def test_training_classification(spark_context, mode, parameter_server_mode,
mnist_data, classification_model):
# Define basic parameters
batch_size = 64
epochs = 10
# Load data
x_train, y_train, x_test, y_test = mnist_data
x_train = x_train[:1000]
y_train = y_train[:1000]
sgd = SGD(lr=0.1)
classification_model.compile(sgd, 'categorical_crossentropy', ['acc'])
# Build RDD from numpy features and labels
rdd = to_simple_rdd(spark_context, x_train, y_train)
# Initialize SparkModel from keras model and Spark context
spark_model = SparkModel(classification_model,
frequency='epoch',
mode=mode,
parameter_server_mode=parameter_server_mode,
port=4000 + random.randint(0, 500))
# Train Spark model
spark_model.fit(rdd,
epochs=epochs,
batch_size=batch_size,
verbose=0,
validation_split=0.1)
# run inference on trained spark model
predictions = spark_model.predict(x_test)
# run evaluation on trained spark model
evals = spark_model.evaluate(x_test, y_test)
# assert we can supply rdd and get same prediction results when supplying numpy array
test_rdd = spark_context.parallelize(x_test)
assert [np.argmax(x) for x in predictions
] == [np.argmax(x) for x in spark_model.predict(test_rdd)]
# assert we get the same prediction result with calling predict on keras model directly
assert [np.argmax(x) for x in predictions] == [
np.argmax(x) for x in spark_model.master_network.predict(x_test)
]
# assert we get the same evaluation results when calling evaluate on keras model directly
assert isclose(evals[0],
spark_model.master_network.evaluate(x_test, y_test)[0],
abs_tol=0.01)
assert isclose(evals[1],
spark_model.master_network.evaluate(x_test, y_test)[1],
abs_tol=0.01)
def main():
gps_files = glob.glob('../data/prototype/**/gps_points.csv')
trip_files = glob.glob('../data/prototype/**/gps_trips.csv')
file_results = process_file(trip_file = trip_files[0], gps_file = gps_files[0])
seq_results = build_seq(input_df = file_results['df'], unique_trips = file_results['unique_trips'])
X = seq_results['x']
y = seq_results['y']
print('Bulding training data from files..')
for i in range(1, len(gps_files)):
file_results = process_file(trip_file = trip_files[i], gps_file = gps_files[i])
seq_results = build_seq(input_df = file_results['df'], unique_trips = file_results['unique_trips'])
X = np.vstack((X, seq_results['x']))
y = np.vstack((y, seq_results['y']))
x_train, x_val, y_train, y_val = train_test_split(X, y, random_state=1, train_size=0.8)
rdd = to_simple_rdd(sc, x_train, y_train)
model = build_model()
spark_model = SparkModel(model, frequency='epoch', mode='asynchronous')
spark_model.fit(rdd, epochs=5, batch_size=32, verbose=0, validation_split=0.1)
# model.fit(x_train, y_train, epochs=5, validation_data=(x_val, y_val))
y_pred = spark_model.predict(x_val)
acc = sum(np.argmax(y_pred, axis=1) == np.argmax(y_val, axis=1)) / y_pred.shape[0]
print("Validation Accuracy: {number:.{digits}f}%".format(number=(acc*100), digits=2))
def test_training_regression(spark_context, mode, parameter_server_mode,
boston_housing_dataset, regression_model):
x_train, y_train, x_test, y_test = boston_housing_dataset
rdd = to_simple_rdd(spark_context, x_train, y_train)
# Define basic parameters
batch_size = 64
epochs = 10
sgd = SGD(lr=0.0000001)
regression_model.compile(sgd, 'mse', ['mae'])
spark_model = SparkModel(regression_model,
frequency='epoch',
mode=mode,
parameter_server_mode=parameter_server_mode,
port=4000 + random.randint(0, 500))
# Train Spark model
spark_model.fit(rdd,
epochs=epochs,
batch_size=batch_size,
verbose=0,
validation_split=0.1)
# run inference on trained spark model
predictions = spark_model.predict(x_test)
# run evaluation on trained spark model
evals = spark_model.evaluate(x_test, y_test)
# assert we can supply rdd and get same prediction results when supplying numpy array
test_rdd = spark_context.parallelize(x_test)
assert all(
np.isclose(x, y, 0.01)
for x, y in zip(predictions, spark_model.predict(test_rdd)))
# assert we get the same prediction result with calling predict on keras model directly
assert all(
np.isclose(x, y, 0.01) for x, y in zip(
predictions, spark_model.master_network.predict(x_test)))
# assert we get the same evaluation results when calling evaluate on keras model directly
assert isclose(evals[0],
spark_model.master_network.evaluate(x_test, y_test)[0],
abs_tol=0.01)
assert isclose(evals[1],
spark_model.master_network.evaluate(x_test, y_test)[1],
abs_tol=0.01)
def test_training_custom_activation(mode, spark_context):
def custom_activation(x):
return sigmoid(x) + 1
model = Sequential()
model.add(Dense(1, input_dim=1, activation=custom_activation))
model.add(Dense(1, activation='sigmoid'))
sgd = SGD(lr=0.1)
model.compile(sgd, 'binary_crossentropy', ['acc'])
x_train = np.random.rand(1000)
y_train = np.zeros(1000)
x_test = np.random.rand(100)
y_test = np.zeros(100)
y_train[:500] = 1
rdd = to_simple_rdd(spark_context, x_train, y_train)
spark_model = SparkModel(model, frequency='epoch', mode=mode,
custom_objects={'custom_activation': custom_activation})
spark_model.fit(rdd, epochs=1, batch_size=16, verbose=0, validation_split=0.1)
assert spark_model.predict(x_test)
assert spark_model.evaluate(x_test, y_test)
class HAN(object):
"""
HAN model is implemented here.
"""
def __init__(self,
text,
labels,
pretrained_embedded_vector_path,
max_features,
max_senten_len,
max_senten_num,
embedding_size,
num_categories=None,
validation_split=0.2,
verbose=0):
"""Initialize the HAN module
Keyword arguments:
text -- list of the articles for training.
labels -- labels corresponding the given `text`.
pretrained_embedded_vector_path -- path of any pretrained vector
max_features -- max features embeddeding matrix can have. To more checkout https://keras.io/layers/embeddings/
max_senten_len -- maximum sentence length. It is recommended not to use the maximum one but the one that covers 0.95 quatile of the data.
max_senten_num -- maximum number of sentences. It is recommended not to use the maximum one but the one that covers 0.95 quatile of the data.
embedding_size -- size of the embedding vector
num_categories -- total number of categories.
validation_split -- train-test split.
verbose -- how much you want to see.
"""
try:
self.verbose = verbose
self.max_features = max_features
self.max_senten_len = max_senten_len
self.max_senten_num = max_senten_num
self.embed_size = embedding_size
self.validation_split = validation_split
self.embedded_dir = pretrained_embedded_vector_path
self.text = pd.Series(text)
self.categories = pd.Series(labels)
self.classes = self.categories.unique().tolist()
# Initialize default hyperparameters
# You can change it using `set_hyperparameters` function
self.hyperparameters = {
'l2_regulizer': None,
'dropout_regulizer': None,
'rnn': LSTM,
'rnn_units': 150,
'dense_units': 200,
'activation': 'softmax',
'optimizer': 'adam',
'metrics': ['acc'],
'loss': 'categorical_crossentropy'
}
if num_categories is not None:
assert (num_categories == len(self.classes))
assert (self.text.shape[0] == self.categories.shape[0])
self.data, self.labels = self.preprocessing()
self.x_train, self.y_train, self.x_val, self.y_val = self.split_dataset(
)
self.embedding_index = self.add_glove_model()
self.set_model()
except AssertionError:
print('Input and label data must be of same size')
# Implement this after you have seen all the different kinds of errors
# try:
# conf = SparkConf().setAppName('HANMusicClassifier').setMaster('')
# self.sc = SparkContext(conf=conf)
# except Error:
conf = SparkConf().setAppName('HANMusicClassifier')
self.sc = SparkContext(conf=conf)
def set_hyperparameters(self, tweaked_instances):
"""Set hyperparameters of HAN model.
Keywords arguemnts:
tweaked_instances -- dictionary of all those keys you want to change
"""
for key, value in tweaked_instances.items():
if key in self.hyperparameters:
self.hyperparameters[key] = value
else:
raise KeyError(key + ' does not exist in hyperparameters')
self.set_model()
def show_hyperparameters(self):
"""To check the values of all the current hyperparameters
"""
print('Hyperparameter\tCorresponding Value')
for key, value in self.hyperparameters.items():
print(key, '\t\t', value)
def clean_string(self, string):
"""
Tokenization/string cleaning for dataset
Every dataset is lower cased except
"""
string = re.sub(r"\\", "", string)
string = re.sub(r"\'", "", string)
string = re.sub(r"\"", "", string)
return string.strip().lower()
def add_dataset(self, text, labels):
try:
self.text = pd.concat([self.text, pd.Series(text)])
self.categories = pd.concat([self.categories, pd.Series(labels)])
assert (len(self.classes) == self.categories.unique().tolist())
except AssertionError:
print("New class cannot be added in this manner")
def preprocessing(self):
"""Preprocessing of the text to make it more resonant for training
"""
paras = []
labels = []
texts = []
for idx in range(self.text.shape[0]):
text = self.clean_string(self.text[idx])
texts.append(text)
sentences = tokenize.sent_tokenize(text)
paras.append(sentences)
tokenizer = Tokenizer(num_words=self.max_features, oov_token=True)
tokenizer.fit_on_texts(texts)
data = np.zeros((len(texts), self.max_senten_num, self.max_senten_len),
dtype='int32')
for i, sentences in enumerate(paras):
for j, sent in enumerate(sentences):
if j < self.max_senten_num:
wordTokens = text_to_word_sequence(sent)
k = 0
for _, word in enumerate(wordTokens):
if k < self.max_senten_len and word in tokenizer.word_index and tokenizer.word_index[
word] < self.max_features:
data[i, j, k] = tokenizer.word_index[word]
k = k + 1
self.word_index = tokenizer.word_index
if self.verbose == 1:
print('Total %s unique tokens.' % len(self.word_index))
labels = pd.get_dummies(self.categories)
if self.verbose == 1:
print('Shape of data tensor:', data.shape)
print('Shape of labels tensor:', labels.shape)
assert (len(self.classes) == labels.shape[1])
assert (data.shape[0] == labels.shape[0])
return data, labels
def split_dataset(self):
indices = np.arange(self.data.shape[0])
np.random.shuffle(indices)
self.data = self.data[indices]
self.labels = self.labels.iloc[indices]
nb_validation_samples = int(self.validation_split * self.data.shape[0])
x_train = self.data[:-nb_validation_samples]
y_train = self.labels[:-nb_validation_samples]
x_val = self.data[-nb_validation_samples:]
y_val = self.labels[-nb_validation_samples:]
if self.verbose == 1:
print(
'Number of positive and negative reviews in traing and validation set'
)
print(y_train.columns.tolist())
print(y_train.sum(axis=0).tolist())
print(y_val.sum(axis=0).tolist())
return x_train, y_train, x_val, y_val
def get_model(self):
"""
Returns the HAN model so that it can be used as a part of pipeline
"""
return self.model
def add_glove_model(self):
"""
Read and save Pretrained Embedding model
"""
embeddings_index = {}
try:
f = open(self.embedded_dir)
for line in f:
values = line.split()
word = values[0]
coefs = np.asarray(values[1:], dtype='float32')
assert (coefs.shape[0] == self.embed_size)
embeddings_index[word] = coefs
f.close()
except OSError:
print('Embedded file does not found')
exit()
except AssertionError:
print(
"Embedding vector size does not match with given embedded size"
)
return embeddings_index
def get_embedding_matrix(self):
"""
Returns Embedding matrix
"""
embedding_matrix = np.random.random(
(len(self.word_index) + 1, self.embed_size))
absent_words = 0
for word, i in self.word_index.items():
embedding_vector = self.embedding_index.get(word)
if embedding_vector is not None:
# words not found in embedding index will be all-zeros.
embedding_matrix[i] = embedding_vector
else:
absent_words += 1
if self.verbose == 1:
print('Total absent words are', absent_words, 'which is',
"%0.2f" % (absent_words * 100 / len(self.word_index)),
'% of total words')
return embedding_matrix
def get_embedding_layer(self):
"""
Returns Embedding layer
"""
embedding_matrix = self.get_embedding_matrix()
return Embedding(len(self.word_index) + 1,
self.embed_size,
weights=[embedding_matrix],
input_length=self.max_senten_len,
trainable=False)
def set_model(self):
"""
Set the HAN model according to the given hyperparameters
"""
if self.hyperparameters['l2_regulizer'] is None:
kernel_regularizer = None
else:
kernel_regularizer = regularizers.l2(
self.hyperparameters['l2_regulizer'])
if self.hyperparameters['dropout_regulizer'] is None:
dropout_regularizer = 1
else:
dropout_regularizer = self.hyperparameters['dropout_regulizer']
word_input = Input(shape=(self.max_senten_len, ), dtype='float32')
word_sequences = self.get_embedding_layer()(word_input)
word_lstm = Bidirectional(self.hyperparameters['rnn'](
self.hyperparameters['rnn_units'],
return_sequences=True,
kernel_regularizer=kernel_regularizer))(word_sequences)
word_dense = TimeDistributed(
Dense(self.hyperparameters['dense_units'],
kernel_regularizer=kernel_regularizer))(word_lstm)
word_att = AttentionWithContext()(word_dense)
wordEncoder = Model(word_input, word_att)
sent_input = Input(shape=(self.max_senten_num, self.max_senten_len),
dtype='float32')
sent_encoder = TimeDistributed(wordEncoder)(sent_input)
sent_lstm = Bidirectional(self.hyperparameters['rnn'](
self.hyperparameters['rnn_units'],
return_sequences=True,
kernel_regularizer=kernel_regularizer))(sent_encoder)
sent_dense = TimeDistributed(
Dense(self.hyperparameters['dense_units'],
kernel_regularizer=kernel_regularizer))(sent_lstm)
sent_att = Dropout(dropout_regularizer)(
AttentionWithContext()(sent_dense))
preds = Dense(len(self.classes))(sent_att)
self.model = Model(sent_input, preds)
self.model.compile(loss=self.hyperparameters['loss'],
optimizer=self.hyperparameters['optimizer'],
metrics=self.hyperparameters['metrics'])
self.spark_model = SparkModel(self.model,
frequency='epoch',
mode='asynchronous')
# Currently cannot plot learning curve
def train_model(self,
rdd,
epochs,
batch_size,
verbose=1,
validation_split=0.1):
"""Training the model
rdd -- The actual data
epochs -- Total number of epochs
batch_size -- size of a batch
verbose -- Whether or not we want verbose feedback
validation_split -- What percentage of the data from the rdd is actually used as a validation set
"""
self.spark_model.fit(self,
rdd,
epochs=epochs,
batch_size=batch_size,
verbose=verbose,
validation_split=validation_split)
def predict(self, rdd):
self.spark_model.predict(rdd)
def plot_results(self):
"""
Plotting learning curve of last trained model.
"""
# summarize history for accuracy
plt.subplot(211)
plt.plot(self.history.history['acc'])
plt.plot(self.history.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
# summarize history for loss
plt.subplot(212)
plt.plot(self.history.history['val_loss'])
plt.plot(self.history.history['loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
time.sleep(10)
plt.close()
spark_model.fit(rdd, epochs=50, batch_size=500, validation_split=0.01)
LOGGER.info('Spark model trained')
LOGGER.info(model.summary())
test_data = scaled_data[training_data_len - 30:, :]
x_test = []
y_test = dataset[training_data_len:, :]
y_test = dataset[training_data_len:, :]
for i in range(30, len(test_data)):
x_test.append(test_data[i - 30:i, 0])
x_test = np.array(x_test)
x_test = np.reshape(x_test, (x_test.shape[0], x_test.shape[1], 1))
predictions = spark_model.predict(x_test)
predictions = scaler.inverse_transform(predictions)
rmse = np.sqrt(np.mean(((predictions - y_test)**2)))
rmse
train = data_pd[:training_data_len]
valid = data_pd[training_data_len:]
valid['Predictions'] = predictions
plt.figure(figsize=(16, 8))
plt.title('Model prediction on EUR/HUF')
plt.xlabel('Date', fontsize=18)
plt.ylabel('Close Price EUR/HUF', fontsize=18)
plt.plot(train['Close'])
plt.plot(valid[['Close', 'Predictions']])
plt.legend(['Train', 'Val', 'Predictions'], loc='lower right')
model = Sequential()
model.add(Dense(18, input_dim=26))
model.add(Activation('sigmoid'))
model.add(Dense(6))
model.add(Activation('sigmoid'))
model.add(Dense(1))
model.add(Activation('sigmoid'))
spark = SparkSession.builder.appName('ElephasTest').getOrCreate()
rdd = to_simple_rdd(spark.sparkContext, train, y_train)
sgd = SGD(lr=0.1)
adagrad = elephas_optimizers.Adagrad()
spark_model = SparkModel(spark.sparkContext,
model,
optimizer=adagrad,
frequency='epoch',
mode='asynchronous',
master_loss='mse',
num_workers=2, master_optimizer=sgd)
# Train Spark model
spark_model.train(rdd, nb_epoch=nb_epoch, batch_size=batch_size, verbose=2, validation_split=0.1)
# Evaluate Spark model by evaluating the underlying model
score = spark_model.master_network.evaluate(test.values, y_test, verbose=2)
print('Test accuracy:', score[1])
print spark_model.predict(test.values)
print y_test
#---(i.e. in training each worker will train on part of the data)
rdd = to_simple_rdd(sc, X_train, y_train)
#---Initialize SparkModel from Keras model and Spark context
#---there are two optimizers needed:
sgd = SGD(lr=0.1) #<---the master optimizer
adagrad = elephas_optimizers.Adagrad() #<---the elephas opimizer
spark_model = SparkModel(sc,
model,
optimizer=adagrad,
frequency='epoch',
mode='asynchronous',
num_workers=args.N_workers,
master_optimizer=sgd)
#---Train Spark model
spark_model.train(rdd,
nb_epoch=args.nb_epoch,
batch_size=args.batch_size,
verbose=1,
validation_split=0.25)
#---Evaluate Spark model by evaluating the underlying Keras master model
pred = spark_model.predict(X_test)
print np.shape(pred)
print np.shape(y_test)
acc = accuracy_score([np.argmax(y) for y in y_test],
[np.argmax(p) for p in pred])
print "--->test accuracy: ", acc
print "--->number of workers: ", args.N_workers
print "--->time: ", time.time() - start_time
class KerasNeuralNetworkSpark(object):
def __init__(self, layers, spark, batch_size=64, epoch=10, num_workers=2, predictionCol='prediction',
labelCol='target', featuresCol='feature'):
self._batch_size = batch_size
self._epoch = epoch
self._model = None
self._spark = spark
self._labels = labelCol
self._features = featuresCol
self._prediction = predictionCol
self._layers = layers
self._worker_num = num_workers
self._build_model()
def _build_model(self):
model = Sequential()
adam = elephas_optimizers.Adam()
layers = self._layers
model.add(Dense(layers[1], input_dim=layers[0], init='normal', activation='relu'))
for i in range(2, len(layers) - 1):
model.add(Dense(layers[i], activation='relu'))
model.add(Dense(layers[-1], activation='sigmoid'))
self._model = SparkModel(self._spark.sparkContext, model,
optimizer=adam,
frequency='epoch',
mode='asynchronous',
master_loss='mse',
num_workers=self._worker_num)
def fit(self, df):
if hasattr(self._model, 'server'):
self._model.server.terminate()
pdf = df.toPandas()
rdd = to_simple_rdd(self._spark.sparkContext, pdf[self._features], pdf[self._labels])
self._model.train(rdd, self._epoch, self._batch_size, 0, 0.1)
def transform(self, df):
pdf = df.toPandas()
# df.write.save('test_df.parquet')
pnparray = pdf[self._features].values
container = np.zeros((pnparray.shape[0], len(pnparray[0])))
for i in range(pnparray.shape[0]):
container[i, :] = pnparray[i][:]
result = self._model.predict(container)
pdf[self._prediction] = result
# import pickle
# with open('ann_result.p', 'w') as f:
# pickle.dump(result, f)
# result_df = pd.DataFrame(pdf
new_df = self._spark.createDataFrame(pdf)
# df.join(new_df)
return new_df
def stop_server(self):
if hasattr(self._model, 'server') and hasattr(self._model.server, 'terminate'):
self._model.server.terminate()