コード例 #1
0
def twenty_one_cm_rmse(y_true, y_pred, discretise = False):
	"""
	used in 21cm paper to evaluate performance
	"""
	if discretise:
		y_p = tools.round_probabilities(y_pred)
	else:
		y_p = y_pred
	return np.sqrt(np.mean((y_true - y_p)**2.)) / np.max(np.abs(y_true))
コード例 #2
0
def regression_r2_score(y_true, y_pred, discretise = False):
	"""
	returns well-known R^2 regression coefficient.
	calculates separately for each output, then takes average over these
	"""
	if discretise:
		y_p = tools.round_probabilities(y_pred)
	else:
		y_p = y_pred
	y_t = y_true
	return sklearn.metrics.r2_score(y_t, y_p, multioutput='uniform_average')
コード例 #3
0
def explained_variance_score(y_true, y_pred, discretise = False):
	"""
	not entirely sure what it is, but see
	https://scikit-learn.org/stable/modules/model_evaluation.html#explained-variance-score
	not sure whether it makes sense to discretise
	"""
	if discretise:
		y_p = tools.round_probabilities(y_pred)
	else:
		y_p = y_pred
	y_t = y_true
	return sklearn.metrics.explained_variance_score(y_t, y_p, multioutput='uniform_average')
コード例 #4
0
def mean_absolute_error(y_true, y_pred, discretise = False):
	"""
	requires input arrays to be same np.dtype.
	returns average, not sum of errors
	discretising (for classification problems) makes little sense to me,
	but may be necessary in some obscure scenarios
	"""
	if discretise:
		y_p = tools.round_probabilities(y_pred)
	else:
		y_p = y_pred
	mae_a = tf.Session().run(tf.keras.losses.mean_absolute_error(y_true, y_p))
	return mae_a.mean()
コード例 #5
0
def twenty_one_cm_rmse_ts(y_true, y_pred, n_z = 136, discretise = False):
	"""
	used in 21cm paper to evaluate performance,
	returns rmse per timeseries (assumes 136 contiguous elements form ts)
	assumes m is divisible by n_z
	"""
	if discretise:
		y_p = tools.round_probabilities(y_pred)
	else:
		y_p = y_pred
	errs = []
	for i in range(y_pred.shape[0] / n_z - 1):
		errs.append(np.sqrt(np.mean((y_true[i * n_z:(i + 1) * n_z] - y_p[i * n_z:(i + 1) * n_z])**2.)) / np.max(np.abs(y_true[i * n_z:(i + 1) * n_z])))
	errs.append(np.sqrt(np.mean((y_true[-1 * n_z:] - y_p[-1 * n_z:])**2.)) / np.max(np.abs(y_true[-1 * n_z:])))
	return np.array(errs)
コード例 #6
0
def decode_onehot_outputs(y_true, y_pred):
	"""
	converts one-hot true ys and predicted ys (probabilities)
	into decoded vectors, for use in sklearn.metric functions.
	e.g.
	y_true = [[0,0,1], [0,1,0], [1,0,0]]
	y_pred = [[0.3,0.4,0.3], [0.1,0.5,0.4], [0.2,0.1,0.7]]
	- >
	y_true = [2,1,0]
	y_pred = [1,1,2]
	"""
	y_pred_r = tools.round_probabilities(y_pred)
	y_pred_d = tools.decode_onehot(y_pred_r)
	y_true_d = tools.decode_onehot(y_true)
	return y_pred_d, y_true_d
コード例 #7
0
def categorical_crossentropy(y_true, y_pred, discretise = False):	
	"""
	requires input arrays to be np.float64 type.
	returns average, not sum of crossentropy.
	discretising (for classification problems) makes little sense to me,
	but may be necessary in some obscure scenarios

	"""
	if discretise:
		y_p = tools.round_probabilities(y_pred)
	else:
		y_p = y_pred
	y_t = y_true
	y_true_t = tf.Variable(y_t, dtype=tf.float64)
	y_pred_t = tf.Variable(y_p, dtype=tf.float64)
	with tf.Session() as sess:
		sess.run(tf.global_variables_initializer())
		cce_a = sess.run(tf.keras.losses.categorical_crossentropy(y_true_t, y_pred_t))
	return cce_a.mean()