Genericize RNNs

4 files changed, 109 insertions, 36 deletions

diff --git a/model/rnn.py b/model/rnn.py
index be17a95..bfc3122 100644
--- a/model/rnn.py
+++ b/model/rnn.py
@@ -16,16 +16,16 @@ from data.hdf5 import TaxiDataset, TaxiStream
 import error
 
 
-class Model(Initializable):
+class RNN(Initializable):
     @lazy()
-    def __init__(self, config, **kwargs):
-        super(Model, self).__init__(**kwargs)
+    def __init__(self, config, rec_input_len=2, output_dim=2, **kwargs):
+        super(RNN, self).__init__(**kwargs)
         self.config = config
 
         self.pre_context_embedder = ContextEmbedder(config.pre_embedder, name='pre_context_embedder')
         self.post_context_embedder = ContextEmbedder(config.post_embedder, name='post_context_embedder')
 
-        in1 = 2 + sum(x[2] for x in config.pre_embedder.dim_embeddings)
+        in1 = rec_input_len + sum(x[2] for x in config.pre_embedder.dim_embeddings)
         self.input_to_rec = MLP(activations=[Tanh()], dims=[in1, config.hidden_state_dim], name='input_to_rec')
 
         self.rec = LSTM(
@@ -34,7 +34,7 @@ class Model(Initializable):
             )
 
         in2 = config.hidden_state_dim + sum(x[2] for x in config.post_embedder.dim_embeddings)
-        self.rec_to_output = MLP(activations=[Tanh()], dims=[in2, 2], name='rec_to_output')
+        self.rec_to_output = MLP(activations=[Tanh()], dims=[in2, output_dim], name='rec_to_output')
 
         self.sequences = ['latitude', 'latitude_mask', 'longitude']
         self.context = self.pre_context_embedder.inputs + self.post_context_embedder.inputs
@@ -55,59 +55,69 @@ class Model(Initializable):
     def get_dim(self, name):
         return self.rec.get_dim(name)
 
-    @application
-    def initial_state(self, *args, **kwargs):
-        return self.rec.initial_state(*args, **kwargs)
+    def process_rto(self, rto):
+        return rto
 
-    @recurrent(states=['states', 'cells'], outputs=['destination', 'states', 'cells'], sequences=['latitude', 'longitude', 'latitude_mask'])
-    def predict_all(self, latitude, longitude, latitude_mask, **kwargs):
-        latitude = (latitude - data.train_gps_mean[0]) / data.train_gps_std[0]
-        longitude = (longitude - data.train_gps_mean[1]) / data.train_gps_std[1]
+    def rec_input(self, latitude, longitude, **kwargs):
+        return (tensor.shape_padright(latitude), tensor.shape_padright(longitude))
 
+    @recurrent(states=['states', 'cells'], outputs=['destination', 'states', 'cells'])
+    def predict_all(self, **kwargs):
         pre_emb = tuple(self.pre_context_embedder.apply(**kwargs))
-        latitude = tensor.shape_padright(latitude)
-        longitude = tensor.shape_padright(longitude)
-        itr = self.input_to_rec.apply(tensor.concatenate(pre_emb + (latitude, longitude), axis=1))
+
+        itr_in = tensor.concatenate(pre_emb + self.rec_input(**kwargs), axis=1)
+        itr = self.input_to_rec.apply(itr_in)
         itr = itr.repeat(4, axis=1)
-        (next_states, next_cells) = self.rec.apply(itr, kwargs['states'], kwargs['cells'], mask=latitude_mask, iterate=False)
+        (next_states, next_cells) = self.rec.apply(itr, kwargs['states'], kwargs['cells'], mask=kwargs['latitude_mask'], iterate=False)
 
         post_emb = tuple(self.post_context_embedder.apply(**kwargs))
         rto = self.rec_to_output.apply(tensor.concatenate(post_emb + (next_states,), axis=1))
 
-        rto = (rto * data.train_gps_std) + data.train_gps_mean
+        rto = self.process_rto(rto)
         return (rto, next_states, next_cells)
 
+    @predict_all.property('sequences')
+    def predict_all_sequences(self):
+        return self.sequences
+
+    @application(outputs=predict_all.states)
+    def initial_states(self, *args, **kwargs):
+        return self.rec.initial_states(*args, **kwargs)
+
     @predict_all.property('contexts')
-    def predict_all_inputs(self):
+    def predict_all_context(self):
         return self.context
 
+    def before_predict_all(self, kwargs):
+        kwargs['latitude'] = (kwargs['latitude'].T - data.train_gps_mean[0]) / data.train_gps_std[0]
+        kwargs['longitude'] = (kwargs['longitude'].T - data.train_gps_mean[1]) / data.train_gps_std[1]
+        kwargs['latitude_mask'] = kwargs['latitude_mask'].T
+
     @application(outputs=['destination'])
-    def predict(self, latitude, longitude, latitude_mask, **kwargs):
-        latitude = latitude.T
-        longitude = longitude.T
-        latitude_mask = latitude_mask.T
-        res = self.predict_all(latitude, longitude, latitude_mask, **kwargs)[0]
-        return res[-1]
+    def predict(self, **kwargs):
+        self.before_predict_all(kwargs)
+        res = self.predict_all(**kwargs)[0]
+
+        last_id = tensor.cast(kwargs['latitude_mask'].sum(axis=0) - 1, dtype='int64')
+        return res[last_id]
 
     @predict.property('inputs')
     def predict_inputs(self):
         return self.inputs
 
     @application(outputs=['cost_matrix'])
-    def cost_matrix(self, latitude, longitude, latitude_mask, **kwargs):
-        latitude = latitude.T
-        longitude = longitude.T
-        latitude_mask = latitude_mask.T
+    def cost_matrix(self, **kwargs):
+        self.before_predict_all(kwargs)
 
-        res = self.predict_all(latitude, longitude, latitude_mask, **kwargs)[0]
+        res = self.predict_all(**kwargs)[0]
         target = tensor.concatenate(
                     (kwargs['destination_latitude'].dimshuffle('x', 0, 'x'),
                      kwargs['destination_longitude'].dimshuffle('x', 0, 'x')),
                 axis=2)
-        target = target.repeat(latitude.shape[0], axis=0)
+        target = target.repeat(kwargs['latitude'].shape[0], axis=0)
         ce = error.erdist(target.reshape((-1, 2)), res.reshape((-1, 2)))
-        ce = ce.reshape(latitude.shape)
-        return ce * latitude_mask
+        ce = ce.reshape(kwargs['latitude'].shape)
+        return ce * kwargs['latitude_mask']
 
     @cost_matrix.property('inputs')
     def cost_matrix_inputs(self):
@@ -123,8 +133,8 @@ class Model(Initializable):
 
     @application(outputs=['cost'])
     def valid_cost(self, **kwargs):
-        # Only works when batch_size is 1.
-        return self.cost_matrix(**kwargs)[-1,0]
+        last_id = tensor.cast(kwargs['latitude_mask'].sum(axis=1) - 1, dtype='int64')
+        return self.cost_matrix(**kwargs)[last_id].mean()
 
     @valid_cost.property('inputs')
     def valid_cost_inputs(self):
@@ -158,7 +168,7 @@ class Stream(object):
         stream = transformers.add_destination(stream)
         stream = transformers.Select(stream, tuple(v for v in req_vars if not v.endswith('_mask')))
 
-        stream = Batch(stream, iteration_scheme=ConstantScheme(1))
+        stream = Batch(stream, iteration_scheme=ConstantScheme(self.config.batch_size))
         stream = Padding(stream, mask_sources=['latitude', 'longitude'])
         stream = transformers.Select(stream, req_vars)
         return stream
@@ -169,7 +179,7 @@ class Stream(object):
         stream = transformers.taxi_remove_test_only_clients(stream)
         stream = transformers.Select(stream, tuple(v for v in req_vars if not v.endswith('_mask')))
 
-        stream = Batch(stream, iteration_scheme=ConstantScheme(1))
+        stream = Batch(stream, iteration_scheme=ConstantScheme(self.config.batch_size))
         stream = Padding(stream, mask_sources=['latitude', 'longitude'])
         stream = transformers.Select(stream, req_vars)
         return stream
diff --git a/model/rnn_direct.py b/model/rnn_direct.py
new file mode 100644
index 0000000..8123327
--- /dev/null
+++ b/model/rnn_direct.py
@@ -0,0 +1,13 @@
+from blocks.bricks.base import lazy
+
+from model.rnn import RNN, Stream
+import data
+
+
+class Model(RNN):
+    @lazy()
+    def __init__(self, config, **kwargs):
+        super(Model, self).__init__(config, **kwargs)
+
+    def process_rto(self, rto):
+        return (rto * data.train_gps_std) + data.train_gps_mean
diff --git a/model/rnn_lag_tgtcls.py b/model/rnn_lag_tgtcls.py
new file mode 100644
index 0000000..49d1f01
--- /dev/null
+++ b/model/rnn_lag_tgtcls.py
@@ -0,0 +1,31 @@
+import numpy
+import theano
+from theano import tensor
+from blocks.bricks.base import lazy
+from blocks.bricks import Softmax
+
+from model.rnn import RNN, Stream
+
+
+class Model(RNN):
+    @lazy()
+    def __init__(self, config, **kwargs):
+        super(Model, self).__init__(config, rec_input_len=4, output_dim=config.tgtcls.shape[0], **kwargs)
+        self.classes = theano.shared(numpy.array(config.tgtcls, dtype=theano.config.floatX), name='classes')
+        self.softmax = Softmax()
+        self.sequences.extend(['latitude_lag', 'longitude_lag'])
+        self.children.append(self.softmax)
+
+    def before_predict_all(self, kwargs):
+        super(Model, self).before_predict_all(kwargs)
+        kwargs['latitude_lag'] = tensor.extra_ops.repeat(kwargs['latitude'], 2, axis=0)
+        kwargs['longitude_lag'] = tensor.extra_ops.repeat(kwargs['longitude'], 2, axis=0)
+
+    def process_rto(self, rto):
+        return tensor.dot(self.softmax.apply(rto), self.classes)
+
+    def rec_input(self, latitude, longitude, latitude_lag, longitude_lag, **kwargs):
+        return (tensor.shape_padright(latitude),
+                tensor.shape_padright(longitude),
+                tensor.shape_padright(latitude_lag),
+                tensor.shape_padright(longitude_lag))
diff --git a/model/rnn_tgtcls.py b/model/rnn_tgtcls.py
new file mode 100644
index 0000000..0c0faf2
--- /dev/null
+++ b/model/rnn_tgtcls.py
@@ -0,0 +1,19 @@
+import numpy
+import theano
+from theano import tensor
+from blocks.bricks.base import lazy
+from blocks.bricks import Softmax
+
+from model.rnn import RNN, Stream
+
+
+class Model(RNN):
+    @lazy()
+    def __init__(self, config, **kwargs):
+        super(Model, self).__init__(config, output_dim=config.tgtcls.shape[0], **kwargs)
+        self.classes = theano.shared(numpy.array(config.tgtcls, dtype=theano.config.floatX), name='classes')
+        self.softmax = Softmax()
+        self.children.append(self.softmax)
+
+    def process_rto(self, rto):
+        return tensor.dot(self.softmax.apply(rto), self.classes)