.classes)
 
-        # Initialization
-        for tbl in embed_tables:
-            tbl.weights_init = config.embed_weights_init
-        mlp.weights_init = config.mlp_weights_init
-        mlp.biases_init = config.mlp_biases_init
+    @predict.property('inputs')
+    def predict_inputs(self):
+        return self.inputs
 
-        for tbl in embed_tables:
-            tbl.initialize()
-        mlp.initialize()
+    @application(outputs=['cost'])
+    def cost(self, **kwargs):
+        y_hat = self.predict(**kwargs)
+        y = tensor.concatenate((kwargs['destination_latitude'][:, None],
+                                kwargs['destination_longitude'][:, None]), axis=1)
 
-        self.cost = cost
-        self.monitor = [cost, hcost]
-        self.outputs = outputs
-        self.pred_vars = ['destination_latitude', 'destination_longitude']
+        return error.erdist(y_hat, y).mean()
 
+    @cost.property('inputs')
+    def cost_inputs(self):
+        return self.inputs + ['destination_latitude', 'destination_longitude']
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']
diff --git a/model/mlp.py b/model/mlp.py
new file mode 100644
index 0000000..9c84ef9
--- /dev/null
+++ b/model/mlp.py
@@ -0,0 +1,103 @@
+from theano import tensor
+
+from fuel.transformers import Batch
+from fuel.streams import DataStream
+from fuel.schemes import ConstantScheme, ShuffledExampleScheme
+from blocks.bricks import application, MLP, Rectifier, Initializable
+
+import data
+from data import transformers
+from data.hdf5 import TaxiDataset, TaxiStream
+from model import ContextEmbedder
+
+
+class FFMLP(Initializable):
+    def __init__(self, config, output_layer=None, **kwargs):
+        super(FFMLP, self).__init__(**kwargs)
+        self.config = config
+
+        self.context_embedder = ContextEmbedder(config)
+
+        output_activation = [] if output_layer is None else [output_layer()]
+        output_dim = [] if output_layer is None else [config.dim_output]
+        self.mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden] + output_activation,
+                       dims=[config.dim_input] + config.dim_hidden + output_dim)
+
+        self.extremities = {'%s_k_%s' % (side, ['latitude', 'longitude'][axis]): axis for side in ['first', 'last'] for axis in [0, 1]}
+        self.inputs = self.context_embedder.inputs + self.extremities.keys()
+        self.children = [ self.context_embedder, self.mlp ]
+
+    def _push_initialization_config(self):
+        self.mlp.weights_init = self.config.mlp_weights_init
+        self.mlp.biases_init = self.config.mlp_biases_init
+
+    @application(outputs=['prediction'])
+    def predict(self, **kwargs):
+        embeddings = tuple(self.context_embedder.apply(**{k: kwargs[k] for k in self.context_embedder.inputs }))
+        extremities = tuple((kwargs[k] - data.train_gps_mean[v]) / data.train_gps_std[v] for k, v in self.extremities.items())
+
+        inputs = tensor.concatenate(extremities + embeddings, axis=1)
+        outputs = self.mlp.apply(inputs)
+
+        return outputs
+
+    @predict.property('inputs')
+    def predict_inputs(self):
+        return self.inputs
+
+class Stream(object):
+    def __init__(self, config):
+        self.config = config
+
+    def train(self, req_vars):
+        stream = TaxiDataset('train')
+        stream = DataStream(stream, iteration_scheme=ShuffledExampleScheme(stream.num_examples))
+
+        valid = TaxiDataset(self.config.valid_set, 'valid.hdf5', sources=('trip_id',))
+        valid_trips_ids = valid.get_data(None, slice(0, valid.num_examples))[0]
+
+        stream = transformers.TaxiExcludeTrips(valid_trips_ids, stream)
+        stream = transformers.TaxiGenerateSplits(stream, max_splits=100)
+
+        stream = transformers.TaxiAddDateTime(stream)
+        stream = transformers.TaxiAddFirstLastLen(self.config.n_begin_end_pts, stream)
+        stream = transformers.Select(stream, tuple(req_vars))
+        return Batch(stream, iteration_scheme=ConstantScheme(self.config.batch_size))
+
+    def valid(self, req_vars):
+        stream = TaxiStream(self.config.valid_set, 'valid.hdf5')
+
+        stream = transformers.TaxiAddDateTime(stream)
+        stream = transformers.TaxiAddFirstLastLen(self.config.n_begin_end_pts, stream)
+        stream = transformers.Select(stream, tuple(req_vars))
+        return Batch(stream, iteration_scheme=ConstantScheme(1000))
+
+    def test(self, req_vars):
+        stream = TaxiStream('test')
+        
+        stream = transformers.TaxiAddDateTime(stream)
+        stream = transformers.TaxiAddFirstLastLen(self.config.n_begin_end_pts, stream)
+
+        return Batch(stream, iteration_scheme=ConstantScheme(1))
+
+    def inputs(self):
+        return {'call_type': tensor.bvector('call_type'),
+                'origin_call': tensor.ivector('origin_call'),
+                'origin_stand': tensor.bvector('origin_stand'),
+                'taxi_id': tensor.wvector('taxi_id'),
+                'timestamp': tensor.ivector('timestamp'),
+                'day_type': tensor.bvector('day_type'),
+                'missing_data': tensor.bvector('missing_data'),
+                'latitude': tensor.matrix('latitude'),
+                'longitude': tensor.matrix('longitude'),
+                'destination_latitude': tensor.vector('destination_latitude'),
+                'destination_longitude': tensor.vector('destination_longitude'),
+                'travel_time': tensor.ivector('travel_time'),
+                'first_k_latitude': tensor.matrix('first_k_latitude'),
+                'first_k_longitude': tensor.matrix('first_k_longitude'),
+                'last_k_latitude': tensor.matrix('last_k_latitude'),
+                'last_k_longitude': tensor.matrix('last_k_longitude'),
+                'input_time': tensor.ivector('input_time'),
+                'week_of_year': tensor.bvector('week_of_year'),
+                'day_of_week': tensor.bvector('day_of_week'),
+                'qhour_of_day': tensor.bvector('qhour_of_day')}
diff --git a/model/time_simple_mlp.py b/model/time_simple_mlp.py
index 9e1b10a..0ce6b29 100644
--- a/model/time_simple_mlp.py
+++ b/model/time_simple_mlp.py
@@ -1,63 +1,29 @@
-from blocks.bricks import MLP, Rectifier, Linear, Sigmoid, Identity
-from blocks.bricks.lookup import LookupTable
+from blocks.bricks import application, Identity
 
-from theano import tensor
-
-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]
-
-        input_list = [x_firstk_latitude, x_firstk_longitude, x_lastk_latitude, x_lastk_longitude]
-        embed_tables = []
-
-        self.require_inputs = ['first_k_latitude', 'first_k_longitude', 'last_k_latitude', 'last_k_longitude']
-
-        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))
-
-        y = tensor.lvector('travel_time')
-
-        # Define the model
-        mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden] + [Identity()],
-                           dims=[config.dim_input] + config.dim_hidden + [config.dim_output])
-
-        # Create the Theano variables
-        inputs = tensor.concatenate(input_list, axis=1)
-        # inputs = theano.printing.Print("inputs")(inputs)
-        outputs = config.exp_base ** mlp.apply(inputs)
-
-        # outputs = theano.printing.Print("outputs")(outputs)
-        # y = theano.printing.Print("y")(y)
 
-        outputs.name = 'outputs'
+class Model(FFMLP):
+    def __init__(self, config, **kwargs):
+        super(Model, self).__init__(config, output_layer=Identity, **kwargs)
+        self.inputs.append('input_time')
 
-        # Calculate the cost
-        cost = error.rmsle(outputs.flatten(), y.flatten())
-        cost.name = 'cost'
+    @application(outputs=['duration'])
+    def predict(self, **kwargs):
+        outputs = super(Model, self).predict(**kwargs)
+        return kwargs['input_time'] + self.config.exp_base ** outputs
 
-        # Initialization
-        for tbl in embed_tables:
-            tbl.weights_init = config.embed_weights_init
-        mlp.weights_init = config.mlp_weights_init
-        mlp.biases_init = config.mlp_biases_init
+    @predict.property('inputs')
+    def predict_inputs(self):
+        return self.inputs
 
-        for tbl in embed_tables:
-            tbl.initialize()
-        mlp.initialize()
+    @application(outputs=['cost'])
+    def cost(self, **kwargs):
+        y_hat = self.predict(**kwargs)
+        y = kwargs['travel_time']
+        return error.rmsle(y_hat.flatten(), y.flatten())
 
-        self.cost = cost
-        self.monitor = [cost]
-        self.outputs = outputs
-        self.pred_vars = ['travel_time']
+    @cost.property('inputs')
+    def cost_inputs(self):
+        return self.inputs + ['travel_time']
diff --git a/model/time_simple_mlp_tgtcls.py b/model/time_simple_mlp_tgtcls.py
index 1f1eab7..35c8d8a 100644
--- a/model/time_simple_mlp_tgtcls.py
+++ b/model/time_simple_mlp_tgtcls.py
@@ -1,67 +1,33 @@
-from blocks.bricks import MLP, Rectifier, Linear, Sigmoid, Identity, Softmax
-from blocks.bricks.lookup import LookupTable
-
 import numpy
 import theano
 from theano import tensor
+from blocks.bricks import application, 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]
-
-        input_list = [x_firstk_latitude, x_firstk_longitude, x_lastk_latitude, x_lastk_longitude]
-        embed_tables = []
-
-        self.require_inputs = ['first_k_latitude', 'first_k_longitude', 'last_k_latitude', 'last_k_longitude']
-
-        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))
-
-        y = tensor.lvector('travel_time')
-
-        # Define the model
-        mlp = MLP(activations=[Rectifier() for _ in config.dim_hidden] + [Softmax()],
-                           dims=[config.dim_input] + config.dim_hidden + [config.dim_output])
-        classes = theano.shared(numpy.array(config.tgtcls, dtype=theano.config.floatX), name='classes')
-
-        # Create the Theano variables
-        inputs = tensor.concatenate(input_list, axis=1)
-        # inputs = theano.printing.Print("inputs")(inputs)
-        cls_probas = mlp.apply(inputs)
-        outputs = tensor.dot(cls_probas, classes)
-
-        # outputs = theano.printing.Print("outputs")(outputs)
-        # y = theano.printing.Print("y")(y)
 
-        outputs.name = 'outputs'
+class Model(FFMLP):
+    def __init__(self, config, **kwargs):
+        super(Model, self, output_layer=Softmax).__init__(config, **kwargs)
+        self.classes = theano.shared(numpy.array(config.tgtcls, dtype=theano.config.floatX), name='classes')
+        self.inputs.append('input_time')
 
-        # Calculate the cost
-        cost = error.rmsle(outputs.flatten(), y.flatten())
-        cost.name = 'cost'
+    @application(outputs=['duration'])
+    def predict(self, **kwargs):
+        cls_probas = super(Model, self).predict(**kwargs)
+        return kwargs['input_time'] + tensor.dot(cls_probas, self.classes)
 
-        # Initialization
-        for tbl in embed_tables:
-            tbl.weights_init = config.embed_weights_init
-        mlp.weights_init = config.mlp_weights_init
-        mlp.biases_init = config.mlp_biases_init
+    @predict.property('inputs')
+    def predict_inputs(self):
+        return self.inputs
 
-        for tbl in embed_tables:
-            tbl.initialize()
-        mlp.initialize()
+    @application(outputs=['cost'])
+    def cost(self, **kwargs):
+        y_hat = self.predict(**kwargs)
+        y = kwargs['travel_time']
+        return error.rmsle(y_hat.flatten(), y.flatten())
 
-        self.cost = cost
-        self.monitor = [cost]
-        self.outputs = outputs
-        self.pred_vars = ['travel_time']
+    @cost.property('inputs')
+    def cost_inputs(self):
+        return self.inputs + ['travel_time']