def __init__(self, env):
# Feature preprocessing: Normalize to zero mean and unit variance
# We use a few samples from the observation space to do this
observation_examples = np.array(
[env.observation_space.sample() for _ in range(10000)],
dtype='float64')
self.scaler = preprocessing.StandardScaler()
self.scaler.fit(observation_examples)
# Used to convert a state to a featurized represenation.
# We use RBF kernels with different variances to cover different parts of the space
self.featurizer = pipeline.FeatureUnion([
('rbf1', RBFSampler(gamma=5.0, n_components=100)),
('rbf2', RBFSampler(gamma=2.0, n_components=100)),
('rbf3', RBFSampler(gamma=1.0, n_components=100)),
('rbf4', RBFSampler(gamma=0.5, n_components=100)),
])
self.featurizer.fit(self.scaler.transform(observation_examples))
self.models = []
for _ in range(env.action_space.n):
model = SGDRegressor(learning_rate='constant')
model.partial_fit([self.featurize_state(env.reset())], [0])
self.models.append(model)
def __init__(self, env):
"""
SGD function approximator, with preprocessing steps from:
https://github.com/dennybritz/reinforcement-learning/blob/master/FA/Q-Learning%20with%20Value%20Function%20Approximation%20Solution.ipynb
"""
# Feature preprocessing: Normalize to zero mean and unit variance
# We use a few samples from the observation space to do this
observation_examples = np.array(
[env.observation_space.sample() for _ in range(10000)], dtype="float64"
)
self.scaler = preprocessing.StandardScaler()
self.scaler.fit(observation_examples)
# Used to convert a state to a featurized represenation.
# We use RBF kernels with different variances to cover different parts of the space
self.featurizer = pipeline.FeatureUnion(
[
("rbf1", RBFSampler(gamma=5.0, n_components=100)),
("rbf2", RBFSampler(gamma=2.0, n_components=100)),
("rbf3", RBFSampler(gamma=1.0, n_components=100)),
("rbf4", RBFSampler(gamma=0.5, n_components=100)),
]
)
self.featurizer.fit(self.scaler.transform(observation_examples))
self.models = []
for _ in range(env.action_space.n):
model = SGDRegressor(learning_rate="constant")
model.partial_fit([self.featurize_state(env.reset())], [0])
self.models.append(model)
def create_estimator(ml_obj, numeric_features, cat_features, date_features):
estimator = pipeline.Pipeline(steps=[
('Feature_processing',
pipeline.FeatureUnion(transformer_list=[
('Numeric_features',
pipeline.Pipeline(steps=[(
'selecting',
preprocessing.FunctionTransformer(
lambda data: data[:, numeric_features], validate=True)),
('scaling',
preprocessing.StandardScaler(
with_mean=0., with_std=1))])),
('Categical_features',
pipeline.Pipeline(steps=[(
'selecting',
preprocessing.FunctionTransformer(
lambda data: data[:, cat_features], validate=True)),
('hot_encoding',
preprocessing.OneHotEncoder(
handle_unknown='ignore'))])),
('Date_features',
pipeline.Pipeline(steps=[(
'selecting',
preprocessing.FunctionTransformer(
lambda data: data[:, date_features], validate=True)),
('hot_encoding',
preprocessing.OneHotEncoder(
handle_unknown='ignore'))]))
])), ('Model_fitting', ml_obj)
])
return estimator
#TODO:
#make custom score
def get_default_features(self):
tfidfvect = TfidfVectorizer()
#features = skpipeline.FeatureUnion([('word_tfidf', tfidfvect)])
features = skpipeline.FeatureUnion([
('word_tfidf', tfidfvect),
])
return features
def main():
env = LunarLander()
# Add transformers for tile coding or extra features.
transformer = pipeline.FeatureUnion([
# ('scaler', preprocessing.StandardScaler()),
# ('square', preprocessing.FunctionTransformer(lambda x: x**2, validate=False)),
# ('dummy', DummyTransformer()),
# ('poly', preprocessing.PolynomialFeatures(2)),
# ('cos', preprocessing.FunctionTransformer(np.cos, validate=False)),
# ('inverter', preprocessing.FunctionTransformer(lambda x: 1. /(x + 1.), validate=False)),
# ('quantile', preprocessing.KBinsDiscretizer(strategy='uniform', n_bins=20, encode='onehot')),
# ('quantile-poly', pipeline.Pipeline([
# ('poly', preprocessing.PolynomialFeatures(2, interaction_only=True)),
# ('quantile', preprocessing.KBinsDiscretizer(strategy='quantile', n_bins=20, encode='onehot-dense', )),
# ])),
# ('quantile', pipeline.Pipeline([
# # ('poly', preprocessing.PolynomialFeatures(2)),
# ('quantile', preprocessing.KBinsDiscretizer(strategy='uniform', n_bins=200, encode='ordinal', )),
# # ('ohe', preprocessing.OneHotEncoder(sparse=False, categories='auto'))
#
# ])),
# ('power', preprocessing.PowerTransformer()),
])
s = env.reset()
# a = 1/(1000*s.T@s)
print('Learning rate:', LR)
agent = GMCAgent(lr=LR,
init_epsilon=EPS,
max_steps=MAX_STEPS,
gamma=GAMMA,
threshold=0.0,
transformer=None,
success_count=SUCCESS_COUNT,
success_criteria=SUCCESS_CRITERIA)
# agent = SemiGradientAgent(lr=LR, init_epsilon=EPS, max_steps=800, gamma=GAMMA, threshold=0.0,
# transformer=None, success_count=3, success_criteria=220)
# agent = EpisodicSemiGradient(lr=a, init_epsilon=0.3, max_steps=800, gamma=0.9999, threshold=0.0,
# transformer=None, success_count=3, success_criteria=220)
agent.fit(env, render_train=False, verbose=True, episodes=10000)
agent.land(env, verbose=False)
now = datetime.now().strftime("%Y-%m-%d")
filename = f'weights/trained_agent_weights_{agent.alpha}_{agent.max_steps}_{now}'
np.save(filename, agent.get_weights())
print("Weights saved to", filename)
def main():
books = datasets.load_files("data/Book/",
shuffle=True,
encoding="ISO-8859-1",
random_state=1337)
X_train, X_test, y_train, y_test = model_selection.train_test_split(
books.data, books.target, test_size=0.10)
model = pipeline.Pipeline(
[('union',
pipeline.FeatureUnion(transformer_list=[
('other_features',
util_q1.AddOtherFeatures(feature_to_add="pos_neg_count")),
('text_data',
pipeline.Pipeline([
('remove_words',
util_q1.RemoveWords(words_to_remove="none")),
("normalisation",
util_q1.NormaliseWords(normalise_type="lemmatize")),
("preprocess",
util_q1.PreprocessData(attribute="frequency_filtering",
attribute_values="tf-idf")),
]))
])),
("classifier",
GradientBoostingClassifier(n_estimators=60,
max_features="sqrt",
subsample=0.8))])
scoring = {
"accuracy": "accuracy",
"recall": "recall",
"precision": "precision"
}
grid_search_model = model_selection.GridSearchCV(
model, {
"classifier__max_depth": range(5, 11, 5),
"classifier__min_samples_split": range(5, 11, 5)
},
n_jobs=-1,
verbose=10,
scoring=scoring,
refit=False)
grid_search_model.fit(X_train, y_train)
joblib.dump(grid_search_model, "outputs/gridsearch_xgboost_aws.pkl")
def main():
model = pipeline.Pipeline([
("features",
pipeline.FeatureUnion(
transformer_list=[("other_features",
AddOtherFeatures(feature_to_add="None")),
("text_data",
pipeline.Pipeline((
"remove_words",
RemoveWords(words_to_remove="None"),
"normalize_words",
NormalizeWords(normalize_type="None"),
"vectorize_text",
VectorizeText(vectorize_type="None"),
"reduce_dimension",
ReduceDimension(reduction_type="None"),
"normalize_features",
NormalizeFeatures(
normalize_type="None"))))])),
("classifier", CurrentModel(model_name="knn"))
])
scoring = calculate_score()
grid_search_model = model_selection.GridSearchCV(model, {
"features__other_features__feature_to_add": ["None"],
"features__text_data__remove_words":
["None", "tool_words", "closed_class", "tool_words_and_closed_class"],
"features__text_data__normalize_words":
["None", "Stemming", "Lemmatization"],
"features__text_data__vectorize_text":
["None", "Presence", "Frequency", "td_idf"],
"features__text_data__reduce_dimension": ["None", "PCA"],
"features__text_data__normalize_features": ["None", "min_max_scale"]
},
n_jobs=-1,
verbose=10,
scoring=scoring,
refit=False)
grid_search_model.fit(X_train, y_train)
pickle.dump(grid_search_model, open("fitted_pipeline", "wb"))
def build_model():
"""
Build an NLP pipeline for multi-label text classification.
"""
# text processing and model pipeline
pipeline = skpipe.Pipeline([
('nlp',
skpipe.FeatureUnion([
('tfif',
skpipe.Pipeline([('feat',
skfet.TfidfVectorizer(strip_accents='unicode',
tokenizer=tokenize)),
('lsa',
skdec.TruncatedSVD(n_components=200,
algorithm='arpack'))])),
('uppr', RatioUpperExtractor()), ('verb', CountVerbExtractor()),
('noun', RatioNounExtractor())
])), ('norm', skprep.StandardScaler()),
('clf',
MLPClassifier(activation='logistic',
learning_rate='adaptive',
early_stopping=True,
random_state=RANDOM_SEED,
verbose=1))
])
# define grid search parameters
params = {
'clf__learning_rate_init': [5e-3, 7.5e-3, 1e-2],
'clf__hidden_layer_sizes': [(100), (200, ), (300, )]
}
# instantiate GridSearchCV object
cv = skms.GridSearchCV(estimator=pipeline,
param_grid=params,
n_jobs=-1,
refit=True,
return_train_score=True)
return cv
def get_estimator(self):
binary = ('binary_variables_processing',
preprocessing.FunctionTransformer(
lambda data: data[:, Model.binary_data_indices],
validate=True))
categorial = (
'categorical_variables_processing',
pipeline.Pipeline(
steps=[(
'selecting',
preprocessing.FunctionTransformer(
lambda data: data[:, Model.categorical_data_indices],
validate=True)),
('hot_encoding',
preprocessing.OneHotEncoder(handle_unknown='ignore',
sparse=False))]))
estimator = pipeline.Pipeline(
steps=[('feature_processing',
pipeline.FeatureUnion(
transformer_list=[binary, categorial])
), ('model_fitting', self.regressor)])
return estimator
def __init__(self, env, use_kernel=False, **agent_params):
self.env = env
self.use_kernel = use_kernel
if use_kernel:
# Sample feature space and define scaler to detrend data
observation_samples = np.array(
[env.observation_space.sample() for x in range(10000)])
self.detrend = preprocessing.StandardScaler()
self.detrend.fit(observation_samples)
# Use detrended data to generate feature space with RBF kernels
self.featurizer = pipeline.FeatureUnion([
("rbf1", RBFSampler(gamma=3.0, n_components=100)),
("rbf2", RBFSampler(gamma=2.0, n_components=100)),
("rbf3", RBFSampler(gamma=1.0, n_components=100)),
("rbf4", RBFSampler(gamma=0.5, n_components=100))
])
self.featurizer.fit(self.detrend.transform(observation_samples))
# Generate linear value function model for each action in our action space
self.models = []
initReward = np.array(0)
for k in range(env.action_space.n):
self.models.append(
linear_model.SGDRegressor(learning_rate="constant"))
random_features = self.map_to_features(self.env.reset())
self.models[k].partial_fit(random_features.reshape(1, -1),
initReward.ravel())
self.agent_params = {
"epsilon_min": 0.01,
"decay_rate": 0.02,
"discount": 0.99,
"iter": 1000
}
self.agent_params.update(agent_params)
])
#transformed_data=cat_feature_pipeline.fit_transform(X_train[['ENRL_CERT_NBR']])
num_feature_pipeline = pipeline.Pipeline([
('imputation', impute.SimpleImputer()),
('standardscalar', preprocessing.StandardScaler())
])
#transformed_data=num_feature_pipeline.fit_transform(X_train[['TOT_BLNG_AMT']])
feature_preprocessing = compose.ColumnTransformer(
[('cat_feature_pipeline', cat_feature_pipeline, cat_features_list),
('num_feature_pipeline', num_feature_pipeline, num_features_list)],
n_jobs=10)
features_pipeline = pipeline.FeatureUnion(
[('pca_selector', decomposition.PCA(n_components=0.90)),
('et_selector',
feature_selection.SelectFromModel(ensemble.ExtraTreesClassifier()))],
n_jobs=20)
classifier = tree.DecisionTreeClassifier()
#build complete pipeline with feature selection and ml algorithms
complete_pipeline = PMMLPipeline([('preprocess', feature_preprocessing),
('zv_filter',
feature_selection.VarianceThreshold()),
('features', features_pipeline),
('tree', classifier)])
pipeline_grid = {}
grid_estimator = model_selection.GridSearchCV(complete_pipeline,
pipeline_grid,
scoring="accuracy",
def _generate_feature_extraction_pipeline(self):
lang = self.feature_config.lang
feature_weights = self.feature_config.weights
prep_params = self.feature_config.prepchoice
# features found in the processed tokens
preprocessor = prep.Preprocessor(
lang=lang,
stopword=prep_params.stopword,
more_stopwords=prep_params.more_stopwords,
spellcheck=prep_params.spellcheck,
stemming=prep_params.stemming,
remove_numbers=prep_params.remove_numbers,
deasciify=prep_params.deasciify,
remove_punkt=prep_params.remove_punkt,
lowercase=prep_params.lowercase)
tfidfvect = sktext.TfidfVectorizer(
tokenizer=prep.identity,
preprocessor=None,
lowercase=False,
use_idf=prep_params.use_idf,
ngram_range=prep_params.wordngramrange,
max_features=prep_params.nmaxfeature)
polpipe3 = toktrans.get_lexicon_count_pipeline(tokenizer=prep.identity,
lexicontype=lang)
token_weights = dict(tfidfvect=feature_weights["word_tfidf"],
polpipe3=feature_weights["lexicon_count"])
token_transformers_dict = dict(
tfidfvect=
tfidfvect, # not to lose above integrity if we change variable names
polpipe3=polpipe3)
token_transformers = [(k, v)
for k, v in token_transformers_dict.items()]
tokenbasedpipe = skpipeline.Pipeline([
('preprocessor', preprocessor),
# ('nadropper', tbt.DropNATransformer()),
('union1',
skpipeline.FeatureUnion(transformer_list=token_transformers,
transformer_weights=token_weights)),
])
charngramvect = sktext.TfidfVectorizer(
analyzer='char_wb',
ngram_range=prep_params.charngramrange,
lowercase=False)
polpipe1 = txtrans.get_polylglot_polarity_count_pipe(lang)
polpipe2 = txtrans.get_polylglot_polarity_value_pipe(lang)
text_weights = dict(
charngramvect=feature_weights["char_tfidf"], # @TODO hardcoded
polpipe1=feature_weights["polyglot_count"],
polpipe2=feature_weights["polyglot_value"])
text_transformers_dict = dict(charngramvect=charngramvect,
polpipe1=polpipe1,
polpipe2=polpipe2)
text_transformers = [(k, v) for k, v in text_transformers_dict.items()]
'''
textpipes = [('charngramvect', charngramvect),]
textpweights = {'charngramvect' : 1.5}
textpweights = dict(charngramvect = 1 if charngramvect else 0)
'''
textbasedpipe = skpipeline.Pipeline([(
'union2',
skpipeline.FeatureUnion(transformer_list=text_transformers,
transformer_weights=text_weights),
)])
print(text_weights)
final_transformers_dict = dict(tokenbasedpipe=tokenbasedpipe,
textbasedpipe=textbasedpipe)
final_transformers = [(k, v)
for k, v in final_transformers_dict.items()]
'''
#tweights = {k : 1 if v else 0 for k,v in final_transformers.items()}
check_zero = lambda x : 1 if sum(x) > 0 else 0
x = list(tokenbasedpipe.get_params(False).values())
print(len(x), x[0])
print(x[0][1]) # convert x[0] tuple to dict, then get transformer weights
print("**")
print(x,"\n--")
print(list(textbasedpipe.get_params(False).values()))
tweights = {k : check_zero(list(k.get_params(False).values())[0][0][1].get_params(False)["transformer_weights"].values())
for _, k in final_transformers_dict.items()}
'''
feature_union = skpipeline.FeatureUnion(
transformer_list=final_transformers,
# transformer_weights=tweights # weight assignment is not necessary as the number of features is small
)
return feature_union
def run_prep():
classifier = sklinear.SGDClassifier(loss='hinge',
penalty='l2',
alpha=1e-3,
n_iter=5,
random_state=42)
lang = "tr"
stopword_choice = True
more_stopwords_list = None
spellcheck_choice = False
stemming_choice = False
number_choice = False
deasc_choice = True
punct_choice = True
case_choice = True
ngramrange = (1, 2) # tuple
nmaxfeature = 10000 # int or None
norm = "l2"
use_idf = True
preprocessor = Preprocessor(lang=lang,
stopword=stopword_choice,
more_stopwords=more_stopwords_list,
spellcheck=spellcheck_choice,
stemming=stemming_choice,
remove_numbers=number_choice,
deasciify=deasc_choice,
remove_punkt=punct_choice,
lowercase=case_choice)
tfidfvect = TfidfVectorizer(tokenizer=identity,
preprocessor=None,
lowercase=False,
use_idf=use_idf,
ngram_range=ngramrange,
max_features=nmaxfeature)
keyword = "arıza"
apipe = tbt.get_keyword_pipeline(keyword)
keyword2 = "pstn"
pstnpipe = tbt.get_keyword_pipeline(keyword2)
polpipe1 = tbt.get_polylglot_polarity_count_pipe(lang)
polpipe2 = tbt.get_polylglot_polarity_value_pipe(lang)
polpipe3 = obt.get_lexicon_count_pipeline(tokenizer=identity)
tokenizedpipe = skpipeline.Pipeline([
('preprocessor', preprocessor),
('union1',
skpipeline.FeatureUnion(transformer_list=[
('vect', tfidfvect),
('polarity3', polpipe3),
])),
])
textbasedpipe = skpipeline.Pipeline([(
'union2',
skpipeline.FeatureUnion([
('has_ariza', apipe),
('has_pstn', pstnpipe),
('polarity1', polpipe1),
('polarity2', polpipe2),
]),
)])
model = skpipeline.Pipeline([
# ('preprocessor', preprocessor),
("union",
skpipeline.FeatureUnion(transformer_list=[
('tfidf', tokenizedpipe),
('txtpipe', textbasedpipe),
])),
('classifier', classifier),
])
t0 = time()
print("Read data")
instances, labels = get_data.get_data()
N = 100
instances, labels = corpus_io.select_N_instances(N, instances, labels)
# instances_train, instances_test, ytrain, ytest = cv.train_test_split(instances, labels, test_size=0.30, random_state=20)
print("Start classification\n..")
nfolds = 5
ypred = cv.cross_val_predict(model, instances, labels, cv=nfolds)
tc_utils.get_performance(labels, ypred, verbose=True)
t1 = time()
print("Classification took ", round(t1 - t0, 2), "sec.")
def _email_features_pipeline(lang,
stopword_choice=True,
more_stopwords_list=None,
spellcheck_choice=False,
stemming_choice=False,
number_choice=False,
deasc_choice=True,
punct_choice=True,
case_choice=True,
ngramrange=(1, 2), # tuple
nmaxfeature=10000, # int or None
norm="l2",
use_idf=True,
keywords=[], # ["arıza", "pstn"]
final_weights=dict(text_based=1, token_based=1)
):
# use a list of (pipeline, pipeline_name, weight)
# features found in the processed tokens
token_features = []
token_weights = {}
preprocessor = prep.Preprocessor(lang=lang,
stopword=stopword_choice, more_stopwords=more_stopwords_list,
spellcheck=spellcheck_choice,
stemming=stemming_choice,
remove_numbers=number_choice,
deasciify=deasc_choice,
remove_punkt=punct_choice,
lowercase=case_choice
)
tfidfvect = TfidfVectorizer(tokenizer=prep.identity, preprocessor=None, lowercase=False,
use_idf=use_idf, ngram_range=ngramrange, max_features=nmaxfeature)
tfidfvect_name = 'word_tfidfvect'
token_features.append((tfidfvect_name, tfidfvect))
token_weights[tfidfvect_name] = 1
# features found in the whole raw text
text_features = []
text_weights = {}
# charngramvect = TfidfVectorizer(analyzer='char_wb', ngram_range=(2, 2), lowercase=False)
# keyword presence features
if keywords:
for keyword in keywords:
keywordpipe = txbt.get_keyword_pipeline(keyword)
feature_name = "has_" + keyword
text_features.append((feature_name, keywordpipe))
text_weights[feature_name] = 1
tokenbasedpipe = skpipeline.Pipeline([('preprocessor', preprocessor),
# ('nadropper', tbt.DropNATransformer()),
('union1', skpipeline.FeatureUnion(
transformer_list=token_features ,
transformer_weights=token_weights
)),
])
textbasedpipe = skpipeline.Pipeline([('union2', skpipeline.FeatureUnion(
transformer_list=text_features,
transformer_weights=text_weights
),
)
])
#######
# add the feature pipes to final_features if all the component weights are non-zero.
########
check_zero_list = lambda x : 1 if sum(x) > 0 else 0
# l = [0,0,0] => check_zero(l) gives 0 and l=[0,0,1] => check_zero(l) gives 1.
final_features_dict = {}
tkweights = list(token_weights.values())
if(check_zero_list(tkweights) != 0):
final_features_dict["token_based"] = tokenbasedpipe
else:
final_weights["token_based"] = 0
txweights = list(text_weights.values())
if(check_zero_list(txweights) != 0):
final_features_dict["text_based"] = textbasedpipe
else:
final_weights["text_based"] = 0
final_features = list(final_features_dict.items())
fweights = list(final_weights.values())
if((check_zero_list(fweights) == 0) or (len(final_features) == 0)):
return None
'''
features = skpipeline.FeatureUnion(transformer_list=[
('tokenbasedfeatures', tokenbasedpipe),
('textbasedfeatures', textbasedpipe),
],
transformer_weights=final_weights)
'''
features = skpipeline.FeatureUnion(transformer_list=final_features,
transformer_weights=final_weights)
return features
def _tr_sentiment_features_pipeline(
lang="tr",
feature_weights={
"word_tfidf": 1,
"polyglot_value": 0,
"polyglot_count": 0,
"lexicon_count": 0,
"char_tfidf": 1
},
stopword_choice=True,
more_stopwords_list=None,
spellcheck_choice=False,
stemming_choice=False,
number_choice=False,
deasc_choice=True,
punct_choice=True,
case_choice=True,
word_ngramrange=(1, 2), # tuple
char_ngramrange=(2, 2),
nmaxfeature=10000, # int or None
norm="l2",
use_idf=True):
preprocessor = prep.Preprocessor(lang=lang,
stopword=stopword_choice,
more_stopwords=more_stopwords_list,
spellcheck=spellcheck_choice,
stemming=stemming_choice,
remove_numbers=number_choice,
deasciify=deasc_choice,
remove_punkt=punct_choice,
lowercase=case_choice)
tfidfvect = TfidfVectorizer(tokenizer=prep.identity,
preprocessor=None,
lowercase=False,
use_idf=use_idf,
ngram_range=word_ngramrange,
max_features=nmaxfeature)
polpipe3 = obt.get_lexicon_count_pipeline(tokenizer=prep.identity)
token_weights = dict(tfidfvect=feature_weights["word_tfidf"],
polpipe3=feature_weights["lexicon_count"])
token_transformers_dict = dict(
tfidfvect=
tfidfvect, # not to lose above integrity if we change variable names
polpipe3=polpipe3)
token_transformers = [(k, v) for k, v in token_transformers_dict.items()]
tokenbasedpipe = skpipeline.Pipeline([
('preprocessor', preprocessor),
# ('nadropper', tbt.DropNATransformer()),
('union1',
skpipeline.FeatureUnion(transformer_list=token_transformers,
transformer_weights=token_weights)),
])
charngramvect = TfidfVectorizer(analyzer='char_wb',
ngram_range=char_ngramrange,
lowercase=False)
polpipe1 = tbt.get_polylglot_polarity_count_pipe(lang)
polpipe2 = tbt.get_polylglot_polarity_value_pipe(lang)
text_weights = dict(charngramvect=feature_weights["char_tfidf"],
polpipe1=feature_weights["polyglot_count"],
polpipe2=feature_weights["polyglot_value"])
text_transformers_dict = dict(charngramvect=charngramvect,
polpipe1=polpipe1,
polpipe2=polpipe2)
text_transformers = [(k, v) for k, v in text_transformers_dict.items()]
'''
textpipes = [('charngramvect', charngramvect),]
textpweights = {'charngramvect' : 1.5}
textpweights = dict(charngramvect = 1 if charngramvect else 0)
'''
textbasedpipe = skpipeline.Pipeline([(
'union2',
skpipeline.FeatureUnion(transformer_list=text_transformers,
transformer_weights=text_weights),
)])
final_transformers_dict = dict(tokenbasedpipe=tokenbasedpipe,
textbasedpipe=textbasedpipe)
final_transformers = [(k, v) for k, v in final_transformers_dict.items()]
'''
#tweights = {k : 1 if v else 0 for k,v in final_transformers.items()}
check_zero = lambda x : 1 if sum(x) > 0 else 0
x = list(tokenbasedpipe.get_params(False).values())
print(len(x), x[0])
print(x[0][1]) # convert x[0] tuple to dict, then get transformer weights
print("**")
print(x,"\n--")
print(list(textbasedpipe.get_params(False).values()))
tweights = {k : check_zero(list(k.get_params(False).values())[0][0][1].get_params(False)["transformer_weights"].values())
for _, k in final_transformers_dict.items()}
'''
features = skpipeline.FeatureUnion(
transformer_list=final_transformers,
# transformer_weights=tweights # weight assignment is not necessary as the number of features is small
)
'''
tokenbasedpipe = skpipeline.Pipeline([('preprocessor', preprocessor),
#('nadropper', tbt.DropNATransformer()),
('union1', skpipeline.FeatureUnion(
transformer_list=[
('tfidfvect', tfidfvect),
#('polarity3', polpipe3),
])),]
)
textbasedpipe = skpipeline.Pipeline([('union2', skpipeline.FeatureUnion([
#('polarity1', polpipe1),
#('polarity2', polpipe2),
('charngramvect', charngramvect),
]),)])
features = skpipeline.FeatureUnion(transformer_list=[
('tokenbasedfeatures', tokenbasedpipe),
('textbasedfeatures', textbasedpipe),
])
'''
return features
def _ar_txt_clf_features_pipeline2(
feature_params_config_dict # {feature_params: {lang: .., weights : .., prep : {}, keywords : []}} see EMAIL_CONF for an example.
):
lang = feature_params_config_dict[conf.lang_key]
feature_weights = feature_params_config_dict[conf.weights_key]
prep_params = feature_params_config_dict[conf.prep_key]
#print(feature_weights)
# features found in the processed tokens
preprocessor = prep.Preprocessor(
lang=lang,
stopword=prep_params[conf.stopword_key],
more_stopwords=prep_params[conf.more_stopwords_key],
spellcheck=prep_params[conf.spellcheck_key],
stemming=prep_params[conf.stemming_key],
remove_numbers=prep_params[conf.remove_numbers_key],
deasciify=prep_params[conf.deasciify_key],
remove_punkt=prep_params[conf.remove_punkt_key],
lowercase=prep_params[conf.lowercase_key])
tfidfvect = TfidfVectorizer(
tokenizer=prep.identity,
preprocessor=None,
lowercase=False,
use_idf=prep_params[conf.use_idf_key],
ngram_range=prep_params[conf.wordngramrange_key],
max_features=prep_params[conf.nmaxfeature_key])
token_weights = dict(tfidfvect=feature_weights["word_tfidf"], )
token_transformers_dict = dict(
tfidfvect=
tfidfvect, # not to lose above integrity if we change variable names
)
token_transformers = [(k, v) for k, v in token_transformers_dict.items()]
tokenbasedpipe = skpipeline.Pipeline([
('preprocessor', preprocessor),
# ('nadropper', tbt.DropNATransformer()),
('union1',
skpipeline.FeatureUnion(transformer_list=token_transformers,
transformer_weights=token_weights)),
])
charngramvect = TfidfVectorizer(
analyzer='char_wb',
ngram_range=prep_params[conf.charngramrange_key],
lowercase=False)
# stylistic
'''
# BUG
named_entity_pipe = tbt.get_named_entity_weight_pipeline(lang)
text_weights = dict(charngramvect=feature_weights["char_tfidf"], # @TODO hardcoded
polpipe1=feature_weights["polyglot_count"],
polpipe2=feature_weights["polyglot_value"],
named_entity_pipe=feature_weights["named_entity_rate"])
text_transformers_dict = dict(charngramvect=charngramvect,
polpipe1=polpipe1,
polpipe2=polpipe2,
named_entity_pipe=named_entity_pipe)
'''
text_weights = dict(
charngramvect=feature_weights["char_tfidf"], # @TODO hardcoded
)
text_transformers_dict = dict(charngramvect=charngramvect, )
text_transformers = [(k, v) for k, v in text_transformers_dict.items()]
'''
textpipes = [('charngramvect', charngramvect),]
textpweights = {'charngramvect' : 1.5}
textpweights = dict(charngramvect = 1 if charngramvect else 0)
'''
textbasedpipe = skpipeline.Pipeline([(
'union2',
skpipeline.FeatureUnion(transformer_list=text_transformers,
transformer_weights=text_weights),
)])
final_transformers_dict = dict(tokenbasedpipe=tokenbasedpipe,
textbasedpipe=textbasedpipe)
final_transformers = [(k, v) for k, v in final_transformers_dict.items()]
#print(textbasedpipe.named_steps)
'''
#tweights = {k : 1 if v else 0 for k,v in final_transformers.items()}
check_zero = lambda x : 1 if sum(x) > 0 else 0
x = list(tokenbasedpipe.get_params(False).values())
print(len(x), x[0])
print(x[0][1]) # convert x[0] tuple to dict, then get transformer weights
print("**")
print(x,"\n--")
print(list(textbasedpipe.get_params(False).values()))
tweights = {k : check_zero(list(k.get_params(False).values())[0][0][1].get_params(False)["transformer_weights"].values())
for _, k in final_transformers_dict.items()}
'''
features = skpipeline.FeatureUnion(
transformer_list=final_transformers,
# transformer_weights=tweights # weight assignment is not necessary as the number of features is small
)
#print("0000000000", feature_params_config_dict)
return features
return self
def transform(self, X, y=None):
return X.fillna(self.most_frequent_)
cat_pipeline = pipeline.Pipeline([
("select_cat", DataFrameSelector(["Pclass", "Sex", "Embarked"])),
("imputer", MostFrequentImputer()),
("cat_encoder", OneHotEncoder(sparse=False)),
])
cat_pipeline.fit_transform(train_data)
preprocess_pipeline = pipeline.FeatureUnion(transformer_list=[
("num_pipeline", num_pipeline),
("cat_pipeline", cat_pipeline),
])
X_train = preprocess_pipeline.fit_transform(train_data)
print('X_train')
print(X_train[:5])
y_train = train_data["Survived"]
# SVC
svm_clf = svm.SVC()
svm_clf.fit(X_train, y_train)
print('svm_clf')
print(svm_clf)
# CHECK PREDICION
INPUT: Dataframe with features (X), target variable dataframe (y), polynomial degree (parameter)
OUTPUT: Score of XGB Regressor
'''
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
model_xgb = xgb.XGBRegressor(n_estimators=n_est,learning_rate=0.15,max_depth=2)
model_xgb.fit(X_train, y_train, eval_metric='rmse')
return model_xgb.score(X_test,y_test)
def regression_pipeline(x,y,degree,pca_comps):
'''
INPUT: Dataframe with features (X), target variable dataframe (y), polynomial degree, PCA components
OUTPUT: Score of regression pipeline using polynomial features and PCA
'''
# lets try using some feature tranforms, to our original X
combined_features = pipeline.FeatureUnion([('poly',PolynomialFeatures(degree=degree)),
('pca', decomposition.PCA(n_components=pca_comps))])
X_train,X_test,y_train,y_test = train_test_split(x,y,test_size = .3)
# now lets lay out the steps of our model
# First we do the feature transforms with our feature union defined above
# Second we do feature selection with the built-in SelectFromModel
# Third we train the actual model
steps = [
('features', combined_features),
('feature_selection', feature_selection.SelectFromModel(Lasso(alpha=.5))),
('model', LinearRegression())]
#but at this point we have only defined what reg, no training has happened yet
regression_pipeline = pipeline.Pipeline(steps)
# calling fit here calls fit on the entire pipeline which in turn executes all its members
regression_pipeline.fit(X_train,y_train)
clf = XGBClassifier(max_depth=5, n_estimators=100)
estimator = pipeline.Pipeline(steps=[
(
'feature_preprocessing',
pipeline.FeatureUnion(transformer_list=[
#real
('numeric_variables_processing',
pipeline.Pipeline(
steps=[('selecting',
preprocessing.FunctionTransformer(
lambda data: data[:, real_data_indices])),
('scaling',
preprocessing.StandardScaler(with_mean=0))])),
#categorical
('categorical_variables_processing',
pipeline.Pipeline(
steps=[('selecting',
preprocessing.FunctionTransformer(
lambda data: data[:, cat_data_indices])),
('hot_encoding',
preprocessing.OneHotEncoder(
handle_unknown='ignore'))])),
])),
('model_fitting', clf)
])
estimator.fit(train_data, y)
pred = estimator.predict(test_data)
write_to_csv('check.csv', ['PassengerId', 'Survived'], pred)
counters_pipe = pipeline.FeatureUnion(
n_jobs=-1,
transformer_list=[
('chars_features',
pipeline.Pipeline([('chars_counter',
TfidfVectorizer(analyzer=u'char',
ngram_range=(2, 5),
tokenizer=None,
max_features=config.max_features,
strip_accents=None,
max_df=0.9,
min_df=2,
lowercase=False)),
('chars_tsvd',
TruncatedSVD(n_components=config.svd_n_components,
n_iter=25,
random_state=42))])),
('words_features',
pipeline.Pipeline([('words_counter',
TfidfVectorizer(analyzer=u'word',
ngram_range=(1, 3),
tokenizer=None,
use_idf=True,
max_features=config.max_features,
strip_accents=None,
max_df=0.9,
min_df=2,
lowercase=False)),
('words_tsvd',
TruncatedSVD(n_components=config.svd_n_components,
n_iter=25,
random_state=42))])),
])
def _generate_feature_extraction_pipeline(self):
lang = self.feature_config.lang
final_weights = self.feature_config.weights
prep_params = self.feature_config.prepchoice
# features found in the processed tokens
token_features = []
token_weights = {}
preprocessor = prep.Preprocessor(
lang=lang,
stopword=prep_params.stopword,
more_stopwords=prep_params.more_stopwords,
spellcheck=prep_params.spellcheck,
stemming=prep_params.stemming,
remove_numbers=prep_params.remove_numbers,
deasciify=prep_params.deasciify,
remove_punkt=prep_params.remove_punkt,
lowercase=prep_params.lowercase)
tfidfvect = TfidfVectorizer(tokenizer=prep.identity,
preprocessor=None,
lowercase=False,
use_idf=prep_params.use_idf,
ngram_range=prep_params.wordngramrange,
max_features=prep_params.nmaxfeature)
tfidfvect_name = 'word_tfidfvect'
token_features.append((tfidfvect_name, tfidfvect))
token_weights[tfidfvect_name] = 1
# features found in the whole raw text
text_features = []
text_weights = {}
# charngramvect = TfidfVectorizer(analyzer='char_wb', ngram_range=(2, 2), lowercase=False)
tokenbasedpipe = skpipeline.Pipeline([
('preprocessor', preprocessor),
# ('nadropper', tbt.DropNATransformer()),
('union1',
skpipeline.FeatureUnion(transformer_list=token_features,
transformer_weights=token_weights)),
])
if text_weights:
textbasedpipe = skpipeline.Pipeline([(
'union2',
skpipeline.FeatureUnion(transformer_list=text_features,
transformer_weights=text_weights),
)])
'''
features = skpipeline.FeatureUnion(transformer_list=[
('tokenbasedfeatures', tokenbasedpipe),
('textbasedfeatures', textbasedpipe),
],
transformer_weights=final_weights)
'''
#######
# add the feature pipes to final_features if all the component weights are non-zero.
########
check_zero_list = lambda x: 1 if sum(x) > 0 else 0
# l = [0,0,0] => check_zero(l) gives 0 and l=[0,0,1] => check_zero(l) gives 1.
final_features_dict = {}
tkweights = list(token_weights.values())
if (check_zero_list(tkweights) != 0):
final_features_dict["token_based"] = tokenbasedpipe
else:
final_weights["token_based"] = 0
txweights = list(text_weights.values())
if (check_zero_list(txweights) != 0):
final_features_dict["text_based"] = textbasedpipe
else:
final_weights["text_based"] = 0
final_features = list(final_features_dict.items())
fweights = list(final_weights.values())
#print(final_weights)
if ((check_zero_list(fweights) == 0) or (len(final_features) == 0)):
return None
features = skpipeline.FeatureUnion(transformer_list=final_features,
transformer_weights=final_weights)
return features
def _build_feature_pipeline(self,
sample_mode='rollouts',
num_components=50,
gammas=None,
num_obs=10000,
use_standard_scaler=True,
featurizer_max_env_steps=10000):
"""Build the feature pipeline.
Args:
sample_mode: A string rerpresenting how to collect data from the
environment to build features. Must be {'rollouts', 'reset', 'random'}.
- `rollouts` will collect observations by executing a random policy in
the env.
- `reset` will collect rollouts by repeatedly resetting the env.
- `random` will just sample the env observation space randomly.
num_components: The number of components in each RBF.
gammas: A list containing the frequency of each RBF. If None will default
to `[0.5, 1.0, 2.5, 5.0]`.
num_obs: The integer number of observations to use to fit the Kernels.
use_standard_scaler: Boolean indicating if the observations should be
normalized.
featurizer_max_env_steps: Maximum number of steps to be taken in each
rollout to estimate the kernels in the featurizer.
Raises:
ValueError: If the `sample_mode` is unknown.
"""
env = self._env._envs[0] # pylint: disable=protected-access
if gammas is None:
gammas = [0.5, 1.0, 2.5, 5.0]
features = []
for i, gamma in enumerate(gammas):
features.append(
('rbf{}'.format(i),
kernel_approximation.RBFSampler(gamma=gamma,
n_components=num_components)))
self.featurizer = pipeline.FeatureUnion(features)
if use_standard_scaler:
self.scaler = skl_preprocessing.StandardScaler()
if sample_mode == 'random':
# Randomly sample from the observation space to fit the featurizers.
observation_examples = np.array([env.observation_space.sample() for _ in range(num_obs)]) # pylint: disable=line-too-long
elif sample_mode == 'reset':
# Just reset the environment to obtain the observations.
observation_examples = np.array(
[env.reset() for _ in range(num_obs)])
elif sample_mode == 'rollouts':
# Rollout mode.
observations = []
while True:
observations.append(env.reset())
done = False
t = 0
while not done and t < featurizer_max_env_steps:
action = env.action_space.sample()
obs, _, done, _ = env.step(action)
observations.append(obs)
if len(observations) > num_obs:
break # Collected enough observations.
observation_examples = np.array(observations)
else:
raise ValueError('Unknown `sample_mode`!')
if use_standard_scaler: self.scaler.fit(observation_examples)
if use_standard_scaler: self.scaler.transform(observation_examples)
self.featurizer.fit(observation_examples)
self.use_standard_scaler = use_standard_scaler
def main():
'''
The main function reads in data, sets up the union features
'''
train_file = 'train_data.csv'
blind_file = 'test_features_2013-03-07.csv'
#read csv files into a dataframe
train_df = pd.read_csv(train_file)
blind_df = pd.read_csv(blind_file)
#do print-out checks of the file
print(train_df.shape)
print(blind_df.shape)
print(train_df.head(5))
#fix column names in the blind_df column to be all lowercase to match with train_df
for col in blind_df.columns:
blind_df.rename(columns={col: col.lower()}, inplace=True)
print(blind_df.head(5))
#take out the zero-target values from the train data
X_total = train_df[train_df['target'] > 0]
#set the predicted y values to be the 'target' column from the train dataframe
y_total = X_total['target']
print(X_total.shape)
print(y_total.shape)
#divide train data into 'train' set and validation test set, where validation set is 20% of data from the training dataframe
X_train, X_test, y_train, y_test = cross_validation.train_test_split(X_total, y_total, test_size=0.2)
#make a dictionary of weights for each step in the model pipeline
transform_dict = {'job type':1, 'years exp': 1, 'location': 1,
'degree': 1, 'major':1, 'industry':1}
#combine features into one large set
all_features = pipeline.FeatureUnion([
('job type', FullModelTransformer(CategTransformer('jobtype'))),
('years exp', FullModelTransformer(LinFitTransformer('yearsexperience'))),
('location', FullModelTransformer(LinFitTransformer('milesfrommetropolis'))),
('degree', FullModelTransformer(CategTransformer('degree'))),
('major', FullModelTransformer(CategTransformer('major'))),
('industry', FullModelTransformer(CategTransformer('industry')))
],
transformer_weights=transform_dict)
#make a pipeline which performs fitting to all features and then fits those predictions to a an overall model (n nearest neighbors)
k_union = pipeline.Pipeline([
("features", all_features),
('modelfit', KNeighborsRegressor(n_neighbors=3))
#("linreg", LinearRegression(fit_intercept=True))
])
#fit the train data
k_union.fit(X_train, y_train.values.reshape(-1,1))
#print the R^2 score of the fit
print k_union.score(X_train, y_train.values.reshape(-1,1))
#fit the validation test data
k_union.fit(X_test, y_test.values.reshape(-1,1))
#print the R^2 score of the fit
print k_union.score(X_test, y_test.values.reshape(-1,1))
#predict on the blind data
result = k_union.predict(blind_df)
#add the prediction result as a column in the blind dataframe
blind_df['target'] = result
#write out resulting dataframe to a new csv file
header = ["jobid", "target"]
#blind_df.to_csv('test_target.csv', columns = header, index=False)
#predict on the entire input dataset
result = k_union.predict(X_total)
X_total['target_pred'] = result
#write resuling dataframe to new csv file
header = ["jobid", "target_pred"]
X_total.to_csv('train_target_pred.csv', columns = header, index=False)
#send results to plot
make_plot(X_total, blind_df)
return
# specify cross-validation
k = 10 # number of folds
cvsplitter = ms.KFold(n_splits=k, shuffle=True,
random_state=0) # cross-validation splitter
score = ms.cross_val_score(model, X, y, cv=cvsplitter)
print('Standardized linear discriminant analysis mean accuracy: {0:.4f}'.
format(score.mean()))
# define steps in a feature selection pipeline
features = list()
features.append(('pca', decomp.PCA(
n_components=3))) # use PCA to select 3 of the best features
features.append(('select_best', fs.SelectKBest(
k=6))) # use chi-squared test to select 6 of the best features
feature_union = pipeline.FeatureUnion(
features) # create the feature selection pipeline
estimators = list()
estimators.append((
'feature_union',
feature_union)) # add the feature selection pipleine to a new pipeline
estimators.append(('logistic', sl.LogisticRegression(
max_iter=1000))) # use logistic regression as the model
model = pipeline.Pipeline(
estimators
) # logistic regression with automatic feature selection by pca and chi-squared test
# specify cross-validation
score = ms.cross_val_score(model, X, y, cv=cvsplitter)
print(
'Logistic regression with automatic feature selection by PCA and chi2 test mean accuracy: {0:.4f}'
.format(score.mean()))
pipeline.FeatureUnion(
n_jobs=-1,
transformer_list=[
('standard', cust_regression_vals()),
('pi1',
pipeline.Pipeline([
('Gene', cust_txt_col('Gene')),
('count_Gene',
feature_extraction.text.CountVectorizer(analyzer=u'char',
ngram_range=(1,
8))),
('tsvd1',
decomposition.TruncatedSVD(n_components=20,
n_iter=25,
random_state=12))
])),
('pi2',
pipeline.Pipeline([
('Variation', cust_txt_col('Variation')),
('count_Variation',
feature_extraction.text.CountVectorizer(analyzer=u'char',
ngram_range=(1,
8))),
('tsvd2',
decomposition.TruncatedSVD(n_components=20,
n_iter=25,
random_state=12))
])),
('pi3',
pipeline.Pipeline([
('Text', cust_txt_col('Text')),
('hv',
feature_extraction.text.HashingVectorizer(
decode_error='ignore',
n_features=2**16,
non_negative=True,
ngram_range=(1, 3))),
('tfidf_Text', feature_extraction.text.TfidfTransformer()),
('tsvd3',
decomposition.TruncatedSVD(n_components=300,
n_iter=25,
random_state=12))
]))
]))
'NumUniqueBadges', 'NumPosts', 'MeanPostScore', 'MeanPostViews',
'MeanPostFavorites', 'MeanPostComments', 'NumComments', 'MeanCommentScore'
]
feature_cols = numeric_cols + ['AboutMe']
model = skl_pipeline.Pipeline([
('feat',
skl_pipeline.FeatureUnion([
('num',
skl_pipeline.Pipeline([
('get',
skl_preproc.FunctionTransformer(itemgetter(numeric_cols),
validate=False)),
('poly', skl_preproc.PolynomialFeatures()),
('std', skl_preproc.StandardScaler()),
])),
('text',
skl_pipeline.Pipeline([
('get',
skl_preproc.FunctionTransformer(itemgetter('AboutMe'),
validate=False)),
('tfidf', skl_featext.text.TfidfVectorizer()),
]))
])), ('reg', skl_linear.ElasticNet(alpha=1.0, l1_ratio=1.0))
])
param_grid = [
dict(
feat__num__poly__degree=[1, 2, 3],
feat__num__poly__interaction_only=[True, False],
feat__text__tfidf__max_df=[0.25, 0.5, 1.0],
_estimators.append(
(cluster.KMeans(random_state=42),
pickle.loads(pickle.dumps(pd_cluster.KMeans(random_state=42))), True))
_estimators.append((neighbors.KNeighborsClassifier(),
pd_neighbors.KNeighborsClassifier(), True))
_estimators.append(
(ensemble.GradientBoostingClassifier(random_state=42),
pd_ensemble.GradientBoostingClassifier(random_state=42), True))
_estimators.append((pipeline.make_union(decomposition.PCA(n_components=2),
feature_selection.SelectKBest(k=1)),
pd_decomposition.PCA(n_components=2) +
pd_feature_selection.SelectKBest(k=1), True))
_estimators.append(
(pipeline.FeatureUnion([('pca', decomposition.PCA(n_components=2)),
('kbest', feature_selection.SelectKBest(k=1))],
transformer_weights={
'pca': 3,
'kbest': 4
}),
pd_pipeline.FeatureUnion(
[('pca', pd_decomposition.PCA(n_components=2)),
('kbest', pd_feature_selection.SelectKBest(k=1))],
transformer_weights={
'pca': 3,
'kbest': 4
}), True))
_estimators.append((pipeline.make_union(decomposition.PCA(n_components=1),
decomposition.PCA(n_components=2),
feature_selection.SelectKBest(k=1)),
pd_decomposition.PCA(n_components=1) +
pd_decomposition.PCA(n_components=2) +
pd_feature_selection.SelectKBest(k=1), True))
def _email_features_pipeline2(feature_params_config_dict # {feature_params: {lang: .., weights : .., prep : {}, keywords : []}} see EMAIL_CONF for an example.
):
lang = feature_params_config_dict[conf.lang_key]
final_weights = feature_params_config_dict[conf.weights_key]
prep_params = feature_params_config_dict[conf.prep_key]
keywords = feature_params_config_dict[conf.keyword_key]
# features found in the processed tokens
token_features = []
token_weights = {}
preprocessor = prep.Preprocessor(lang=lang,
stopword=prep_params[conf.stopword_key], more_stopwords=prep_params[conf.more_stopwords_key],
spellcheck=prep_params[conf.spellcheck_key],
stemming=prep_params[conf.stemming_key],
remove_numbers=prep_params[conf.remove_numbers_key],
deasciify=prep_params[conf.deasciify_key],
remove_punkt=prep_params[conf.remove_punkt_key],
lowercase=prep_params[conf.lowercase_key]
)
tfidfvect = TfidfVectorizer(tokenizer=prep.identity, preprocessor=None, lowercase=False,
use_idf=prep_params[conf.use_idf_key],
ngram_range=prep_params[conf.ngramrange_key],
max_features=prep_params[conf.nmaxfeature_key])
tfidfvect_name = 'word_tfidfvect'
token_features.append((tfidfvect_name, tfidfvect))
token_weights[tfidfvect_name] = 1
# features found in the whole raw text
text_features = []
text_weights = {}
# charngramvect = TfidfVectorizer(analyzer='char_wb', ngram_range=(2, 2), lowercase=False)
# keyword presence features
if keywords:
for keyword in keywords:
keywordpipe = txbt.get_keyword_pipeline(keyword)
feature_name = "has_" + keyword
text_features.append((feature_name, keywordpipe))
text_weights[feature_name] = 1
tokenbasedpipe = skpipeline.Pipeline([('preprocessor', preprocessor),
# ('nadropper', tbt.DropNATransformer()),
('union1', skpipeline.FeatureUnion(
transformer_list=token_features ,
transformer_weights=token_weights
)),
])
textbasedpipe = skpipeline.Pipeline([('union2', skpipeline.FeatureUnion(
transformer_list=text_features,
transformer_weights=text_weights
),
)
])
'''
features = skpipeline.FeatureUnion(transformer_list=[
('tokenbasedfeatures', tokenbasedpipe),
('textbasedfeatures', textbasedpipe),
],
transformer_weights=final_weights)
'''
#######
# add the feature pipes to final_features if all the component weights are non-zero.
########
check_zero_list = lambda x : 1 if sum(x) > 0 else 0
# l = [0,0,0] => check_zero(l) gives 0 and l=[0,0,1] => check_zero(l) gives 1.
final_features_dict = {}
tkweights = list(token_weights.values())
if(check_zero_list(tkweights) != 0):
final_features_dict["token_based"] = tokenbasedpipe
else:
final_weights["token_based"] = 0
txweights = list(text_weights.values())
if(check_zero_list(txweights) != 0):
final_features_dict["text_based"] = textbasedpipe
else:
final_weights["text_based"] = 0
final_features = list(final_features_dict.items())
fweights = list(final_weights.values())
#print(final_weights)
if((check_zero_list(fweights) == 0) or (len(final_features) == 0)):
return None
features = skpipeline.FeatureUnion(transformer_list=final_features,
transformer_weights=final_weights)
return features
def __init__(self, env, use_kernel=False, **agent_params):
self.env = env
self.use_kernel = use_kernel
self.agent_params = {
"epsilon_min": 0.01,
"decay_rate": 0.01,
"discount": 0.99,
"iter": 200,
}
self.agent_params.update(agent_params)
# Generating feature space of RBF kernels
if self.use_kernel:
observation_samples = np.array(
[env.observation_space.sample() for x in range(10000)])
self.detrend = preprocessing.StandardScaler()
self.detrend.fit(observation_samples)
self.featurizer = pipeline.FeatureUnion([
("rbf1", RBFSampler(gamma=5.0, n_components=100)),
("rbf2", RBFSampler(gamma=2.0, n_components=100)),
("rbf3", RBFSampler(gamma=1.0, n_components=100)),
("rbf4", RBFSampler(gamma=0.5, n_components=100))
])
self.featurizer.fit(self.detrend.transform(observation_samples))
self.n_features = len(
self.featurizer.transform(env.observation_space.sample())[0])
else:
self.n_features = len(env.observation_space.sample())
print(self.n_features)
# Generating linear model approximation for value function
with tf.variable_scope("value_function"):
self.value_features = tf.placeholder(tf.float32, [self.n_features],
name="value_features")
self.value_reward_target = tf.placeholder(
tf.float32, name="value_reward_target")
value_output_layer = tf.contrib.layers.fully_connected(
inputs=tf.expand_dims(self.value_features, 0),
num_outputs=1,
activation_fn=None,
weights_initializer=tf.zeros_initializer)
self.value_estimate = tf.squeeze(value_output_layer)
self.value_loss = tf.squared_difference(self.value_estimate,
self.value_reward_target)
self.value_optimizer = tf.train.AdamOptimizer()
self.value_train_op = self.value_optimizer.minimize(
self.value_loss)
# Generating linear model approximation for policy function
with tf.variable_scope("policy_function"):
self.action = tf.placeholder(tf.int32, name="action")
self.policy_features = tf.placeholder(tf.float32,
[self.n_features],
name="policy_features")
self.policy_reward_target = tf.placeholder(
tf.float32, name="policy_reward_target")
policy_output_layer = tf.contrib.layers.fully_connected(
inputs=tf.expand_dims(self.policy_features, 0),
num_outputs=env.action_space.n,
activation_fn=None,
weights_initializer=tf.zeros_initializer)
self.action_probabilities = tf.squeeze(
tf.nn.softmax(policy_output_layer))
self.max_action_probability = tf.gather(self.action_probabilities,
self.action)
self.policy_loss = -tf.log(
self.max_action_probability) * self.policy_reward_target
self.policy_optimizer = tf.train.AdamOptimizer()
self.policy_train_op = self.policy_optimizer.minimize(
self.policy_loss)
train.fillna('', inplace=True)
test.fillna('', inplace=True)
counters_pipe = pipeline.FeatureUnion(
n_jobs=-1,
transformer_list=[
('chars_features',
TfidfVectorizer(analyzer=u'char',
ngram_range=(2, 5),
tokenizer=None,
max_features=config.max_features,
strip_accents=None,
max_df=0.9,
min_df=2,
lowercase=False)),
('words_features',
TfidfVectorizer(analyzer=u'word',
ngram_range=(1, 3),
tokenizer=None,
use_idf=True,
max_features=config.max_features,
strip_accents=None,
max_df=0.9,
min_df=2,
lowercase=False)),
])
cv = KFold(n_splits=config.nfolds, shuffle=True, random_state=42)
splits = list(cv.split(train))