def __init__(self, cached_features):
base_model.BaseModel.__init__(self, cached_features)
self.weights_file = "working/models/lstm_model.pickle.weights"
self.n_epochs = 4
self.padding_size = 256
self.max_features = 43980
self.batch_size = 32
self.model = self.SetupModel()
self.X_train = None
self.y_train = numpy.array([])
self.feature_extractor = CachedFeatureExtractor("working/models/sequence_train_batches")
self.test_feature_extractor = CachedFeatureExtractor("working/models/sequence_test_batches")
class BaseModel(object):
"""
This is the Base class for all models
There are two ways of training any of the derived models:
1) Online: override the _fit_internal() method that trains on a batch
2) Complete: override the fit() method that trains on entire training set
"""
def __init__(self, cached_features=True):
self.model = None
self.classes = numpy.array([0, 1])
self.cached_features = cached_features
self.feature_extractor = CachedFeatureExtractor("working/models/train_batches")
self.test_feature_extractor = CachedFeatureExtractor("working/models/test_batches")
def fit(self):
print " Fitting ", self.__class__.__name__, "..."
counter = 0
for (X_train, y_train) in self.feature_extractor.nextBatch():
self._fit_internal(X_train, y_train)
def _fit_internal(self, X_train, y_train):
self.model.partial_fit(X_train, y_train, classes=self.classes)
def predict_proba(self):
print " Predicting ", self.__class__.__name__, "..."
counter = 0
y_ans = numpy.array([])
for (X_test, y_test_ignored) in self.test_feature_extractor.nextBatch():
y_ans = numpy.append(y_ans, self._predict_internal(X_test))
return y_ans
def _predict_internal(self, X_test):
pass
def predict_and_dump(self, submission_file):
pass
def dump(self, filename):
print "Dumping model to", filename, "..."
with open(filename, "wb") as f:
cPickle.dump(self, f)
class LSTMModel(base_model.BaseModel):
def __init__(self, cached_features):
base_model.BaseModel.__init__(self, cached_features)
self.weights_file = "working/models/lstm_model.pickle.weights"
self.n_epochs = 4
self.padding_size = 256
self.max_features = 43980
self.batch_size = 32
self.model = self.SetupModel()
self.X_train = None
self.y_train = numpy.array([])
self.feature_extractor = CachedFeatureExtractor("working/models/sequence_train_batches")
self.test_feature_extractor = CachedFeatureExtractor("working/models/sequence_test_batches")
def SetupModel(self):
model = Sequential()
model.add(Embedding(self.max_features, 128))
model.add(LSTM(128, 128))
model.add(Dropout(0.5))
model.add(Dense(128, 1))
model.add(Activation('sigmoid'))
# try using different optimizers and different optimizer configs
model.compile(loss='binary_crossentropy', optimizer='adam', class_mode="binary")
return model
def _fit_internal(self, X_train, y_train):
if self.X_train is None:
self.X_train = sequence.pad_sequences(X_train, maxlen=self.padding_size)
else:
self.X_train = numpy.vstack((self.X_train, sequence.pad_sequences(X_train, maxlen=self.padding_size)))
self.y_train = numpy.append(self.y_train, y_train)
print self.X_train.shape, self.y_train.shape
def fit(self):
print "Fitting ", self.__class__.__name__, "..."
counter = 0
for (X_train, y_train) in self.feature_extractor.nextBatch():
print " Adding batch ", counter, "...",
self._fit_internal(X_train, y_train)
counter += 1
self.model.fit(self.X_train, self.y_train, batch_size=self.batch_size, nb_epoch=self.n_epochs, show_accuracy=True)
def predict_proba(self):
print " Predicting ", self.__class__.__name__, "..."
print "Creating model using SetupModel..."
self.model = self.SetupModel()
print "Loading model weights from ", self.weights_file, "..."
self.model.load_weights(self.weights_file)
counter = 0
y_ans = numpy.array([])
#self.test_feature_extractor = CachedFeatureExtractor("working/models/sequence_test_batches")
for (X_test, y_test_ignored) in self.test_feature_extractor.nextBatch():
y_ans = numpy.append(y_ans, self._predict_internal(X_test))
counter += 1
return y_ans
def _predict_internal(self, X_test):
X_new = sequence.pad_sequences(X_test, maxlen=self.padding_size)
return self.model.predict(X_new, batch_size=self.batch_size)[:, 0]
def dump(self, filename):
print "saving weights to: ", self.weights_file
self.model.save_weights(self.weights_file, overwrite=True)
self.model = None
print "Dumping model to ", filename, "..."
with open(filename, "wb") as f:
cPickle.dump(self, f)
def __init__(self, cached_features=True):
self.model = None
self.classes = numpy.array([0, 1])
self.cached_features = cached_features
self.feature_extractor = CachedFeatureExtractor("working/models/train_batches")
self.test_feature_extractor = CachedFeatureExtractor("working/models/test_batches")
