def fit(self,
X,
y,
batch_size=32,
epochs=1,
verbose=1,
validation_split=0.0,
validation_data=None,
shuffle=True):
"""Trains the model for a fixed number of epochs.
Arguments:
X: tensor or array-like
Input data.
y: tensor or array-like
Target data.
batch_size: integer, Default: 32
Number of sample in gradient update.
epochs: integer, Default: 1
Number of epochs to train the model.
verbose: integer, Default: 1
Verbosity mode. 0 = silent, 1 = one line per epoch.
validation_split: float, Default: 0
Fraction of the training data to be used as validation data.
The validation data is selected from the last samples in the
`x` and `y` data, before shuffling.
validation_data: tuple `(x_val, y_val)`, Default: None
Data on which to evaluate the loss and any model metrics at
the end of each epoch.
`validation_data` will override `validation_split`.
shuffle: boolean, Default: True
Whether to shuffle the data before each epoch.
"""
if validation_data is None:
split = train_test_split(X, y, validation_split, shuffle=False)
X, X_val, y, y_val = split
else:
X_val, y_val = validation_data
progbar = Progbar(epochs, X.shape[0], batch_size)
for epoch in range(epochs):
X, y = shuffle_data(X, y)
losses = []
metrics = []
for output, target in batch_iterator(X, y, batch_size):
loss, metric = self.train_on_batch(output, target)
losses.append(loss)
metrics.append(metric)
if verbose >= 1:
progbar.update(epoch, loss, metric)
loss = M.mean(losses)
metric = M.mean(metrics)
val_loss, val_metric = self.evaluate(X_val, y_val, batch_size)
self.losses['training'].append(loss)
self.metrics['training'].append(metric)
self.losses['validation'].append(val_loss)
self.metrics['validation'].append(val_metric)
if verbose >= 1:
progbar.update(epoch, loss, metric, val_loss, val_metric)
def evaluate(self, X=None, y=None, batch_size=32, verbose=0):
"""Returns the loss value & metrics values for the model in test mode.
Arguments:
X: array-like, Default: None
Input data.
y: array-like, Default: None
Target data.
batch_size: integer, Default: 32
Number of sample in gradient update.
verbose: integer, Default: 0
Verbosity mode. 0 = silent, 1 = one line per batch.
Returns:
A scalar test loss or list of scalars with metrics.
"""
losses = []
metrics = []
for output, target in batch_iterator(X, y, batch_size):
loss, metric, output = self.test_on_batch(output, target)
losses.append(loss)
metrics.append(metric)
if verbose >= 1:
print(f"-> {target} - raw{M.round(output, decimals=3)}")
return M.mean(losses), M.mean(metrics)
def call(self, y_true, y_pred):
target = y_true
output = y_pred
if self.from_logits:
output = 1 / (1 + M.exp(-y_pred))
output = M.clip(output, M.epsilon(), 1.0 - M.epsilon())
output = -target * M.log(output) - (1.0 - target) * M.log(1.0 - output)
return M.mean(output, axis=-1)
def fit(self, X):
"""Compute the mean and std to be used for later scaling.
Arguments:
X: array-like
The data used to compute the mean and standard deviation used
for later scaling along the features axis.
"""
self.mean = M.array([])
self.std = M.array([])
for i in range(X.shape[1]):
self.mean = M.append(self.mean, M.mean(X[:, i]))
self.std = M.append(self.std, M.std(X[:, i]))
return self
def test_on_batch(self, X, y):
"""Test the model on a single batch of samples.
Arguments:
X: array-like
Input data.
y: array-like
Target data.
Returns:
A scalar test loss or list of scalars with metrics.
"""
output = X
for layer in self.layers:
trainable = layer.trainable
layer.trainable = False
output = layer(output)
layer.trainable = trainable
loss = M.mean(self.loss(y, output))
metric = self.compile_metrics(y, output)
return loss, metric, output
def train_on_batch(self, X, y):
"""Runs a single gradient update on a single batch of data.
Arguments:
X: array-like
Input data.
y: array-like
Target data.
Returns:
A scalar test loss or list of scalars with metrics.
"""
output = X
for layer in self.layers:
output = layer(output)
loss = M.mean(self.loss(y, output))
metric = self.compile_metrics(y, output)
grads = self.loss.gradient(y, output)
for layer in reversed(self.layers):
grads = layer.backward(grads)
return loss, metric
def call(self, y_true, y_pred):
return M.mean(M.square(y_pred - y_true), axis=-1)
def binary_accuracy(self, y_true, y_pred, threshold=0.5):
if threshold != 0.5:
y_pred = (y_pred > threshold)
return M.mean(M.equal(y_true, M.round(y_pred)), axis=-1)
