def __init__(self, n_inputs, layersizes=None, max_iter=100,
learning_rate=0.05,
activation="ReLU", output="softmax", loss="crossentropy",
l2_penalty=0.0, l1_penalty=0.0,
dropout=0.0, input_dropout=0.0, batch_size=64,
momentum=0.0, fraction_validation_set=0.15,
convergence_iter_tol=30, early_stopping=True,
shuffle_data=True,
learning_rate_schedule="constant", learning_rate_decay=None,
snapshot_interval=None,
verbose=False, random_state=None, dtype=np.float32,
activationparams=None,
layerclass="default", logger=None,
output_weights=None):
self.batch_size = batch_size
self.max_iter = max_iter
# we need a seed that is random even in multiple threads
# started very close to each other in time
if random_state is None:
random_state = op.rand_int()
self.n_inputs = n_inputs
self.dropout = dropout
self.verbose = verbose
self.random_state = random_state
self.l2_penalty = l2_penalty
self.l1_penalty = l1_penalty
self.shuffle_data = shuffle_data
self.learning_rate = learning_rate # used for adaptive schedule
self.momentum = momentum
self.learning_rate_schedule = learning_rate_schedule
self.learning_rate_decay = learning_rate_decay
self.layersizes = layersizes
self.input_dropout = input_dropout
self.activation = activation.lower()
self.loss = loss.lower()
self.output = output.lower()
self.dtype = dtype
self.layerclass = layerclass
self.logger = logger
self.output_weights = output_weights
self.convergence_iter_tol = convergence_iter_tol
self.fraction_validation_set = fraction_validation_set
self.early_stopping = early_stopping
self.activationparams = activationparams
self.snapshot_interval = snapshot_interval
if learning_rate_schedule not in ('constant', 'adaptive', 'simple', 'invscale', 'linear', 'power'):
raise ValueError("Unknown learning rate schedule.")
if self.layersizes is not None:
self.setup_layers(self.activationparams)
self.ignore_last_minibatch_if_smaller=False
op.set_seed(self.random_state)
self.reset()
self._epoch_callbacks = []
self._minibatch_callbacks = []
# use self.__class__ instead of NeuralNet in case
# subclasses implement track_progress themselves
self._epoch_callbacks.append(self.__class__.track_progress)
def reset(self, random_state=None):
self.statistics = pd.DataFrame(dtype=np.float64,
columns=('train_loss', 'val_loss', 'val_score', 'time'))
self.statistics.index.name = "epoch"
self.current_epoch = 0 # number of iterations
self.update_count = 0
self._best_params = None
self._best_score_va = np.finfo(np.float32).min
self._no_improvement_since = 0
if random_state is not None:
op.set_seed(self.random_state)
if len(self.layers) > 0:
self.layers[0].setup(self.n_inputs, batch_size=self.batch_size)
for l1, l2 in zip(self.layers[:-1], self.layers[1:]):
l2.setup(l1.size, batch_size=self.batch_size)
def reset(self, random_state=None):
self.statistics = pd.DataFrame(dtype=np.float64,
columns=('train_loss', 'val_loss',
'val_score', 'time'))
self.statistics.index.name = "epoch"
self.current_epoch = 0 # number of iterations
self.update_count = 0
self._best_params = None
self._best_score_va = np.finfo(np.float32).min
self._no_improvement_since = 0
if random_state is not None:
op.set_seed(self.random_state)
if len(self.layers) > 0:
self.layers[0].setup(self.n_inputs, batch_size=self.batch_size)
for l1, l2 in zip(self.layers[:-1], self.layers[1:]):
l2.setup(l1.size, batch_size=self.batch_size)
def __init__(self,
n_inputs,
layersizes=None,
max_iter=100,
learning_rate=0.05,
activation="ReLU",
output="softmax",
loss="crossentropy",
l2_penalty=0.0,
l1_penalty=0.0,
dropout=0.0,
input_dropout=0.0,
batch_size=64,
momentum=0.0,
fraction_validation_set=0.15,
convergence_iter_tol=30,
early_stopping=True,
shuffle_data=True,
learning_rate_schedule="constant",
learning_rate_decay=None,
snapshot_interval=None,
verbose=False,
random_state=None,
dtype=np.float32,
activationparams=None,
layerclass="default",
logger=None,
output_weights=None):
self.batch_size = batch_size
self.max_iter = max_iter
# we need a seed that is random even in multiple threads
# started very close to each other in time
if random_state is None:
random_state = op.rand_int()
self.n_inputs = n_inputs
self.dropout = dropout
self.verbose = verbose
self.random_state = random_state
self.l2_penalty = l2_penalty
self.l1_penalty = l1_penalty
self.shuffle_data = shuffle_data
self.learning_rate = learning_rate # used for adaptive schedule
self.momentum = momentum
self.learning_rate_schedule = learning_rate_schedule
self.learning_rate_decay = learning_rate_decay
self.layersizes = layersizes
self.input_dropout = input_dropout
self.activation = activation.lower()
self.loss = loss.lower()
self.output = output.lower()
self.dtype = dtype
self.layerclass = layerclass
self.logger = logger
self.output_weights = output_weights
self.convergence_iter_tol = convergence_iter_tol
self.fraction_validation_set = fraction_validation_set
self.early_stopping = early_stopping
self.activationparams = activationparams
self.snapshot_interval = snapshot_interval
if learning_rate_schedule not in ('constant', 'adaptive', 'simple',
'invscale', 'linear', 'power'):
raise ValueError("Unknown learning rate schedule.")
self.layers = []
if self.layersizes is not None:
self.setup_layers(self.activationparams)
op.set_seed(self.random_state)
self.reset()
self._epoch_callbacks = []
self._minibatch_callbacks = []
self._epoch_callbacks.append(NeuralNet.track_progress)