def fmeasure(y_true, y_pred):
"""Computes the f-measure, the harmonic mean of precision and recall."""
to_keep = y_pred[2] == 0
y_pred = y_pred[to_keep]
y_true = y_true[to_keep]
pre = precision(y_true, y_pred)
rec = recall(y_true, y_pred)
return (2 * pre * rec) / 2
def test_recall():
y_true = K.variable(np.array([0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 0, 1, 0, 0]))
y_pred = K.variable(np.array([1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0]))
# Calculated using sklearn.metrics.recall_score
expected = 0.2857142857142857
actual = K.eval(metrics.recall(y_true, y_pred))
epsilon = 1e-05
assert expected - epsilon <= actual <= expected + epsilon
def test_recall():
y_true = K.variable(np.array([0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 0, 1, 0, 0]))
y_pred = K.variable(np.array([1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0]))
# Calculated using sklearn.metrics.recall_score
expected = 0.2857142857142857
actual = K.eval(metrics.recall(y_true, y_pred))
epsilon = 1e-05
assert expected - epsilon <= actual <= expected + epsilon
def predict_and_eval(model, X_train, y_train, X_test, y_test, threshold = None):
logging.info("Performing prediction on train and test for evaluation ...")
y_pred_train = model.predict(X_train, batch_size=batch_size, verbose=1)
y_pred_test = model.predict(X_test, batch_size=batch_size, verbose=1)
eval_types = ['Train', 'Test']
logs = {}
for e, eval_type in enumerate(eval_types):
# print "[%s]" % eval_type
metric_prefix = '' if e == 0 else 'val_'
X_eval = X_train if e == 0 else X_test
y_eval = y_train if e == 0 else y_test
y_pred_eval = y_pred_train if e == 0 else y_pred_test
# threshold = 0.48
# y_pred_eval = (0.5 - threshold) + y_pred_eval
y_eval = y_eval.astype(float)
if threshold != None:
y_eval = K.clip((0.5 - threshold) + y_eval, 0., 1.)
logs[metric_prefix + 'loss'] = metrics.binary_crossentropy(y_eval, y_pred_eval).eval()
logs[metric_prefix + 'acc'] = metrics.binary_accuracy(y_eval, y_pred_eval).eval()
logs[metric_prefix + 'precision'] = metrics.precision(y_eval, y_pred_eval).eval()
logs[metric_prefix + 'recall'] = metrics.recall(y_eval, y_pred_eval).eval()
logs[metric_prefix + 'fbeta_score'] = metrics.fmeasure(y_eval, y_pred_eval).eval()
# log_file.write("%d,%.5f,%s,%.5f\n" % (epoch, threshold, eval_type, average_faux_jaccard_similarity))
# print "%d,%.5f,%s,%.4f" % (epoch, threshold, eval_type, average_faux_jaccard_similarity)
metrics_line = ''
for s in ['loss', 'acc', 'precision', 'recall', 'fbeta_score']:
metrics_line += "%s: %.5f %s: %.5f - " %(s, logs[s], 'val_'+s, logs['val_' +s])
logging.info(metrics_line)
return logs
def batch_pairwise_metrics(y_true, y_pred):
#assert K.get_variable_shape(y_true)[1] == K.get_variable_shape(y_pred)[1]
num_classes = K.get_variable_shape(y_pred)[1]
preds_cats = K.argmax(y_pred, axis=1)
preds_one_hot = K.one_hot(preds_cats, num_classes)
overall_precision = [None for _ in range(num_classes)]
overall_recall = [None for _ in range(num_classes)]
overall_fmeasure = [None for _ in range(num_classes)]
out_dict = {}
for cc in range(num_classes):
#Metrics should take 1D arrays which are 1 for positive, 0 for negative
two_true, two_pred = y_true[:, cc], preds_one_hot[:, cc]
cur_dict = {
'precision/%02d' % cc:
kmetrics.precision(two_true, two_pred),
'recall/%02d' % cc:
kmetrics.recall(two_true, two_pred),
'fmeasure/%02d' % cc:
kmetrics.fmeasure(two_true, two_pred),
'binary_accuracy/%02d' % cc:
kmetrics.binary_accuracy(two_true, two_pred),
'act_pos/%02d' % cc:
K.sum(two_true),
'pred_pos/%02d' % cc:
K.sum(two_pred)
}
out_dict.update(cur_dict)
overall_precision[cc] = cur_dict['precision/%02d' % cc]
overall_recall[cc] = cur_dict['recall/%02d' % cc]
overall_fmeasure[cc] = cur_dict['fmeasure/%02d' % cc]
out_dict.update(make_stats('precision', overall_precision))
out_dict.update(make_stats('recall', overall_recall))
out_dict.update(make_stats('fmeasure', overall_fmeasure))
return out_dict