def main_eval():
print("load specified model")
model = load_model(args.model, custom_objects=evaluation.get_metrics())
print("load evaluation image")
img = dataset.load_image_eval(args.data)
print("run evaluation on final year")
y_pred = evaluation.predict_image(model, img, args.area_size)
visualize.save_image_as(y_pred, "res/out.png")
def main_train_h5():
print("check for data.h5")
try:
open(args.h5data, "r")
except FileNotFoundError:
h5dataset.make_dataset(args.h5data)
print("load remaining data")
sat_images = dataset.load_sat_images(args.data)
alt, slp = dataset.load_static_data(args.data)
print("initialize training generator")
train_gen = h5dataset.patch_generator_from_h5(args.h5data,
sat_images,
alt,
slp,
size=args.area_size,
batch_size=args.batch_size,
p=args.p_train)
print("initialize validation generator")
val_gen = h5dataset.patch_generator_from_h5(args.h5data,
sat_images,
alt,
slp,
size=args.area_size,
batch_size=args.batch_size,
p=args.p_val)
print("get network")
model = networks.get_model_by_name(args.model_type)(args)
print("compile")
custom_metrics = list(evaluation.get_metrics().values())
model.compile(optimizer="adam",
loss="binary_crossentropy",
metrics=["accuracy"] + custom_metrics)
print(model.summary())
print("start training")
model.fit_generator(train_gen,
steps_per_epoch=args.steps_per_epoch,
epochs=args.epochs,
validation_data=val_gen,
validation_steps=args.steps_per_val,
verbose=True,
max_q_size=args.queue_size,
workers=1)
print("store model")
model.save(args.model)
def new_model(data1, data2):
"""
Training and testing a tree on the noisy dataset without those observations above
Results in a 97% accuracy
"""
diff_obs = check_signals_only(data1, data2)
df2 = pd.DataFrame(data2)
clean_removed = pd.concat([df2, diff_obs,
diff_obs]).drop_duplicates(keep=False)
clean_removed_dataset = clean_removed.to_numpy()
np.random.shuffle(clean_removed_dataset)
split = 0.7
train = clean_removed_dataset[:int(len(clean_removed_dataset) * split)]
test = clean_removed_dataset[int(len(clean_removed_dataset) * split):]
model = trees.binarySearchTree(train)
print('Max depth is', model.get_max_depth())
y_pred = model.predict(test[:, :-1])
cm = ev.confusion_matrix(test[:, -1], y_pred)
i = ev.get_metrics(cm, printout=True)
ev.plot_conf_matrix(cm)
def train_eval_all_folds(self, x_val, y_val):
"""Trains and evaluates models over all folds.
Args:
timestamp [string]: timestamp of the time the model is run,
used as model identifier
x_val [ndarray]: feature matrix
y_val [ndarray]: label vector
Returns:
auc [float]: area under the ROC curve
mean_fpr [list of floats]: false positive rate averaged over
all kfolds
mean_tpr [list of floats]: true positive rate averaged over
all kfolds
trained_model [sklearn object]: trained object to be pickled
so that it can be used for
scoring
"""
if self.world_type == "closed":
# Why we use stratified k-fold here:
# http://stats.stackexchange.com/questions/49540/understanding-stratified-cross-validation
cv = cross_validation.StratifiedKFold(y_val, n_folds=self.k,
shuffle=True)
elif self.world_type == "open":
pass # TODO
fpr_arr, tpr_arr, metrics_all_folds = [], [], []
for i, (train, test) in enumerate(cv):
fold_timestamp = datetime.datetime.now().isoformat()
y_train, y_test = y_val[train], y_val[test]
if self.feature_scaling:
scaler = preprocessing.StandardScaler().fit(x_val[train])
x_train = scaler.transform(x_val[train])
x_test = scaler.transform(x_val[test])
else:
x_train, x_test = x_val[train], x_val[test]
trained_model = self.train_single_fold(x_train, y_train)
pred_probs = self.score(x_test, trained_model)
filename_kfold = '{}_{}_undefended_frontpage_{}_model_{}_fold_{}_world.pkl'.format(
fold_timestamp, self.model_timestamp, self.model_type, i, self.world_type)
fold_to_save = {'trained_object': trained_model,
'y_true': y_test, 'y_predicted': pred_probs}
self.pickle_results(filename_kfold, fold_to_save)
# Metrics computation
# Compute ROC curve and area under the ROC curve
eval_metrics = evaluation.get_metrics(y_test, pred_probs)
metrics_all_folds.append(eval_metrics)
fpr_arr.append(eval_metrics['fpr'])
tpr_arr.append(eval_metrics['tpr'])
# Save results of metrics in database
self.db.save_fold_of_model(eval_metrics, self.model_timestamp, fold_timestamp)
auc = evaluation.plot_allkfolds_ROC(self.model_timestamp, cv,
fpr_arr, tpr_arr)
print("Classifier {} trained! AUC: {}".format(self.model_timestamp,
auc))
avg_metrics = evaluation.get_average_metrics(metrics_all_folds)
# Save results of experiment (model evaluation averaged over all
# folds) into the database
self.db.save_full_model(avg_metrics, self.model_timestamp, self.__dict__)
def train_eval_all_folds(self, x_val, y_val):
"""Trains and evaluates models over all folds.
Args:
timestamp [string]: timestamp of the time the model is run,
used as model identifier
x_val [ndarray]: feature matrix
y_val [ndarray]: label vector
Returns:
auc [float]: area under the ROC curve
mean_fpr [list of floats]: false positive rate averaged over
all kfolds
mean_tpr [list of floats]: true positive rate averaged over
all kfolds
trained_model [sklearn object]: trained object to be pickled
so that it can be used for
scoring
"""
if self.world_type == "closed":
# Why we use stratified k-fold here:
# http://stats.stackexchange.com/questions/49540/understanding-stratified-cross-validation
cv = cross_validation.StratifiedKFold(y_val,
n_folds=self.k,
shuffle=True)
elif self.world_type == "open":
pass # TODO
fpr_arr, tpr_arr, metrics_all_folds = [], [], []
for i, (train, test) in enumerate(cv):
fold_timestamp = datetime.datetime.now().isoformat()
y_train, y_test = y_val[train], y_val[test]
if self.feature_scaling:
scaler = preprocessing.StandardScaler().fit(x_val[train])
x_train = scaler.transform(x_val[train])
x_test = scaler.transform(x_val[test])
else:
x_train, x_test = x_val[train], x_val[test]
trained_model = self.train_single_fold(x_train, y_train)
pred_probs = self.score(x_test, trained_model)
filename_kfold = '{}_{}_undefended_frontpage_{}_model_{}_fold_{}_world.pkl'.format(
fold_timestamp, self.model_timestamp, self.model_type, i,
self.world_type)
fold_to_save = {
'trained_object': trained_model,
'y_true': y_test,
'y_predicted': pred_probs
}
self.pickle_results(filename_kfold, fold_to_save)
# Metrics computation
# Compute ROC curve and area under the ROC curve
eval_metrics = evaluation.get_metrics(y_test, pred_probs)
metrics_all_folds.append(eval_metrics)
fpr_arr.append(eval_metrics['fpr'])
tpr_arr.append(eval_metrics['tpr'])
# Save results of metrics in database
self.db.save_fold_of_model(eval_metrics, self.model_timestamp,
fold_timestamp)
auc = evaluation.plot_allkfolds_ROC(self.model_timestamp, cv, fpr_arr,
tpr_arr)
print("Classifier {} trained! AUC: {}".format(self.model_timestamp,
auc))
avg_metrics = evaluation.get_average_metrics(metrics_all_folds)
# Save results of experiment (model evaluation averaged over all
# folds) into the database
self.db.save_full_model(avg_metrics, self.model_timestamp,
self.__dict__)
print("Loading data...")
X, Y = load_dataset()
print("Training model")
t0 = time()
transformer = TfidfVectorizer(ngram_range=(1, 2), max_df=0.5)
X_train, X_test, y_train, y_test = train_test_split(X,
Y,
test_size=0.2,
random_state=42)
X_train = transformer.fit_transform(X_train)
X_test = transformer.transform(X_test)
model = MultinomialNB()
clf = model.fit(X_train, y_train)
train_time = time() - t0
print("Finished")
print("\t- train time: %0.3fs" % train_time)
t0 = time()
y_pred = clf.predict(X_test)
print(classification_report(y_test, y_pred))
test_time = time() - t0
print("\t- test time: %0.3fs" % test_time)
get_metrics(y_test, y_pred)
save_model("model/model.pkl", clf)
if limit == '':
print('No limit entered')
limit = None
else:
limit = int(limit)
np.random.shuffle(data)
train = data[:int(len(data) * split)]
test = data[int(len(data) * split):]
model = binarySearchTree(train, limit=limit)
print('Max depth of tree is', model.get_max_depth())
y_pred = model.predict(test[:, :-1])
cm = ev.confusion_matrix(test[:, -1], y_pred)
i = ev.get_metrics(cm, printout=True)
print('To continue, you may need to close the plot windows first')
ev.plot_conf_matrix(cm)
print('Visualising the pruned trees')
model.visualise_tree()
input('\nTo restart, hit enter\n')
if model == '2':
split = float(input('Enter training data split value, eg 0.7\n'))
while True:
if split < 0 or split > 1:
print('Invalid split entered')
else:
break
print('You have entered ' + str(split) + '\n')
def ensemble_learning(directory_name,
data,
X,
y,
baseline=-1,
model_num=None,
resample=0,
feature_set=None,
feature_importance=0,
average_method='macro',
path=None):
"""
Store the results calculated according to the arguments and store them in a file.
Arguments:
directory_name (str): the directory under which the files should be stored
data (dataframe): the whole dataset
X (dataframe): examples
y (dataframe): target/label
baseline (int): -1 for no baseline, 1 for all predictions as 1, 0 for all predictions as 0
model_num (int): classification model
1:
2:
3:
4:
5:
6:
resample (int): -1 for undersampling, 1 for oversampling and 0 for no resampling
feature_set (list): list of features to be considered
feature_importance (int): 0 for absent, 1 for present
average_method: macro by default
path: the path to the directory where the recordings should be stored
"""
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=42)
#prepare the dictionary to be written to the file
data_dict = dict()
metrics_dict = dict()
dir_name = path + directory_name + '/'
os.mkdir(dir_name)
#open the config file for writing
config_file = open(dir_name + 'config.json', 'w')
#open the metrics file for writing
metrics_file = open(dir_name + 'metrics.json', 'w')
data_dict = {'model_num': model_num}
data_dict = {'baseline': baseline}
data_dict.update({'resample': resample})
data_dict.update({'feature_set': feature_set})
data_dict.update({'n_features': n_features})
data_dict.update({'feature_importance': feature_importance})
'''
#create test set labels for the baseline if applicable
if baseline == 0:
y_test = y_test.replace(1,0)
elif baseline == 1:
y_test = y_test.replace(0,1)
'''
#resample the training set (if applicable)
if resample == -1:
#undersample
'''NearMiss 3 . NearMiss-3 is a 2-step algorithm: first, for each minority sample,
their :m nearest-neighbors will be kept; then, the majority samples selected are the
on for which the average distance to the k nearest neighbors is the largest.'''
nm = NearMiss(version=3)
print(sorted(Counter(y_train).items()))
X_resampled, y_resampled = nm.fit_resample(X_train, y_train)
X_train = X_resampled
y_train = y_resampled
print(str(sorted(Counter(y_train).items())))
elif resample == 1:
#oversample
X_resampled, y_resampled = SMOTE().fit_resample(X_train, y_train)
X_train = X_resampled
y_train = y_resampled
print(sorted(Counter(y_resampled).items()))
#write the training dataset class distribution to the file
file = open(dir_name + 'train_val_dist.csv', 'a')
file.write(str(sorted(Counter(y_train).items())))
file.write('\n')
file.close()
model = get_model(model_num)
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
if baseline == 0:
y_pred = y_pred.replace(1, 0)
elif baseline == 1:
y_pred = y_pred.replace(0, 1)
plot_lc(model=model,
cv=StratifiedKFold(n_splits=5, shuffle=True, random_state=777),
X=X,
y=y)
#evaluation
metrics = get_metrics(y_test, y_pred)
for key, value in metrics.items():
metrics_dict[key] = value
#correlation
correlation(data)
#linearity
test_for_linearity(X_train, y_train)
#homoscedasticity
test_for_homoscedasticity(X_train, y_train, X_test, y_test)
'''
#learning curve
#if model_num == 7:
cv = ShuffleSplit(n_splits=5, test_size=0.2, random_state=0)
#else:
#cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=777)
train_sizes, train_scores, test_scores = learning_curve(estimator = model, X = data[feature_set], y = data['label'], cv = cv, scoring = 'f1_macro', train_sizes=np.linspace(.1, 1.0, 10))
# Create means and standard deviations of training set scores
train_mean = np.mean(train_scores, axis=1)
train_std = np.std(train_scores, axis=1)
print('scores: ', train_scores, train_mean)
# Create means and standard deviations of test set scores
test_mean = np.mean(test_scores, axis=1)
test_std = np.std(test_scores, axis=1)
# Draw lines
print('Learning Curve')
plt.plot(train_sizes, train_mean, '--', color="#111111", label="Training score")
plt.plot(train_sizes, test_mean, color="#111111", label="Cross-validation score")
# Draw bands
plt.fill_between(train_sizes, train_mean - train_std, train_mean + train_std, color="#DDDDDD")
plt.fill_between(train_sizes, test_mean - test_std, test_mean + test_std, color="#DDDDDD")
# Create plot
plt.title("Learning Curve")
plt.xlabel("Training Set Size"), plt.ylabel("Macro-F1 Score"), plt.legend(loc="best")
plt.tight_layout()
plt.show()
'''
plot_learning_curves(X_train,
y_train,
X_test,
y_test,
model,
scoring='f1_macro')
plt.show()
if feature_importance == 1:
feat_importances = pd.Series(model.feature_importances_,
index=feature_set)
print(feature_set)
print('Feat: ', feat_importances)
feat_importances.nlargest(20).plot(kind='barh')
#plot_importance(model)
plt.show()
perm = PermutationImportance(model,
random_state=1).fit(X_train, y_train)
display(eli5.show_weights(perm,
feature_names=X_train.columns.tolist()))
#write the training dataset class distribution to the file
file = open(dir_name + 'feature_importances.csv', 'a')
for ind in range(0, len(feature_set)):
file.write(feature_set[ind] + ',' + str(feat_importances[ind]) +
'\n')
file.close()
#write the permutation feature importance decrease in error values to the file
file = open(dir_name + 'permutation_feature_importances.csv', 'a')
print(perm.feature_importances_)
for ind in range(0, len(feature_set)):
file.write(feature_set[ind] + ',' +
str(perm.feature_importances_[ind]) + '\n')
file.close()
#write the scores to the file
json.dump(metrics_dict, metrics_file)
metrics_file.close()
#write the configuration values to the file
json.dump(data_dict, config_file)
config_file.close()
"""
run eval on a pair of corpus files
"""
import sys
sys.path.append('../../src/style_transfer_baseline')
import evaluation
import models
src_path = sys.argv[1]
pred_path = sys.argv[2]
tgt_path = sys.argv[3]
classifier_path = "../../data/v2/eval_classifier"
eval_classifier = models.TextClassifier.from_pickle(
"../../data/v2/eval_classifier")
src = [x.strip().split() for x in open(src_path)]
pred = [x.strip().split() for x in open(pred_path)]
tgt = [x.strip().split() for x in open(tgt_path)]
print(evaluation.get_metrics(src, pred, tgt, classifier=eval_classifier))
with open('log.txt', 'w') as file:
file.write('k, knn_frac, min_overlap, map_k, cosine\n')
for i in range(len(k_vals)):
for j in range(len(knn_frac_vals)):
for k in range(len(min_overlap_vals)):
print(song_df.shape)
tuning_model = ALSpkNN(user_df,
song_df,
k_vals[i],
knn_frac_vals[j],
min_overlap_vals[k],
cf_weighting_alpha=1)
print("Fitting model...")
tuning_model.fit(train_plays)
metrics = get_metrics(
metrics=['MAP@K', 'mean_cosine_list_dissimilarity'],
N=20,
model=tuning_model,
train_user_items=train_plays.transpose(),
test_user_items=test_plays.transpose(),
song_df=song_df,
limit=10)
mapk = metrics['MAP@K']
cosdis = metrics['cosine_list_dissimilarity']
with open('log.txt', 'a') as file:
file.write(
f'{k_vals[i]},{knn_frac_vals[j]},{min_overlap_vals[k]},{mapk},{cosdis}\n'
)
def crossvalidate(directory_name,
splits,
data,
X,
y,
baseline=-1,
model_num=None,
resample=0,
feature_set=None,
feature_importance=0,
average_method='macro',
path=None):
"""
Store the results calculated according to the arguments and store them in a file.
Arguments:
directory_name (str): the directory under which the files should be stored
splits (int): number of folds
data (dataframe): the whole dataset
X (dataframe): examples
y (dataframe): target/label
baseline (int): -1 for no baseline, 1 for all predictions as 1, 0 for all predictions as 0
model_num (int): classification model
1:
2:
3:
4:
5:
6:
resample (int): -1 for undersampling, 1 for oversampling and 0 for no resampling
feature_set (list): list of features to be considered
feature_importance (int): 0 for absent, 1 for present
average_method: macro by default
path: the path to the directory where the recordings should be stored
"""
#prepare the dictionary to be written to the file
data_dict = dict()
metrics_dict = dict()
dir_name = path + directory_name + '/'
os.mkdir(dir_name)
#create a directory for each split
for fold in range(1, splits + 1):
os.mkdir(dir_name + str(fold))
print(dir_name + str(fold))
#open the config file for writing
config_file = open(dir_name + 'config.json', 'w')
#open the metrics file for writing
metrics_file = open(dir_name + 'metrics.json', 'w')
data_dict = {'model_num': model_num}
data_dict = {'baseline': baseline}
data_dict.update({'resample': resample})
data_dict.update({'feature_set': feature_set})
data_dict.update({'n_features': n_features})
data_dict.update({'feature_importance': feature_importance})
metrics_dict = dict()
metrics_dict['f1_macro'] = list()
metrics_dict['tpr'] = list()
metrics_dict['tnr'] = list()
metrics_dict['fpr'] = list()
metrics_dict['precision'] = list()
metrics_dict['recall'] = list()
metrics_dict['accuracy'] = list()
metrics_dict['f1'] = list()
model = get_model(model_num)
kfold = StratifiedKFold(n_splits=splits, shuffle=True, random_state=777)
#if model_num == 3:
#kfold = ShuffleSplit(n_splits=splits, test_size=0.2, random_state=0)
plot_lc(model=model, cv=kfold, X=X, y=y, resample=resample)
#linearity
test_for_linearity(X, y)
i = 0
for train_index, test_index in kfold.split(X, y):
#create train-test splits
X_train, y_train = X.iloc[train_index], y.iloc[train_index]
X_test, y_test = X.iloc[test_index], y.iloc[test_index]
'''
#create test set labels for the baseline if applicable
if baseline == 0:
y_test = y_test.replace(1,0)
elif baseline == 1:
y_test = y_test.replace(0,1)
'''
#resample the training set (if applicable)
if resample == -1:
#undersample
'''NearMiss 3 . NearMiss-3 is a 2-step algorithm: first, for each minority sample,
their :m nearest-neighbors will be kept; then, the majority samples selected are the
on for which the average distance to the k nearest neighbors is the largest.'''
nm = NearMiss(version=3)
print(str(sorted(Counter(y_train).items())))
X_resampled, y_resampled = nm.fit_resample(X_train, y_train)
X_train = X_resampled
y_train = y_resampled
print(sorted(Counter(y_train).items()))
elif resample == 1:
#oversample
X_resampled, y_resampled = SMOTE().fit_resample(X_train, y_train)
X_train = X_resampled
y_train = y_resampled
print(sorted(Counter(y_resampled).items()))
#write the training dataset class distribution to the file
file = open(dir_name + str(i + 1) + '/train_val_dist.csv', 'a')
file.write(str(sorted(Counter(y_train).items())))
file.write('\n')
file.close()
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
if baseline == 0:
y_pred = y_pred.replace(1, 0)
elif baseline == 1:
y_pred = y_pred.replace(0, 1)
metrics = get_metrics(y_test, y_pred)
for key, value in metrics.items():
metrics_dict[key].append(value)
#homoscedasticity
test_for_homoscedasticity(X_train, y_train, X_test, y_test)
#correlation
correlation(data)
if feature_importance == 1:
if model_num == 1:
feat_importances = pd.Series(model.feature_importances_,
index=X.columns)
elif model_num == 3:
feat_importances = pd.Series(abs(svm.coef_[0]),
index=X.columns)
if model_num != 2:
print('Feat. Imp.: ', feat_importances)
feat_importances.nlargest(20).plot(kind='barh')
#plot_importance(model)
plt.show()
#write the feature importance values to the file
file = open(dir_name + str(i + 1) + '/feature_importances.csv',
'a')
for ind in range(0, len(feature_set)):
file.write(feature_set[ind] + ',' +
str(feat_importances[ind]) + '\n')
file.close()
perm = PermutationImportance(model,
random_state=1).fit(X_train, y_train)
print('PERM: ', perm.feature_importances_)
display(
eli5.show_weights(perm,
feature_names=X_train.columns.tolist()))
#write the permutation feature importance decrease in error values to the file
file = open(
dir_name + str(i + 1) + '/permutation_feature_importances.csv',
'a')
for ind in range(0, len(feature_set)):
file.write(feature_set[ind] + ',' +
str(perm.feature_importances_[ind]) + '\n')
file.write('\n')
file.close()
i += 1
for key, values in metrics_dict.items():
metrics_dict[key] = sum(values) / len(values)
#write the scores to the file
json.dump(metrics_dict, metrics_file)
metrics_file.close()
#write the configuration values to the file
json.dump(data_dict, config_file)
config_file.close()