def evaluate_test(x_eval, y_eval, res_eval, model, config, test=False):
total_pred = np.array([], dtype=np.float64)
total_y_test = np.array([], dtype=np.int64)
if (config['old'] == False):
size = 2048
else:
size = 1024
transformed_target = np.zeros(shape=(1, size), dtype=np.float64)
test_batch_num = int(
math.ceil(x_eval.shape[0] / float(config['batch_size'])))
cum_acc = []
cum_f1 = []
with torch.no_grad():
for i in range(test_batch_num):
begin_index = i * config['batch_size']
end_index = min((i + 1) * config['batch_size'], x_eval.shape[0])
batch_test_x = x_eval[begin_index:end_index]
batch_test_y = y_eval[begin_index:end_index]
batch_test_res = res_eval[begin_index:end_index]
if (config['model_type'] == 'dan'
or config['model_type'] == 'dann'):
batch_test_x = batch_test_x.reshape(-1, 1, 1,
batch_test_x.shape[1])
elif (config['model_type'] == 'jdda'
or config['model_type'] == 'jdda_rnn'
or config['model_type'] == 'jdda_ff'):
batch_test_x = batch_test_x.reshape(-1, 1,
batch_test_x.shape[1])
batch_test_x = torch.from_numpy(batch_test_x).float().to(
config['device'])
batch_test_y = torch.from_numpy(batch_test_y).long().to(
config['device'])
batch_test_res = torch.from_numpy(batch_test_res).float()
output_test, hidden_test = model(batch_test_x)
_, predicted = torch.max(output_test.data, 1)
# loss = _loss(batch_test_res, batch_test_y, output_test)
total_pred = np.concatenate((total_pred, predicted.cpu()))
total_y_test = np.concatenate((total_y_test, batch_test_y.cpu()))
batch_acc = accuracy_score(batch_test_y.cpu().numpy(),
predicted.cpu().numpy())
_, _, batch_f1, _ = prfs(batch_test_y.cpu().numpy(),
predicted.cpu().numpy(),
average='weighted')
cum_acc.append(batch_acc)
cum_f1.append(batch_f1)
transformed_target = np.concatenate(
(transformed_target, hidden_test.cpu().numpy()))
if (test == True):
print(classification_report(total_y_test, total_pred, digits=4))
acc = accuracy_score(total_y_test, total_pred)
cum_acc = np.array(cum_acc)
cum_f1 = np.array(cum_f1)
print("Testing accuracy", acc)
print("Average acc and std ", np.mean(cum_acc), np.std(cum_acc))
print("Average F1 and std ", np.mean(cum_f1), np.std(cum_f1))
_, _, f1, _ = prfs(total_y_test, total_pred, average='weighted')
transformed_target = np.delete(transformed_target, (0), axis=0)
print("Target - Eval", transformed_target.shape, x_eval.shape)
return f1, transformed_target
def _compute_metrics(self,
gt_all,
pred_all,
types,
print_results: bool = False):
labels = [t.index for t in types]
per_type = prfs(gt_all,
pred_all,
labels=labels,
average=None,
zero_division=0)
micro = prfs(gt_all,
pred_all,
labels=labels,
average='micro',
zero_division=0)[:-1]
macro = prfs(gt_all,
pred_all,
labels=labels,
average='macro',
zero_division=0)[:-1]
total_support = sum(per_type[-1])
if print_results:
self._print_results(per_type,
list(micro) + [total_support],
list(macro) + [total_support], types)
return [m * 100 for m in micro + macro]
def classify_vuamc(all_features,
features_to_use,
classifier=xgb.XGBClassifier,
n_dev=0,
seed=0,
classifier_args=None):
if classifier_args is None:
classifier_args = {}
print("Extracting features {}".format(', '.join(features_to_use)))
F = preprocess(all_features, features_to_use, n_dev=n_dev, seed=seed)
print("{} total features".format(F['X_train'].shape[1]))
print("{} train, {} dev, {} test".format(F['X_train'].shape[0],
F['X_dev'].shape[0],
F['X_test'].shape[0]))
print("Training {}".format(classifier))
clf = classifier(scale_pos_weight=sum(1 - F['y_train']) /
sum(F['y_train']),
**classifier_args)
clf.fit(F['X_train'], F['y_train'])
print("Scoring")
y_pred_test = clf.predict(F['X_test'])
y_pred_dev = clf.predict(F['X_dev'])
p, r, f1, s = prfs(F['y_test'], y_pred_test, average='binary')
print("Precision:", p)
print("Recall:", r)
print("F1 Score:", f1)
acc = (F['y_test'] == y_pred_test).mean()
print("Accuracy:", acc)
n_features = F['X_train'].shape[1]
# By genre
genres = np.unique(F['test_genre'])
for genre in genres:
print("Genre: {}".format(genre))
genre_mask = F['test_genre'] == genre
y_pred_genre = clf.predict(F['X_test'][genre_mask])
pg, rg, fg, sg = prfs(F['y_test'][genre_mask],
y_pred_genre,
average='binary')
print("Precision:", pg)
print("Recall:", rg)
print("F1 Score:", fg)
accg = (F['y_test'][genre_mask] == y_pred_genre).mean()
print("Accuracy:", accg)
print()
stats = {
'precision': p,
'recall': r,
'f1': f1,
'accuracy': acc,
'n_features': n_features
}
F['y_pred_test'] = y_pred_test
F['y_pred_dev'] = y_pred_dev
return F, stats
def evalute_one(self, metadata, predictions, batches):
def log_loss(y, _p):
eps = 1e-3
p = np.clip(_p, eps, 1. - eps)
return np.mean(-(y * np.log(p) + (1 - y) * np.log(1 - p)))
outputs = ['y_onsets', 'y_frames', 'y_offsets']
result = dict()
for output in outputs:
loss = log_loss(batches[output], predictions[output])
y_true = (batches[output] > 0.5) * 1
y_pred = (predictions[output] > 0.5) * 1
p, r, f, _ = prfs(y_true, y_pred, average='micro')
result[output] = dict(
loss=loss,
p=p,
r=r,
f=f
)
result['adsr'] = self.evaluate_adsr(metadata, predictions)
return result
def on_epoch_end(self, epoch, logs={}):
predict = np.argmax(self.model.predict(
self.validation_data[:2], batch_size=self.params['batch_size']),
axis=1)
targ = np.argmax(self.validation_data[2], axis=1)
prec, rec, f1, _ = prfs(targ, predict, average='macro')
if self.verbose:
info("epoch: {} p: {:0.4f} r: {:0.4f} f:{:0.4f}".format(
epoch, prec, rec, f1))
if f1 > self.best:
self.best = f1
self.best_epoch = epoch
self.wait = 0
self.best_weights = self.model.get_weights()
else:
if self.wait >= self.patience:
self.model.stop_training = True
self.model.set_weights(self.best_weights)
info("Stopping at epoch {}, best: e {}, f: {}.".format(
epoch, self.best_epoch, self.best))
else:
self.wait += 1
if epoch == (self.params['epochs'] - 1):
info("Stopping at final epoch {}, best: e {}, f: {}.".format(
epoch, self.best_epoch, self.best))
if self.best < f1:
self.model.set_weights(self.best_weights)
def fit_eval(m, trnX, trnY, tstX, tstY, average='micro', **param):
m.set_params(**param)
m.fit(trnX, trnY, tstX, tstY)
pred = m.predict(tstX)
prec, rec, f1, _ = prfs(tstY, pred, average=average)
acc = accuracy_score(tstY, pred)
return (acc, prec, rec, f1)
def get_metrics(pred, pred_prob, y):
"""This function calculates the metrics of AUC_PR, Error, Precision, Recall, and F1 score from
true labels y, prediction pred, or predicted P(y=1|x) pred_prob.
Parameters
----------
pred : iterable (list or np.array)
Predicted labels
pred_prob : iterable (list or np.array)
Predicted P(y=1|x)
y : iterable (list or np.array)
True labels.
"""
precision, recall, f1, _ = zip(*prfs(y, pred))[1]
error = 1 - accuracy_score(y, pred)
area_under_curve = auc_score(y, pred_prob)
metrics_dict = {
"AUC": area_under_curve,
"Error": error,
"Precision": precision,
"Recall": recall,
"F1 score": f1,
}
return metrics_dict
def main():
# Reading Data from folders
X, y = load_data(data_path='./crop_dataset/crop_dataset/')
print(f"Data shape: {X.shape}, Labels: {y.shape}\n")
# Displaying random set of images from data
display_random_set(data=X, labels=y)
# Splitting data into training and testing data, training will consist of 70% of the data and 30% of the remaining
# will be testing data.
x_train, x_val, y_train, y_val = train_test_split(X,
y,
test_size=0.3,
random_state=42,
shuffle=True)
print(
f"Training Data: {x_train.shape}, Training labels: {y_train.shape}\nValidation Data: {x_val.shape}, "
f"Validation labels: {y_val.shape}\n")
# Adjusting labels to be represented as categorical data.
y_train = to_categorical(y=y_train, num_classes=len(np.unique(y)))
y_val = to_categorical(y=y_val, num_classes=len(np.unique(y)))
# Creating Neural network model.
model = build_model(num_classes=len(np.unique(y)),
img_dim=x_train[0].shape)
# To train the model again change train value to True, change to False to not train.
train_model(x=x_train,
y=y_train,
x_val=x_val,
y_val=y_val,
model=model,
train=True)
print("[In progress] Loading H5 model and history file...")
classifier = load_model(filepath='traffic_sign_model.h5')
hist_loaded = load_history(file_name='traffic_sign.pickle')
print("[Done] Loading H5 model and history file...")
# Loading data for testing model.
x_test, y_test = load_test_data(test_data_dir='./test_data/test_data',
test_data_labels_dir='./test_labels.csv')
predictions = classifier.predict_classes(x_test)
accuracy = np.array([
1 if predictions[i] == int(y_test[i]) else 0
for i in range(len(predictions))
])
print(f"Accuracy on test data: {np.mean(accuracy) * 100} %.")
# plotting loss and mse curves for training and validation steps
plot_curves(hist_loaded)
# plotting accuracy bar graph per class
labels = np.unique(y)
precision, recall, f1, support = prfs(y_true=y_test,
y_pred=predictions,
average=None)
accuracy_per_class(labels, precision, recall, f1)
def _compute_metrics(gt_all,
pred_all,
labels,
labels_str,
print_results: bool = False):
per_type = prfs(gt_all, pred_all, labels=labels, average=None)
micro = prfs(gt_all, pred_all, labels=labels, average='micro')[:-1]
macro = prfs(gt_all, pred_all, labels=labels, average='macro')[:-1]
total_support = sum(per_type[-1])
if print_results:
_print_results(per_type,
list(micro) + [total_support],
list(macro) + [total_support], labels_str)
metrics = [m * 100 for m in micro + macro]
return dict(zip(METRIC_LABELS, metrics))
def get_prfs(causality_truth, causality_pred):
assert causality_pred.shape == causality_truth.shape
assert len(causality_pred.shape) == 2
from sklearn.metrics import precision_recall_fscore_support as prfs
precision, recall, F1, _ = prfs(
to_np_array(causality_truth).flatten(),
to_np_array(causality_pred).flatten())
return precision[1], recall[1], F1[1]
def score(train, test, opt):
from sklearn.metrics import precision_recall_fscore_support as prfs
pred = _predict(train, test, opt)
gold = test.labels
prec, rec, f1, _ = prfs(gold, pred, average='macro')
print("Precision {}, Recall: {}, F-score: {}".format(
prec, rec, f1))
def evaluate(logger, tag_group, device, recurrent_model, output_model, loaders,
global_step):
recurrent_model.eval()
output_model.eval()
losses = defaultdict(list)
all_prf = defaultdict(list)
with torch.no_grad():
loss_function_bce = nn.BCEWithLogitsLoss(reduction='mean')
loss_function_mse = nn.MSELoss(reduction='mean')
# each loader is for one sequence
for midifilename, loader in loaders:
print('evaluate midifilename', midifilename)
y_true, y_pred = evaluate_aux(device, recurrent_model,
output_model, loader, losses)
print('y_true.shape', y_true.shape)
print('y_pred.shape', y_pred.shape)
y_pred = (y_pred > 0.5) * 1
# import matplotlib.pyplot as plt
# fig, axes = plt.subplots(nrows=2, sharex=True, sharey=True)
# axes[0].imshow(y_true.T)
# axes[1].imshow(y_pred.T)
# plt.show()
# exit()
p, r, f, _ = prfs(y_true, y_pred, average='micro')
print('p {:>4.2f} r {:>4.2f} f {:>4.2f}'.format(p, r, f))
all_prf['p'].append(p)
all_prf['r'].append(r)
all_prf['f'].append(f)
to_log = {
'{}_prf/p'.format(tag_group): np.mean(all_prf['p']),
'{}_prf/r'.format(tag_group): np.mean(all_prf['r']),
'{}_prf/f'.format(tag_group): np.mean(all_prf['f']),
'{}_losses/mse_frames'.format(tag_group):
np.mean(losses['mse_frames']),
'{}_losses/mse_velocity'.format(tag_group):
np.mean(losses['mse_velocity']),
'{}_losses/bce'.format(tag_group): np.mean(losses['bce'])
}
if logger is not None:
for key, value in to_log.items():
logger.add_scalar(key, value, global_step)
return to_log
def rfFitScore(clf, dftrain, dftrain_y, dftest, dftest_y):
'''random forest classifier fit and score.
clf=RandomForestClassifier, dftrain=train data,
dftrain_y=train data Y, dftest=test data,
dftest_y=test data Y'''
clfit = clf.fit(dftrain, dftrain_y['Y']) # clf.fit(X, y)
imp = clfit.feature_importances_ # ndarray of 562
# clfit.fit_transform( X, y=None ) # returns X_new
new_y = clfit.predict( dftest ) # returns predicted Y
test_score = clfit.score( dftest, dftest_y['Y'] )
print("test score:", test_score) # clfit.oob_score_
if (clf.oob_score):
print("oob score", clfit.oob_score_)
# calculate test score by other means
print("predict True %.3f percent, %d out of %d" % \
((100 * sum(dftest_y['Y'] == new_y) / dftest_y.shape[0]), \
sum(dftest_y['Y'] == new_y), dftest_y.shape[0]))
print("predict False %.3f percent, %d out of %d" % \
((100 * sum(dftest_y['Y'] != new_y) / dftest_y.shape[0]), \
sum(dftest_y['Y'] != new_y), dftest_y.shape[0]))
# new_p = clfit.predict_proba( dftest )
# # probability of each X variable to predict each y class
# print("test predict probabilities head:\n", new_p[:5])
# cross table of variable predictions
ptab = pd.crosstab(dftest_y['Y'], new_y, \
rownames=['actual'], colnames=['predicted'])
print("cross table:\n", ptab)
# accuracy: percent labeled correctly
# precision: true positives / (true positives + true negatives)
# recall: true positives / (true positives + false negatives)
precision, recall, fbeta, support = prfs(dftest_y['Y'], new_y)
print("precision", precision, "\nrecall", recall, \
"\nfbeta", fbeta, "\nsupport", support)
if (clf.oob_score):
return test_score, imp, clfit.oob_score_
else:
return test_score, imp
def log_metrics(y_true, y_pred, metrics_write_path, average=""):
"""
Log the precision/recall/f1-score/support
:param y_true: (ndarray)
True labels
:param y_pred: (ndarray)
Classifier predicted labels
:param metrics_write_path: (str)
Place to write metrics
:param average: (str)
Specifies micro/macro averaging
:return:
None
"""
if not average:
# Default write_file.write both micro and macro averages
with open(metrics_write_path, "a") as write_file:
write_file.write("\nModel Report\n")
timestamp = str(
datetime.fromtimestamp((datetime.timestamp(datetime.now()))))
write_file.write("{}\n".format(timestamp))
write_file.write("----------------------------------------\n")
# Micro Avg
log_metrics(y_true, y_pred, metrics_write_path, average="micro")
# Macro Avg
log_metrics(y_true, y_pred, metrics_write_path, average="macro")
else:
metrics = prfs(y_true, y_pred, average=average)
with open(metrics_write_path, "a") as write_file:
write_file.write("\n")
write_file.write("\n- {} averaging -\n".format(average))
write_file.write(classification_report(y_true, y_pred))
write_file.write("\nPrecision: {}\n".format(metrics[0]))
write_file.write("Recall: {}\n".format(metrics[1]))
write_file.write("F1-Score: {}\n".format(metrics[2]))
write_file.write("Support: {}\n".format(metrics[3]))
write_file.write("Accuracy: {}\n".format(
accuracy_score(y_true, y_pred)))
write_file.write("AUC Score ({}): {}\n".format(
average, roc_auc_score(y_true, y_pred, average=average)))
write_file.write(str(confusion_matrix(y_true, y_pred)) + "\n")
def eval_params(self, nparr):
cutoff = int(self.X.shape[0]*.7)
C = nparr[0]
gamma = nparr[1]
if gamma < 0:
gamma = 1e-3
if C < 0:
C = 1e-3
svm = SVC(C=C, gamma=gamma)
X_train, y_train = self.X[:cutoff], self.labels[:cutoff]
X_test, y_test = self.X[cutoff:], np.array(self.labels[cutoff:])
svm.fit(X_train, y_train)
y_pred = svm.predict(X_test)
# fitness = (y_pred == y_test).sum() / (y_test.shape[0]+.0)
fitness = prfs(y_test, y_pred, average='macro')[2]
print "F-score macro: %.6f achieved with C=%.6f and gamma=%.6f" % (fitness, C, gamma)
return fitness
def rfFitScore(clf, dftrain, dftrain_y, dftest, dftest_y):
'''random forest classifier fit and score.
clf=RandomForestClassifier, dftrain=train data,
dftrain_y=train data Y, dftest=test data,
dftest_y=test data Y'''
clfit = clf.fit(dftrain, dftrain_y['Y']) # clf.fit(X, y)
imp = clfit.feature_importances_ # ndarray of 562
# clfit.fit_transform( X, y=None ) # returns X_new
new_y = clfit.predict(dftest) # returns predicted Y
test_score = clfit.score(dftest, dftest_y['Y'])
print("test score:", test_score) # clfit.oob_score_
if (clf.oob_score):
print("oob score", clfit.oob_score_)
# calculate test score by other means
print("predict True %.3f percent, %d out of %d" % \
((100 * sum(dftest_y['Y'] == new_y) / dftest_y.shape[0]), \
sum(dftest_y['Y'] == new_y), dftest_y.shape[0]))
print("predict False %.3f percent, %d out of %d" % \
((100 * sum(dftest_y['Y'] != new_y) / dftest_y.shape[0]), \
sum(dftest_y['Y'] != new_y), dftest_y.shape[0]))
# new_p = clfit.predict_proba( dftest )
# # probability of each X variable to predict each y class
# print("test predict probabilities head:\n", new_p[:5])
# cross table of variable predictions
ptab = pd.crosstab(dftest_y['Y'], new_y, \
rownames=['actual'], colnames=['predicted'])
print("cross table:\n", ptab)
# accuracy: percent labeled correctly
# precision: true positives / (true positives + true negatives)
# recall: true positives / (true positives + false negatives)
precision, recall, fbeta, support = prfs(dftest_y['Y'], new_y)
print("precision", precision, "\nrecall", recall, \
"\nfbeta", fbeta, "\nsupport", support)
if (clf.oob_score):
return test_score, imp, clfit.oob_score_
else:
return test_score, imp
def _score(self,
gold,
pred,
scores={'precision', 'recall', 'f1-score'},
negative_class=None,
average=None):
""" Return the score for the testset.
"""
from sklearn.metrics import precision_recall_fscore_support as prfs
if average is None:
average = 'macro'
if negative_class:
average = 'binary'
scores = [sc \
if ':' in sc or \
sc not in {'precision', 'recall', 'f1-score'}\
else ':'.join((sc, average))\
for sc in scores]
scores = {k: None for k in scores}
for sc_avg in list(scores):
if ':' in sc_avg:
sc, avg = sc_avg.split(':')
else:
sc = sc_avg
avg = None
if scores[sc_avg] is not None:
continue
if sc not in {'precision', 'recall', 'f1-score', 'accuracy'}:
warning("Skipping unknown score `{}'.".format(sc))
continue
if sc in {'precision', 'recall', 'f1-score'}:
if avg not in {'binary', 'micro', 'macro'}:
warning("Skipping `{}': unknown avgeraging method.".format(
sc_avg))
continue
p, r, f, _ = prfs(gold, pred, average=avg)
scores[':'.join(('precision', avg))] = p
scores[':'.join(('recall', avg))] = r
scores[':'.join(('f1-score', avg))] = f
if sc == 'accuracy':
from sklearn.metrics import accuracy_score
scores['accuracy'] = accuracy_score(gold, pred)
return {k: v for k, v in scores.items()}
def evaluate_test(x_eval, y_eval, res_eval, model, config, test=False):
total_pred = np.array([], dtype=np.float64)
total_y_test = np.array([], dtype=np.int64)
test_batch_num = int(
math.ceil(x_eval.shape[0] / float(config['batch_size'])))
with torch.no_grad():
for i in range(test_batch_num):
begin_index = i * config['batch_size']
end_index = min((i + 1) * config['batch_size'], x_eval.shape[0])
batch_test_x = x_eval[begin_index:end_index]
batch_test_y = y_eval[begin_index:end_index]
batch_test_res = res_eval[begin_index:end_index]
if (config['model_type'] == 'dan'
or config['model_type'] == 'dann'):
batch_test_x = batch_test_x.reshape(-1, 1, 1,
batch_test_x.shape[1])
batch_test_x = torch.from_numpy(batch_test_x).float().to(
config['device'])
batch_test_y = torch.from_numpy(batch_test_y).long().to(
config['device'])
batch_test_res = torch.from_numpy(batch_test_res).float()
output_test, _ = model(batch_test_x, batch_test_x)
_, predicted = torch.max(output_test.data, 1)
# loss = _loss(batch_test_res, batch_test_y, output_test)
total_pred = np.concatenate((total_pred, predicted.cpu()))
total_y_test = np.concatenate((total_y_test, batch_test_y.cpu()))
# overall_loss = _loss(res_eval, y_eval, total_pred)
# overall_loss = criterion(total_pred, total_y_test)
# acc = accuracy_score(total_y_test, total_pred)
if (test == True):
print(classification_report(total_y_test, total_pred, digits=4))
acc = accuracy_score(total_y_test, total_pred)
print("Testing accuracy", acc)
_, _, f1, _ = prfs(total_y_test, total_pred, average='weighted')
# print ("Overall testing/evaluation F1 is ", f1)
return f1
def evaluate_one_loader(cuda, net, loader):
net.eval()
loss_function = nn.BCELoss(reduction='mean')
smoothed_loss = 1.
y_true = []
y_pred = []
for x, y in loader:
if cuda:
x = x.cuda()
y = y.cuda()
y_hat = net.predict(x)
loss = loss_function(y_hat, y)
smoothed_loss = smoothed_loss * 0.9 + loss.detach().cpu().item() * 0.1
y_true.append(y.detach().cpu().numpy())
y_pred.append((y_hat.detach().cpu().numpy() > 0.5) * 1)
y_true = np.vstack(y_true)
y_pred = np.vstack(y_pred)
p, r, f, _ = prfs(y_true, y_pred, average='micro')
return smoothed_loss, p, r, f
def success_metrics(model):
"""
Print evaluation of the model
parameter: Pytorch model, (dataloaders['val'])
return: precision, recall, f1, support
"""
model.eval()
original_labels = []
pred_lst = []
with torch.no_grad():
for i, (inputs, labels) in enumerate(dataloaders['val']):
inputs = inputs.to(device)
labels = labels.to(device)
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
original_labels.extend(labels)
pred_lst.extend(preds)
precision, recall, f1, support = prfs(original_labels, pred_lst, average='weighted')
print("Precision: {:.2%}\nRecall: {:.2%}\nF1 score: {:.2%}".format(precision, recall, f1))
model1.add(Dense(6, activation='relu'))
model1.add(Dense(1, activation='sigmoid', input_dim=11))
model1.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
model1.fit(X_train, y_train, batch_size=10, epochs=100)
y_pred = model1.predict(X_test)
y_pred = (y_pred > 0.5)
print('CM:', confusion_matrix(y_test, y_pred))
print('AC:', ac(y_test, y_pred))
print('F1 scores:', f1(y_test, y_pred))
print('PR:', prfs(y_test, y_pred))
#logistic Regression
from sklearn.linear_model import LogisticRegression as lr
model2 = lr().fit(X_train, y_train)
y_pred = model2.predict(X_test)
y_pred = (y_pred > 0.5)
print('CM:', confusion_matrix(y_test, y_pred))
print('AC:', ac(y_test, y_pred))
print('F1 scores:', f1(y_test, y_pred))
print('PR:', prfs(y_test, y_pred))
from sklearn.neighbors import KNeighborsClassifier as knn
model3 = knn().fit(X_train, y_train)
# # best result: epsilon = 2, threshold = 0.46, F1 = 0.515 (P = 0.468, R = 0.573)
# # y_pred = neighbourhood_difference(y_dists)
#
# plot_thresholds(y_true, y_pred, False, 'binary')
# Measurements on test data
if __name__ == '__main__':
data = config.get_seg_data('test')
print("Loading the data")
y = np.load(data['y'])
y_true = np.load(data['y_true_lm'])
T = 0.5
y_pred = compute_distance(y, euclidean_distances) > T
P, R, F, S = prfs(y_true, y_pred, average='binary')
print('euclidean distance: threshold = %.2f' % T)
print_measurements(y_true, y_pred)
y_dists = compute_distance(y, cosine_distances)
T = 0.92
y_pred = y_dists > T
print('cosine distance: threshold = %.2f' % T)
print_measurements(y_true, y_pred)
T = 0.33
y_pred = compute_distance(y, manhattan_distances) > T
print('manhattan distance: threshold = %.2f' % T)
print_measurements(y_true, y_pred)
cd_loss = criterion(cd_preds, labels)
loss = cd_loss
loss.backward()
optimizer.step()
cd_preds = cd_preds[-1]
_, cd_preds = torch.max(cd_preds, 1)
# Calculate and log other batch metrics
cd_corrects = (
100 *
(cd_preds.squeeze().byte() == labels.squeeze().byte()).sum() /
(labels.size()[0] * (opt.patch_size**2)))
cd_train_report = prfs(labels.data.cpu().numpy().flatten(),
cd_preds.data.cpu().numpy().flatten(),
average='binary',
pos_label=1)
train_metrics = set_metrics(train_metrics, cd_loss, cd_corrects,
cd_train_report, scheduler.get_lr())
# log the batch mean metrics
mean_train_metrics = get_mean_metrics(train_metrics)
for k, v in mean_train_metrics.items():
writer.add_scalars(str(k), {'train': v}, total_step)
# clear batch variables from memory
del batch_img1, batch_img2, labels
scheduler.step()
###########################################################
################# ConvNet Evaluation ######################
###########################################################
print "\n fitting cnn model on %d samples..." % x_train.shape[0]
# hist = cnn_100.model.fit(x_train, y_train, nb_epoch=10)
# cnn_100.save_model(hist)
# print 'saved hist: ', str(hist)
cnn_200_pred = cnn_200.model.predict(x_test)
y_true = np.array([np.argmax(i) for i in y_test])
y_pred = np.array([np.argmax(i) for i in cnn_200_pred])
cnn_acc = accuracy_score(y_true, y_pred)
p, r, f, s = prfs(y_true, y_pred, average='macro')
print
print "cnn_20 stats: precision=%.3f, recall=%.3f, f-score=%.3f, acc: %.3f" % (
p, r, f, cnn_acc)
###########################################################
################# LogReg/SVM Preparation ##################
###########################################################
print 'reading text data...'
t = time()
text_reader = TweetCorpusReader('../data/', text_only=False)
t = time() - t
print 'time elapsed: %.0fs' % t
def callback(epoch, model):
if epoch % 250 != 0:
return
perf = []
for X, Z, y in ((experiment.X[ts], experiment.Z[ts],
experiment.y[ts]),
(experiment.X[tr], experiment.Z[tr],
experiment.y[tr])):
yhat = model.predict(X)
y_loss = model.loss_y(X, y)
z_loss = model.loss_z(X, Z)
y_perf = list(prfs(y, yhat, average='binary')[:3])
perf.extend(y_perf + [y_loss, z_loss])
#print(Z[0])
#print(Z_hat[0])
if args.experiment.startswith('color') and args.record_lime:
similarities, dispersions, times_senn, times_lime = [], [], [], []
for i in selection:
x = experiment.X[i].reshape(1, -1)
z_senn, t_senn = model.explain(x, return_runtime=True)
z_senn = z_senn.ravel()
z_lime, Z_lime, t_lime = \
experiment.explain_lime(model, tr, i,
n_repeats=args.lime_repeats,
n_samples=args.lime_samples,
n_features=args.lime_features)
times_senn.append(t_senn)
times_lime.append(t_lime)
n_nonzeros = len(np.nonzero(z_lime)[0])
similarities.append(
_whatever_at_k(z_senn, z_lime, n_nonzeros))
n_repeats = len(Z_lime)
dispersions.append(1 / (n_repeats * (n_repeats - 1)) * \
np.sum(pairwise_distances(Z_lime, Z_lime)))
path = basename + '__fold={}__instance={}__epoch={}'.format(
k, i, epoch)
experiment.dump_explanation(path + '_senn.png',
experiment.X[i],
experiment.Z[i], z_senn)
experiment.dump_explanation(path + '_lime.png',
experiment.X[i],
experiment.Z[i], z_lime)
perf.extend([
np.mean(similarities),
np.mean(dispersions),
np.mean(times_senn),
np.mean(times_lime),
])
print('epoch {} : {}'.format(epoch, perf))
return perf
print '#'*40
print 'NO Twitter Features'
print 'SVM - Linear Kernel'
print '#'*40
print 'ACC\tPR\tRE\tF1'
print '#'*40
i=1
for tr, ts in KFold(n=len(normalized_corpus), n_folds=10):
train = X[tr]
test = X[ts]
clf = LinearSVC()
clf.fit(train, labels[tr])
ytrue = labels[ts]
ypred = clf.predict(test)
acc = (ytrue == ypred).sum() / (len(ypred)+.0)
p, r, f, s = prfs(ytrue, ypred, average='binary')
accs.append(acc)
ps.append(p)
rs.append(r)
fs.append(f)
print "%.2f\t%.2f\t%.2f\t%.2f KFoldRnd%d" % (acc,p,r,f,i)
i += 1
print '#'*40
print 'Mean accuracy: %.2f' % (np.mean(accs))
print 'Mean precision: %.2f' % (np.mean(ps))
print 'Mean recal: %.2f' % (np.mean(rs))
print 'Mean f-score: %.2f' % (np.mean(fs))
accs = []
ps = []
print(
'Loading the paragraph trained classifier trained on data processed by'
' threshold_half_max function')
classifier = load_pickle(config.classifier_par_half_max)
threshold = 0.39
y_true = process_y(data, threshold_half_max)
print("Loading x")
x = load_sparse_csr(data['x'])
else:
threshold = 0.3
vectorizer = load_pickle(config.vectorizer)
binarizer = load_pickle(config.binarizer)
print('Loading the classifier')
classifier = load_pickle(config.classifier)
corpus, topics = build_corpus_and_topics(config.data['test'])
print('Transforming corpus by vectorizer')
x = vectorizer.transform(corpus)
print('Transforming article topics by binarizer')
y_true = binarizer.transform(topics)
del vectorizer, binarizer, corpus, topics
y_pred = predict_tuned(x, classifier, threshold)
P, R, F, S = prfs(y_true, y_pred, average='samples')
print('F1 = %.3f (P = %.3f, R = %.3f)' % (F, P, R))
def performanceSummary(trueLabels,predictedLabels):
accuracy=acu(trueLabels,predictedLabels)
precision,recall,fscore,support=prfs(trueLabels,predictedLabels)
return [accuracy.tolist()]+precision.tolist()+recall.tolist()+fscore.tolist()+support.tolist()
# The twitter-specific tokenizer makes the parsing slow, however # the accuracy is much improved with it. X_train, train_feats = pr.process(X_train, verbose=True) # ~7min vectoring phase X_mat = pr.fit_transform(X_train, saveVectorizer=False, saveMatrix=False, verbose=True) X_test, test_feats = pr.process(X_test, verbose=True) X_test = pr.transform(X_test, saveMatrix=False, verbose=True) # Compare the accuracy with and w/o the twitter-specific features. # Must scale the features matrix before concatenating with the ngrams matrix. print '\nTF-IDF Unigrams and Bigrams || Logistic Regression classifier' print '-'*40 clf = LR() # Roughly 3 minutes on training t0 = time.time() print 'Training on %d samples...' % (X_mat.shape[0]) clf.fit(X_mat, y_train) print 'Training time: %.0fs' % ((time.time()-t0)) print 'Testing on %d samples...' % (X_test.shape[0]) y_pred = clf.predict(X_test) acc = (y_pred==y_test).sum()/(len(y_pred)+.0) f1 = prfs(y_test, y_pred, average="macro")[-2] roc_auc = roc_auc_score(y_test, y_pred) print '\nReport\n'+'-'*40 print 'Accuracy: %.4f\nMacro F-1 Score: %.4f\nROC_AUC Score: %.4f' % (acc, f1, roc_auc)
if classPred[bind]==0:
voxPred = branch_1[bind]
else:
voxPred = branch_2[bind]
revVoxPred = voxPred[:,:,::-1,:]
voxPred = np.maximum(revVoxPred,voxPred)
voxPredLabels = voxPred[grid_xyz[:,0],grid_xyz[:,1],grid_xyz[:,2],0]
tree = ckdt(grid_xyz_pc)
_,inds = tree.query(cloud,k=1)
pcPred = voxPredLabels[inds]
thresh_pr = [] ; thresh_re = []
for thresh in np.arange(0.0,1.01,0.01):
threshPred = np.array(pcPred)
threshPred[threshPred >= thresh] = 1
threshPred[threshPred < thresh] = 0
p,r,f,s = prfs(seg.flatten(),threshPred.flatten(),average='binary')
thresh_pr.append(p)
thresh_re.append(r)
prs.append(thresh_pr)
recs.append(thresh_re)
meanacc += acc
if j%5==0:
print('BATCH ' + str(j) + ' of ' + str(len(trainInstances)/args['batch_size']),' PREDICTION ACCURACY: ',meanacc/(j+1))
except StopIteration:
if not os.path.exists(args['outputDir']):
os.makedirs(args['outputDir'])
prs = np.array(prs)
recs = np.array(recs)
pr_mean = np.mean(prs,axis=0)
rec_mean = np.mean(recs,axis=0)
def permutations(data, shortest, longest, iterations, kernel, prefix, suffix):
#initialize text data vectorizer
if (shortest == 1) & (longest == 1):
AAs = [
'a', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'k', 'l', 'm', 'n', 'p',
'q', 'r', 's', 't', 'v', 'w', 'y'
]
cv = CountVectorizer(analyzer='char',
ngram_range=(shortest, longest),
vocabulary=AAs)
else:
cv = CountVectorizer(analyzer='char', ngram_range=(shortest, longest))
dataVect = cv.transform(data.Sequence)
# repeat for each kernel type:
for k in kernel:
#initialize feature use arrays
precisions = pd.DataFrame(columns=data.Genus.unique())
recalls = pd.DataFrame(columns=data.Genus.unique())
fbetas = pd.DataFrame(columns=data.Genus.unique())
micros = pd.DataFrame(columns=['Precision', 'Recall', 'Fbeta'])
macros = pd.DataFrame(columns=['Precision', 'Recall', 'Fbeta'])
scores = pd.DataFrame(columns=("iteration", "score"))
#initialize the genus classification for this ngram
for i in data.Genus.unique():
data[i] = 0
#repeat n times
for j in range(iterations):
randomstate = j + 5342 #make sure the randomstate input is at least 4 digits and repeatable
#initialize classifier
#uses default settings for the classifier, see scikit-learn documentation sklearn.svm.SVC for details
clf = SVC(kernel=k)
#build training and test data sets
X_train, X_test, y_train, y_test = train_test_split(
dataVect,
data.Genus,
test_size=0.5,
stratify=data.Genus,
random_state=randomstate)
#Scale the data to the training set
StSc = StandardScaler(copy=True, with_mean=False)
StSc.fit(X_train)
X_sc_train = StSc.transform(X_train)
X_sc_test = StSc.transform(X_test)
X_scaled = StSc.transform(dataVect)
#train the classifier
clf.fit(X_sc_train, y_train)
#make predictions for the original dataset
y_pred = clf.predict(X_scaled)
score = clf.score(X_sc_test, y_test)
scores.loc[j] = [j, score]
data['prediction'] = y_pred
#increment the Genus column counter for each Genus
for c in clf.classes_:
data.loc[data.prediction == c, c] += 1
#record simulation data
metrics = prfs(data.Genus, y_pred)
precisions.loc[j] = metrics[0]
recalls.loc[j] = metrics[1]
fbetas.loc[j] = metrics[2]
metrics = prfs(data.Genus, y_pred, average='micro')
micros.loc[j, 'Precision'] = metrics[0]
micros.loc[j, 'Recall'] = metrics[1]
micros.loc[j, 'Fbeta'] = metrics[2]
metrics = prfs(data.Genus, y_pred, average='macro')
macros.loc[j, 'Precision'] = metrics[0]
macros.loc[j, 'Recall'] = metrics[1]
macros.loc[j, 'Fbeta'] = metrics[2]
data.to_csv("{0}_SVM_{1}_{2}.csv".format(prefix, k, suffix))
scores.to_csv("{0}_SVM_{1}_{2}.scores".format(prefix, k, suffix))
precisions.to_csv("{0}_SVM_{1}{2}.precision".format(prefix, k, suffix))
recalls.to_csv("{0}_SVM_{1}_{2}.recall".format(prefix, k, suffix))
fbetas.to_csv("{0}_SVM_{1}_{2}.fbeta".format(prefix, k, suffix))
micros.to_csv("{0}_SVM_{1}_{2}.micro".format(prefix, k, suffix))
macros.to_csv("{0}_SVM_{1}_{2}.macro".format(prefix, k, suffix))
def precision_recall(beta_true, beta_pred):
b_true = beta_true!=0
b_pred = beta_pred!=0
p, r, f, s = prfs(b_true, b_pred, pos_label=1)
return p, r
def train_dnn(x_t, y_t, res_t, x_eval, y_eval, res_eval, x_test, y_test,
config):
# Start training process
counter = 0
best_f1 = 0.0
f1_total = 0.0
epoch_no_improve = 0
n_epoch_stop = 3
#momentum = 0.9
#log_interval = 10
#l2_decay = 5e-4
batch_num = int(x_t.shape[0] / config['batch_size'])
if (config['model_type'] == 'dan'):
training_model = transfer_model.DANNet(config).to(config['device'])
else:
training_model = transfer_model.CNNModel(config).to(config['device'])
loss_class = torch.nn.NLLLoss().to(config['device'])
loss_domain = torch.nn.NLLLoss().to(config['device'])
for epoch in range(config['num_epochs']):
# Start training process
if (config['model_type'] == 'dan'):
LEARNING_RATE = config['learning_rate'] / math.pow(
(1 + 10 * (epoch - 1) / config['num_epochs']), 0.75)
print('learning rate{: .4f}'.format(LEARNING_RATE))
optimizer = torch.optim.SGD([
{
'params': training_model.sharedNet.parameters()
},
{
'params': training_model.cls_fc.parameters(),
'lr': LEARNING_RATE
},
],
lr=LEARNING_RATE / 10,
momentum=config['momentum'],
weight_decay=config['l2_decay'])
elif (config['model_type'] == 'dann'):
optimizer = torch.optim.Adam(training_model.parameters(),
lr=config['learning_rate'])
for batch in range(batch_num):
training_model.train()
batch_index = generate_batch(x_t.shape[0], config['batch_size'])
test_batch_idx = generate_batch(x_test.shape[0],
config['batch_size'])
batch_x = x_t[batch_index]
batch_y = y_t[batch_index]
batch_res = res_t[batch_index]
batch_x = batch_x.reshape(-1, 1, 1, batch_x.shape[1])
batch_x = torch.from_numpy(batch_x).float().to(config['device'])
batch_y = torch.from_numpy(batch_y).long().to(config['device'])
batch_res = torch.from_numpy(np.array(batch_res)).float()
batch_test_x = x_test[test_batch_idx].reshape(
-1, 1, 1, x_test.shape[1])
batch_test_x = torch.from_numpy(batch_test_x).float().to(
config['device'])
# Forward pass
#print ("Shape source and target", batch_x.shape, batch_test_x.shape)
alpha = 0.0
if (config['model_type'] == 'dan'):
label_pred, loss_mmd = training_model(batch_x, batch_test_x)
batch_test_x = x_test[test_batch_idx].reshape(
-1, 1, 1, x_test.shape[1])
batch_test_x = torch.from_numpy(batch_test_x).float().to(
config['device'])
label_pred, loss_mmd = training_model(batch_x, batch_test_x)
#loss = criterion(outputs, batch_y)
# loss = _loss(batch_res, batch_y, outputs)
loss_cls = F.nll_loss(F.log_softmax(label_pred, dim=1),
batch_y)
gamma = 2 / (1 + math.exp(-10 *
(epoch) / config['num_epochs'])) - 1
err = loss_cls + gamma * loss_mmd
# Backward and optimize
optimizer.zero_grad()
err.backward()
optimizer.step()
#print ("Label prediction", label_pred)
tr_pred = np.argmax(label_pred.cpu().detach(), 1)
#tr_pred = np.argmax(outputs.cpu().detach(), 1)
counter += 1
training_model.eval()
_, _, f1, _ = prfs(batch_y.cpu(), tr_pred, average='weighted')
f1_total += f1
elif (config['model_type'] == 'dann'):
p = float(counter + epoch *
x_t.shape[0]) / config['num_epochs'] / x_t.shape[0]
alpha = 2. / (1. + np.exp(-10 * p)) - 1
training_model.zero_grad()
#batch_x = batch_x.reshape(-1, 1, 1, x_test.shape[1])
#class_label = torch.LongTensor(batch_size)
domain_label = torch.zeros(config['batch_size'])
domain_label = domain_label.long().to(config['device'])
# Using source data
class_output, domain_output = training_model(batch_x, alpha)
err_s_label = loss_class(class_output, batch_y)
err_s_domain = loss_domain(domain_output, domain_label)
### Using target data
#x_test_batch = torch.FloatTensor(batch_size, 1, x_test.shape[1], x_test.shape[1])
domain_label = torch.ones(config['batch_size'])
domain_label = domain_label.long().to(config['device'])
_, domain_output = training_model(batch_test_x, alpha)
err_t_domain = loss_domain(domain_output, domain_label)
err = err_t_domain + err_s_domain + err_s_label
#print ("Error is ", err)
err.backward()
optimizer.step()
training_model.eval()
class_output, _ = training_model(batch_x, alpha)
tr_pred = class_output.data.cpu().max(1, keepdim=True)[1]
_, _, f1, _ = prfs(batch_y.cpu(), tr_pred, average='weighted')
f1_total += f1
if (counter + 1) % 100 == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Best F1: {:.4f}'
.format(epoch + 1, config['num_epochs'], batch + 1,
batch_num, err.item(), f1))
counter += 1
print("------epoch : ", epoch, " Loss: ", err.item(), " Training F1:",
round((f1_total / batch_num), 4))
training_model.eval()
eval_f1 = evaluate_test(x_eval, y_eval, res_eval, training_model,
config, False)
f1_total += eval_f1
if (eval_f1 >= best_f1):
best_f1 = eval_f1
torch.save(training_model.state_dict(),
config['ckpt_path'] + 'best_model.pth')
print("Current F1 score (evaluation) ", eval_f1)
print("Best F1 score (evaluation) ", best_f1)
print("All completed")
concat = Dropout(.5)(concat)
output = Dense(2, activation='softmax', name='dense')(concat)
model = Model([input1, input2], output, name='deep_lstm')
model.compile("adam", "binary_crossentropy", metrics=['accuracy'])
model.summary()
from keras.utils import plot_model
# plot_model(model, to_file='attn-cnn-lstm.png', show_layer_names=True, show_shapes=True)
############################################################
################ Training and Eval #########################
############################################################
labels = np.load("labels.npy")
x1_train, x1_test, x2_train, x2_test, labels_train, labels_test = train_test_split(
data1, data2, labels, test_size=0.15)
N_EPOCHS = 10
history = model.fit([x1_train, x2_train],
labels_train,
batch_size=128,
validation_split=.15,
epochs=N_EPOCHS,
callbacks=[EarlyStopping(patience=2)])
preds = model.predict([x1_test, x2_test]).argmax(axis=-1)
y_true = labels_test.argmax(axis=-1)
print(prfs(y_true, preds))
print(accuracy_score(y_true, preds))
for m in range(len(mnb.class_count_)):
temp = np.argsort(mnb.coef_[m])[-20:]
for e in temp[-10:]:
if features[e] in aa_top_10.columns:
aa_top_10[features[e]][mnb.classes_[m]] += 1
else:
aa_top_10[features[e]] = 0
aa_top_10[features[e]][mnb.classes_[m]] = 1
for k in temp[:10]:
if features[k] in aa_next_10.columns:
aa_next_10[features[k]][mnb.classes_[m]] += 1
else:
aa_next_10[features[k]] = 0
aa_next_10[features[k]][mnb.classes_[m]] = 1
metrics = prfs(data.Genus, y_pred)
#record simulation data
precisions.loc[j] = metrics[0]
recalls.loc[j] = metrics[1]
fbetas.loc[j] = metrics[2]
metrics = prfs(data.Genus, y_pred, average='micro')
micros.loc[j, 'Precision'] = metrics[0]
micros.loc[j, 'Recall'] = metrics[1]
micros.loc[j, 'Fbeta'] = metrics[2]
metrics = prfs(data.Genus, y_pred, average='macro')
macros.loc[j, 'Precision'] = metrics[0]
macros.loc[j, 'Recall'] = metrics[1]
macros.loc[j, 'Fbeta'] = metrics[2]
#Model fitting
clf = rfc()
#Calling feature selection methods
fs = feature_selection()
#clf,x_train,x_test,x_final_test,y_out = fs.PCASelection(x_train,y_train_binary,x_test,y_test_binary,x_final_test,clf)
clf, x_train, x_test, x_final_test, y_out = fs.KBest(x_train, y_train_binary,
x_test, y_test_binary,
x_final_test, clf)
clf.fit(x_train, y_train_binary)
y_out = clf.predict(x_test)
#Printing scores
score = clf.score(x_test, y_test_binary)
print "Score : ", score
print "Precision recall f-score support : ", prfs(y_test_binary, y_out)
#Cross validation
folds = 2
print "\nManual ", folds, " fold cross validation score"
CV(x_orig_train, y_orig_train_binary, clf, folds)
scores = cross_val_score(clf, x_orig_train, y_orig_train_binary, cv=10)
#Checking with inbuilt CV function
print "\nChecking with inbuilt function"
skf = KFold(n_splits=folds, shuffle=False)
skfscore = cross_val_score(clf, x_orig_train, y_orig_train_binary, cv=skf)
print skfscore
#Manual Parameter tuning
print "\nManual parameter tuning"