def rmse(y_true, y_pred):
'''
Root mean squared error.
Parameters
----------
y_true: ndarray
Ground truth
y_pred: ndarray
Array of predictions
Returns
-------
rmsle: float
Root mean squared error
References
----------
.. [1] https://www.kaggle.com/wiki/RootMeanSquaredError
'''
# Check shapes
y_true, y_pred = align_shape(y_true, y_pred)
check_shapes(y_true, y_pred)
return np.sqrt(((y_true - y_pred)**2).mean())
def rmsle(y_true, y_pred):
'''
Root mean squared logarithmic error.
Parameters
----------
y_true: ndarray
Ground truth
y_pred: ndarray
Array of predictions
Returns
-------
rmsle: float
Root mean squared logarithmic error
References
----------
.. [1] https://www.kaggle.com/wiki/RootMeanSquaredLogarithmicError
.. [2] https://www.slideshare.net/KhorSoonHin/rmsle-cost-function
'''
# Check shapes
y_true, y_pred = align_shape(y_true, y_pred)
check_shapes(y_true, y_pred)
return np.sqrt(((np.log(y_pred + 1) - np.log(y_true + 1))**2).mean())
def wmae(y_true, y_pred, weights):
'''
Weighted mean absolute error.
Parameters
----------
y_true: ndarray
Ground truth
y_pred: ndarray
Array of predictions
Returns
-------
rmsle: float
Weighted mean absolute error
References
----------
.. [1] https://www.kaggle.com/wiki/WeightedMeanAbsoluteError
'''
# Check shapes
y_true, y_pred = align_shape(y_true, y_pred)
check_shapes(y_true, y_pred)
return (weights * np.abs(y_true - y_pred)).mean()
def log_loss(y_true, y_pred):
'''
Logarithmic loss
Parameters
----------
y_true: numpy.ndarray
Targets
y_pred: numpy.ndarray
Class probability
Returns
-------
score: float
Logarithmic loss score
References
----------
.. [1] https://www.kaggle.com/wiki/LogLoss
.. [2] http://www.exegetic.biz/blog/2015/12/making-sense-logarithmic-loss/
.. [3] http://wiki.fast.ai/index.php/Log_Loss
'''
# Check shapes
check_shapes(y_true, y_pred)
y_true, y_pred = align_shape(y_true, y_pred)
# Checking values
if not (y_pred > 0).all():
raise ValueError("Prediction array contains zeroes!")
return -(y_true * np.log(y_pred)).sum(axis=1).mean()
def mean_average_precision(y_true, y_pred):
# TODO: definition of query!!!
'''
Mean average precision
Parameters
----------
y_true: numpy.ndarray
Targets
y_pred: numpy.ndarray
Class predictions (0 or 1 values only)
Returns
------
score: float
Mean average precision score
References
----------
.. [1] https://en.wikipedia.org/wiki/Information_retrieval#Mean_average_precision
'''
# Check shapes
check_shapes(y_true, y_pred)
y_true, y_pred = align_shape(y_true, y_pred)
def mcc(y_true, y_pred):
'''
Matthews Correlation Coefficient
Parameters
----------
y_true: numpy.ndarray
Targets
y_pred: numpy.ndarray
Class predictions (0 or 1 values only)
Returns
------
score: float
Matthews Correlation Coefficient score
References
----------
.. [1] https://lettier.github.io/posts/2016-08-05-matthews-correlation-coefficient.html
.. [2] https://en.wikipedia.org/wiki/Matthews_correlation_coefficient
'''
# Check shapes
check_shapes(y_true, y_pred)
y_true, y_pred = align_shape(y_true, y_pred)
# Confusion matrix values
tp, tn, fp, fn = confusion_binary(y_true, y_pred)
numerator = tp * tn - fp * fn
denominator = np.sqrt((tp + fp) * (fn + tn) * (fp + tn) * (tp + fn))
return numerator / denominator
def mean_utility(y_true, y_pred, weights):
'''
Mean utility
Parameters
----------
y_true: numpy.ndarray
Targets
y_pred: numpy.ndarray
Class predictions (0 or 1 values only)
Returns
------
score: float
Mean utility score
References
----------
.. [1] https://www.kaggle.com/wiki/MeanUtility
.. [2] https://en.wikipedia.org/wiki/Multi-label_classification
Notes
-----
The higher the better.
'''
# Check shapes
check_shapes(y_true, y_pred)
y_true, y_pred = align_shape(y_true, y_pred)
# Weights assignment
w_tp, w_tn, w_fp, w_fn = weights
tp, tn, fp, fn = confusion_binary(y_true, y_pred)
return w_tp * tp + w_tn * tn + w_fp * fp + w_fn * fn
def hamming_loss(y_true, y_pred):
'''
Hamming loss
Parameters
----------
y_true: numpy.ndarray
Targets
y_pred: numpy.ndarray
Class predictions (0 or 1 values only)
Returns
------
score: float
Hamming loss score
References
----------
.. [1] https://www.kaggle.com/wiki/HammingLoss
.. [2] https://en.wikipedia.org/wiki/Multi-label_classification
Notes
-----
The smaller the better.
'''
# Check shapes
check_shapes(y_true, y_pred)
y_true, y_pred = align_shape(y_true, y_pred)
# Logical values only!
check_binary(y_true, y_pred)
return np.logical_xor(y_pred, y_true).mean(axis=1).mean()
def mce(y_true, y_pred):
'''
Mean consequential error
Parameters
----------
y_true: numpy.ndarray
Targets
y_pred: numpy.ndarray
Class predictions (0 or 1 values only)
Returns
-------
score: float
Mean consequential error score
References
----------
.. [1] https://www.kaggle.com/wiki/MeanConsequentialError
.. [2] http://www.machinelearning.ru/wiki/images/5/59/PZAD2016_04_errors.pdf (RU)
Notes
-----
The higher the better.
'''
# Check shapes
check_shapes(y_true, y_pred)
y_true, y_pred = align_shape(y_true, y_pred)
# Checking binarity
check_binary(y_true, y_pred)
return (y_true.astype(bool) == y_pred.astype(bool)).mean()
batches_per_epoch = len(train_tf_dataset)
val_batches_per_epoch = len(val_tf_dataset)
# tf.debugging.set_log_device_placement(True)
for epoch in range(1, pretrain_epochs + 1):
new_learning_rate = 0.001 - (0.001 - 0.0005) * (epoch -
1) / (pretrain_epochs - 1)
optimizer.lr.assign(new_learning_rate)
iterator = iter(train_dist_dataset)
print('epoch', epoch)
losses = []
for batch_n in tqdm(range(1, batches_per_epoch + 1)):
batch = next(iterator)
if check_shapes(batch):
loss, greedy_seqs = distributed_step(batch, 'train')
# loss = distributed_train_step(batch)
losses.append(loss)
if batch_n % 200 == 0:
# if True:
with tf.device('CPU'):
train_sums = list(
tf.concat(greedy_seqs.values, axis=0).numpy())
train_inds = list(
tf.concat(batch[-1].values, axis=0).numpy().squeeze())
articles = [article[x] for x in train_inds]
gt_summaries = [summary[x] for x in train_inds]
examples_oovs = [oovs[x] for x in train_inds]