class Classifier(Predictor):
model_name = 'classifier'
def __init__(self, mark=None, net_type=Feedforward):
Predictor.__init__(self, mark, net_type)
# Private attributes
self._probabilities = TensorSlot(self, 'Probability')
# TODO: to be deprecated
# self._evaluation_group = Group(self, self._metric, self._probabilities,
# name='evaluation group')
@with_graph
def build(self,
optimizer=None,
loss='cross_entropy',
metric='accuracy',
batch_metric=None,
eval_metric=None,
**kwargs):
Predictor.build(self,
optimizer=optimizer,
loss=loss,
metric=metric,
batch_metric=batch_metric,
eval_metric=eval_metric,
**kwargs)
def _build(self, optimizer=None, metric=None, **kwargs):
# TODO: ... do some compromise
hub.block_validation = True
# If last layer is not softmax layer, add it to model TODO
# if not (isinstance(self.last_function, Activation)
# and self.last_function.abbreviation == 'softmax'):
# self.add(Activation('softmax'))
# Call parent's method to build using the default loss function
# -- cross entropy
Predictor._build(self, optimizer=optimizer, metric=metric, **kwargs)
assert self.outputs.activated
# Plug tensor into probabilities slot
self._probabilities.plug(self.outputs.tensor)
@with_graph
def classify(self,
data,
batch_size=None,
extractor=None,
return_probs=False,
verbose=False):
probs = self.evaluate(self._probabilities.tensor,
data,
batch_size,
extractor,
verbose=verbose)
if return_probs: return probs
# TODO: make clear data shape here and add comments
if self.input_type is InputTypes.RNN_BATCH:
preds = [np.argmax(p, axis=-1) for p in probs]
else:
preds = np.argmax(probs, axis=-1)
return preds
@with_graph
def evaluate_model(self, data, batch_size=None, extractor=None, **kwargs):
"""This method is a mess."""
if hub.take_down_confusion_matrix:
# TODO: (william) please refactor this method
cm = self.evaluate_pro(data,
batch_size,
verbose=kwargs.get('verbose', False),
show_class_detail=True,
show_confusion_matrix=True)
# Take down confusion matrix
from tframe import context
agent = context.trainer.model.agent
agent.take_notes('Confusion Matrix on {}:'.format(data.name),
False)
agent.take_notes('\n' + cm.matrix_table().content)
agent.take_notes('Evaluation Result on {}:'.format(data.name),
False)
agent.take_notes('\n' + cm.make_table().content)
return cm.accuracy
# If not necessary, use Predictor's evaluate_model method
metric_is_accuracy = self.eval_metric.name.lower() == 'accuracy'
if not metric_is_accuracy:
result = super().evaluate_model(data, batch_size, **kwargs)
if metric_is_accuracy: result *= 100
return result
console.show_status('Evaluating classifier on {} ...'.format(
data.name))
acc_slot = self.metrics_manager.get_slot_by_name('accuracy')
assert isinstance(acc_slot, MetricSlot)
acc_foreach = acc_slot.quantity_definition.quantities
results = self.evaluate(acc_foreach,
data,
batch_size,
extractor,
verbose=hub.val_progress_bar)
if self.input_type is InputTypes.RNN_BATCH:
results = np.concatenate([y.flatten() for y in results])
accuracy = np.mean(results) * 100
# Show accuracy
console.supplement('Accuracy on {} is {:.3f}%'.format(
data.name, accuracy))
# Return accuracy
return accuracy
@with_graph
def evaluate_pro(self, data_set, batch_size=None, verbose=False, **kwargs):
"""Evaluate model and give results calculated based on TP, TN, FP, and FN.
'extractor' is not considered cuz the developer forgot what is this used
for.
:param data_set: an instance of dataset contains at least features
:param batch_size: set this value if your (G)RAM is not big enough to
handle the whole dataset
:param verbose: whether to show status or progress bar stuff
:param kwargs: other options which will be recognized by PyCharm
"""
# Get options
show_confusion_matrix = kwargs.get('show_confusion_matrix', False)
show_class_detail = kwargs.get('show_class_detail', False)
export_false = kwargs.get('export_false', False)
top_k = kwargs.get('export_top_k', 3)
# Check data_set before get model prediction
assert self.input_type is InputTypes.BATCH
assert isinstance(data_set, DataSet)
assert data_set.features is not None and data_set.targets is not None
# -------------------------------------------------------------------------
# Calculate predicted classes and corresponding probabilities
# -------------------------------------------------------------------------
probs = self.classify(data_set,
batch_size,
return_probs=True,
verbose=verbose)
# This provides necessary information for image viewer presentation
# i.e., the sorted probabilities for each class
probs_sorted = np.fliplr(np.sort(probs, axis=-1))
class_sorted = np.fliplr(np.argsort(probs, axis=-1))
preds = class_sorted[:, 0]
truths = np.ravel(data_set.dense_labels)
# Produce confusion matrix
cm = ConfusionMatrix(num_classes=data_set.num_classes,
class_names=data_set.properties.get(
pedia.classes, None))
cm.fill(preds, truths)
# Print evaluation results
if show_confusion_matrix:
console.show_info('Confusion Matrix:')
console.write_line(cm.matrix_table(kwargs.get('cell_width', None)))
console.show_info('Evaluation Result:')
console.write_line(
cm.make_table(decimal=4, class_details=show_class_detail))
# Visualize false set if specified
if export_false:
indices = np.argwhere(preds != truths).flatten()
false_set = data_set[indices]
false_set.properties[pedia.predictions] = preds[indices]
false_set.properties[pedia.top_k_label] = class_sorted[
indices, :top_k]
false_set.properties[pedia.top_k_prob] = probs_sorted[
indices, :top_k]
return cm, false_set
else:
return cm
def _build(self,
loss='cross_entropy',
optimizer=None,
metric=None,
metric_is_like_loss=True,
metric_name='Metric'):
Feedforward._build(self)
# Check shapes of branch outputs
output_shape = self._check_branch_outputs()
# Initiate targets placeholder
self._plug_target_in(output_shape)
# Define output tensors
for i, output in enumerate(self.branch_outputs):
if i == 0 or not self.strict_residual:
output_tensor = output
else:
output_tensor = output + self._boutputs[i - 1].tensor
slot = TensorSlot(self, name='output_{}'.format(i + 1))
slot.plug(output_tensor)
self._boutputs.append(slot)
# Define loss tensors
loss_function = losses.get(loss)
with tf.name_scope('Loss'):
for i, output in enumerate(self._boutputs):
assert isinstance(output, TensorSlot)
loss_tensor = loss_function(self._targets.tensor,
output.tensor)
slot = TensorSlot(self, name='loss_{}'.format(i + 1))
slot.plug(loss_tensor)
self._losses.append(slot)
# Add summary
if hub.summary:
name = 'loss_sum_{}'.format(i + 1)
sum_slot = SummarySlot(self, name)
sum_slot.plug(tf.summary.scalar(name, loss_tensor))
self._train_step_summaries.append(sum_slot)
# Define metric tensors
metric_function = metrics.get(metric)
if metric_function is not None:
with tf.name_scope('Metric'):
for i, output in enumerate(self._boutputs):
assert isinstance(output, TensorSlot)
metric_tensor = metric_function(self._targets.tensor,
output.tensor)
slot = Metric(self, name='metric_{}'.format(i + 1))
slot.plug(metric_tensor,
as_loss=metric_is_like_loss,
symbol='{}{}'.format(metric_name, i + 1))
self._metrics.append(slot)
# Add summary
if hub.summary:
name = 'metric_sum_{}'.format(i + 1)
sum_slot = SummarySlot(self, name)
sum_slot.plug(tf.summary.scalar(name, metric_tensor))
self._validation_summaries.append(sum_slot)
# Define train step
self._define_train_step(optimizer)
# Define groups
# TODO when train a single branch with summary on, error may occur
# .. due to that the higher branch summary can not get its value
act_summaries = []
if hub.monitor_preact:
slot = SummarySlot(self, 'act_summary')
slot.plug(
tf.summary.merge(tf.get_collection(
pedia.train_step_summaries)))
act_summaries.append(slot)
self._update_group = Group(self, *self._losses, *self._train_steps,
*self._train_step_summaries, *act_summaries)
self._validate_group = Group(self, *self._metrics,
*self._validation_summaries)
class Predictor(Feedforward, Recurrent):
"""A feedforward or a recurrent predictor"""
model_name = 'Predictor'
def __init__(self, mark=None, net_type=Feedforward):
"""
Construct a Predictor
:param mark: model mark
:param net_type: \in {Feedforward, Recurrent}
"""
if not net_type in (Feedforward, Recurrent):
raise TypeError('!! Unknown net type')
self.master = net_type
# Attributes
self._targets = TensorSlot(self, 'targets')
self._val_targets = TensorSlot(self, 'val_targets')
# Call parent's constructor
net_type.__init__(self, mark)
# region : Properties
@property
def affix(self):
if self.master is Feedforward: return 'forward'
assert self.master is Recurrent
return 'recurrent'
@property
def description(self):
return '{}: {}'.format(self.master.__name__, self.structure_string())
@property
def input_type(self):
if self.master is Feedforward: return InputTypes.BATCH
else: return InputTypes.RNN_BATCH
# endregion : Properties
# region : Build
@with_graph
def build_as_regressor(self,
optimizer=None,
loss='euclid',
metric='rms_ratio',
metric_name='Err %'):
self.build(optimizer=optimizer,
loss=loss,
metric=metric,
metric_name=metric_name)
@with_graph
def build(self,
optimizer=None,
loss='euclid',
metric=None,
batch_metric=None,
eval_metric=None,
**kwargs):
context.metric_name = 'unknown' # TODO: to be deprecated
Model.build(self,
optimizer=optimizer,
loss=loss,
metric=metric,
batch_metric=batch_metric,
eval_metric=eval_metric,
**kwargs)
def _build(self, optimizer=None, loss='euclid', metric=None, **kwargs):
# For some RNN predictors, their last step is counted as the only output
# e.g. RNNs for sequence classification tasks
last_only = False
if 'last_only' in kwargs.keys():
last_only = kwargs.pop('last_only')
if hub.use_gather_indices:
# Initiate gather_indices placeholder
assert context.gather_indices is None
context.gather_indices = tf.placeholder(
tf.int32, [None, 2], 'gather_indices')
tf.add_to_collection(pedia.default_feed_dict,
context.gather_indices)
# Get loss quantity before building
self.loss_quantity = losses.get(loss, last_only)
# This is for calculating loss inside a while-loop
context.loss_function = self.loss_quantity.function
# Call parent's build method to link network
# Usually output tensor has been plugged into Model._outputs slot
self.master._build(self)
assert self.outputs.activated
# Initiate targets and add it to collection
self._plug_target_in(self.outputs.shape_list)
# Define loss. Some tensorflow apis only support calculating logits
with tf.name_scope('Loss'):
loss_tensor = self.loss_quantity(self._targets.tensor,
self.outputs.tensor)
# TODO: with or without regularization loss?
if hub.summary:
tf.add_to_collection(
pedia.train_step_summaries,
tf.summary.scalar('loss_sum', loss_tensor))
# Try to add extra loss which is calculated by the corresponding net
# .. regularization loss is included
if self.extra_loss is not None:
loss_tensor = tf.add(loss_tensor, self.extra_loss)
# Plug in
self.loss.plug(loss_tensor, quantity_def=self.loss_quantity)
# <monitor_grad_step_02: register loss and plug grad_ops in>
if hub.monitor_weight_grads:
context.monitor.register_loss(loss_tensor)
self.grads_slot.plug(context.monitor.grad_ops_list)
self._update_group.add(self.grads_slot)
# Monitor general tensors (currently only activation is included)
if hub.export_activations and context.monitor.tensor_fetches:
self.general_tensor_slot.plug(context.monitor.tensor_fetches)
self._update_group.add(self.general_tensor_slot)
# Initialize metric
if metric is not None:
checker.check_type_v2(metric, (str, Quantity))
# Create placeholder for val_targets if necessary
# Common targets will be plugged into val_target slot by default
self._plug_val_target_in(kwargs.get('val_targets', None))
with tf.name_scope('Metric'):
self._metrics_manager.initialize(metric, last_only,
self._val_targets.tensor,
self._outputs.tensor,
**kwargs)
# Merge summaries
self._merge_summaries()
# Define train step
self._define_train_step(optimizer)
def _plug_target_in(self, shape):
dtype = hub.dtype
if hub.target_dim != 0: shape[-1] = hub.target_dim
# If target is sparse label
if hub.target_dtype is not None: dtype = hub.target_dtype
# if hub.target_dim == 1: dtype = tf.int32 # TODO: X
# Handle recurrent situation
if self._targets.tensor is not None:
# targets placeholder has been plugged in Recurrent._build_while_free
# method
assert self.master == Recurrent
return
target_tensor = tf.placeholder(dtype, shape, name='targets')
self._targets.plug(target_tensor, collection=pedia.default_feed_dict)
def _plug_val_target_in(self, val_targets):
if val_targets is None:
self._val_targets = self._targets
else:
assert isinstance(val_targets, str)
val_target_tensor = tf.placeholder(hub.dtype,
self.outputs.shape_list,
name=val_targets)
self._val_targets.plug(val_target_tensor,
collection=pedia.default_feed_dict)
# endregion : Build
# region : Train
def update_model(self, data_batch, **kwargs):
if self.master is Feedforward:
return Feedforward.update_model(self, data_batch, **kwargs)
# Update recurrent model
feed_dict = self._get_default_feed_dict(data_batch, is_training=True)
results = self._update_group.run(feed_dict)
self.set_buffers(results.pop(self._state_slot), is_training=True)
# TODO: BETA
assert not hub.use_rtrl
if hub.use_rtrl:
self._gradient_buffer_array = results.pop(self._grad_buffer_slot)
if hub.test_grad:
delta = results.pop(self.grad_delta_slot)
_ = None
return results
# endregion : Train
# region : Public Methods
@with_graph
def predict(self, data, batch_size=None, extractor=None, **kwargs):
return self.evaluate(self._outputs.tensor, data, batch_size, extractor,
**kwargs)
@with_graph
def evaluate_model(self, data, batch_size=None, dynamic=False, **kwargs):
"""The word `evaluate` in this method name is different from that in
`self.evaluate` method. Here only eval_metric will be evaluated and
the result will be printed on terminal."""
# Check metric
if not self.eval_metric.activated:
raise AssertionError('!! Metric not defined')
# Do dynamic evaluation if necessary
if dynamic:
from tframe.trainers.eval_tools.dynamic_eval import DynamicEvaluator as de
de.dynamic_evaluate(self, data, kwargs.get('val_set', None),
kwargs.get('delay', None))
return
# If hub.val_progress_bar is True, this message will be showed in
# model.evaluate method
if not hub.val_progress_bar:
console.show_status('Evaluating on {} ...'.format(data.name))
# use val_progress_bar option here temporarily
result = self.validate_model(
data, batch_size, allow_sum=False,
verbose=hub.val_progress_bar)[self.eval_metric]
console.supplement('{} = {}'.format(
self.eval_metric.symbol, hub.decimal_str(result,
hub.val_decimals)))
return result
# endregion : Public Methods
# region : Private Methods
def _evaluate_batch(self, fetch_list, data_batch, **kwargs):
return self.master._evaluate_batch(self, fetch_list, data_batch,
**kwargs)
def _get_default_feed_dict(self, batch, is_training):
return self.master._get_default_feed_dict(self, batch, is_training)
class Classifier(Predictor):
model_name = 'classifier'
def __init__(self, mark=None, net_type=Feedforward):
Predictor.__init__(self, mark, net_type)
# Private attributes
self._probabilities = TensorSlot(self, 'Probability')
# TODO: to be deprecated
# self._evaluation_group = Group(self, self._metric, self._probabilities,
# name='evaluation group')
@with_graph
def build(self, optimizer=None, loss='cross_entropy', metric='accuracy',
batch_metric=None, eval_metric=None, **kwargs):
Predictor.build(
self, optimizer=optimizer, loss=loss, metric=metric,
batch_metric=batch_metric, eval_metric=eval_metric, **kwargs)
def _build(self, optimizer=None, metric=None, **kwargs):
# TODO: ... do some compromise
hub.block_validation = True
# If last layer is not softmax layer, add it to model TODO
# if not (isinstance(self.last_function, Activation)
# and self.last_function.abbreviation == 'softmax'):
# self.add(Activation('softmax'))
# Call parent's method to build using the default loss function
# -- cross entropy
Predictor._build(self, optimizer=optimizer, metric=metric, **kwargs)
assert self.outputs.activated
# Plug tensor into probabilities slot
self._probabilities.plug(self.outputs.tensor)
@with_graph
def evaluate_model(self, data, batch_size=None, extractor=None,
export_false=False, **kwargs):
# If not necessary, use Predictor's evaluate_model method
metric_is_accuracy = self.eval_metric.name.lower() == 'accuracy'
if not export_false or not metric_is_accuracy:
result = super().evaluate_model(data, batch_size, **kwargs)
if metric_is_accuracy: result *= 100
return result
console.show_status('Evaluating classifier on {} ...'.format(data.name))
acc_slot = self.metrics_manager.get_slot_by_name('accuracy')
assert isinstance(acc_slot, MetricSlot)
acc_foreach = acc_slot.quantity_definition.quantities
results = self.evaluate(acc_foreach, data, batch_size, extractor,
verbose=hub.val_progress_bar)
if self.input_type is InputTypes.RNN_BATCH:
results = np.concatenate([y.flatten() for y in results])
accuracy = np.mean(results) * 100
# Show accuracy
console.supplement('Accuracy on {} is {:.3f}%'.format(data.name, accuracy))
# export_false option is valid for images only
if export_false and accuracy < 100.0:
assert self.input_type is InputTypes.BATCH
assert isinstance(data, DataSet)
assert data.features is not None and data.targets is not None
top_k = hub.export_top_k if hub.export_top_k > 0 else 3
probs = self.classify(data, batch_size, extractor, return_probs=True)
probs_sorted = np.fliplr(np.sort(probs, axis=-1))
class_sorted = np.fliplr(np.argsort(probs, axis=-1))
preds = class_sorted[:, 0]
false_indices = np.argwhere(results == 0).flatten()
false_preds = preds[false_indices]
probs_sorted = probs_sorted[false_indices, :top_k]
class_sorted = class_sorted[false_indices, :top_k]
false_set = data[false_indices]
false_set.properties[pedia.predictions] = false_preds
false_set.properties[pedia.top_k_label] = class_sorted
false_set.properties[pedia.top_k_prob] = probs_sorted
from tframe.data.images.image_viewer import ImageViewer
vr = ImageViewer(false_set)
vr.show()
# Return accuracy
return accuracy
@with_graph
def classify(self, data, batch_size=None, extractor=None,
return_probs=False, verbose=False):
probs = self.evaluate(
self._probabilities.tensor, data, batch_size, extractor, verbose=verbose)
if return_probs: return probs
if self.input_type is InputTypes.RNN_BATCH:
preds = [np.argmax(p, axis=-1) for p in probs]
else: preds = np.argmax(probs, axis=-1)
return preds
class Classifier(Predictor):
def __init__(self, mark=None, net_type=Feedforward):
Predictor.__init__(self, mark, net_type)
# Private attributes
self._probabilities = TensorSlot(self, 'Probability')
self._evaluation_group = Group(self,
self._metric,
self._probabilities,
name='evaluation group')
@with_graph
def build(self, optimizer=None, metric='accuracy', **kwargs):
Predictor.build(self,
optimizer=optimizer,
loss='cross_entropy',
metric=metric,
metric_is_like_loss=False,
metric_name='Accuracy')
def _build(self, optimizer=None, metric=None, **kwargs):
# TODO: ... do some compromise
hub.block_validation = True
# If last layer is not softmax layer, add it to model
if not (isinstance(self.last_function, Activation)
and self.last_function.abbreviation == 'softmax'):
self.add(Activation('softmax'))
# Call parent's method to build using the default loss function
# -- cross entropy
Predictor._build(self, optimize=optimizer, metric=metric, **kwargs)
assert self.outputs.activated
# Plug tensor into probabilities slot
self._probabilities.plug(self.outputs.tensor)
def evaluate_model(self,
data,
batch_size=None,
extractor=None,
export_false=False,
**kwargs):
# Feed data set into model and get results
false_sample_list = []
false_label_list = []
true_label_list = []
num_samples = 0
console.show_status('Evaluating classifier ...')
for batch in self.get_data_batches(data, batch_size):
assert isinstance(batch, DataSet) and batch.targets is not None
# Get predictions
preds = self._classify_batch(batch, extractor)
# Get true labels in dense format
if batch.targets.shape[-1] > 1:
targets = batch.targets.reshape(-1, batch.targets.shape[-1])
else:
targets = batch.targets
num_samples += len(targets)
true_labels = misc.convert_to_dense_labels(targets)
if len(true_labels) < len(preds):
assert len(true_labels) == 1
true_labels = np.concatenate((true_labels, ) * len(preds))
# Select false samples
false_indices = np.argwhere(preds != true_labels)
if false_indices.size == 0: continue
features = batch.features
if self.input_type is InputTypes.RNN_BATCH:
features = np.reshape(features, [-1, *features.shape[2:]])
false_indices = np.reshape(false_indices, false_indices.size)
false_sample_list.append(features[false_indices])
false_label_list.append(preds[false_indices])
true_label_list.append(true_labels[false_indices])
# Concatenate
if len(false_sample_list) > 0:
false_sample_list = np.concatenate(false_sample_list)
false_label_list = np.concatenate(false_label_list)
true_label_list = np.concatenate(true_label_list)
# Show accuracy
accuracy = (num_samples - len(false_sample_list)) / num_samples * 100
console.supplement('Accuracy on {} is {:.2f}%'.format(
data.name, accuracy))
# Try to export false samples
if export_false and accuracy < 100:
false_set = DataSet(features=false_sample_list,
targets=true_label_list)
if hasattr(data, 'properties'):
false_set.properties = data.properties
false_set.data_dict[pedia.predictions] = false_label_list
from tframe.data.images.image_viewer import ImageViewer
vr = ImageViewer(false_set)
vr.show()
def classify(self, data, batch_size=None, extractor=None):
predictions = []
for batch in self.get_data_batches(data, batch_size):
preds = self._classify_batch(batch, extractor)
if isinstance(preds, int): preds = [preds]
predictions.append(preds)
return np.concatenate(predictions)
def _classify_batch(self, batch, extractor):
assert isinstance(batch, DataSet) and batch.features is not None
batch = self._sanity_check_before_use(batch)
feed_dict = self._get_default_feed_dict(batch, is_training=False)
probs = self._probabilities.run(feed_dict)
if self.input_type is InputTypes.RNN_BATCH:
assert len(probs.shape) == 3
probs = np.reshape(probs, (-1, probs.shape[2]))
if extractor is None: preds = misc.convert_to_dense_labels(probs)
else: preds = extractor(probs)
return preds
class Predictor(Feedforward, Recurrent):
"""A feedforward or a recurrent predictor"""
model_name = 'Predictor'
def __init__(self, mark=None, net_type=Feedforward):
"""
Construct a Predictor
:param mark: model mark
:param net_type: \in {Feedforward, Recurrent}
"""
if not net_type in (Feedforward, Recurrent):
raise TypeError('!! Unknown net type')
self.master = net_type
# Call parent's constructor
net_type.__init__(self, mark)
# Attributes
self._targets = TensorSlot(self, 'targets')
# region : Properties
@property
def description(self):
return '{}: {}'.format(self.master.__name__, self.structure_string())
@property
def input_type(self):
if self.master is Feedforward: return InputTypes.BATCH
else: return InputTypes.RNN_BATCH
# endregion : Properties
# region : Build
@with_graph
def build_as_regressor(self,
optimizer=None,
loss='euclid',
metric='rms_ratio',
metric_is_like_loss=True,
metric_name='Err %'):
self.build(optimizer=optimizer,
loss=loss,
metric=metric,
metric_name=metric_name,
metric_is_like_loss=metric_is_like_loss)
@with_graph
def build(self,
optimizer=None,
loss='euclid',
metric=None,
metric_is_like_loss=True,
metric_name='Metric',
**kwargs):
Model.build(self,
optimizer=optimizer,
loss=loss,
metric=metric,
metric_name=metric_name,
metric_is_like_loss=metric_is_like_loss)
def _build(self,
optimizer=None,
loss='euclid',
metric=None,
metric_is_like_loss=True,
metric_name='Metric',
**kwargs):
# Call parent's build method
# Usually output tensor has been plugged into Model._outputs slot
self.master._build(self)
assert self.outputs.activated
# Initiate targets and add it to collection
self._plug_target_in(self.outputs.shape_list)
# Define loss
loss_function = losses.get(loss)
with tf.name_scope('Loss'):
if loss == 'cross_entropy':
output_tensor = self.logits_tensor
assert output_tensor is not None
else:
output_tensor = self.outputs.tensor
loss_tensor = loss_function(self._targets.tensor, output_tensor)
# TODO: with or without regularization loss?
if hub.summary:
tf.add_to_collection(
pedia.train_step_summaries,
tf.summary.scalar('loss_sum', loss_tensor))
# Try to add regularization loss
reg_loss = self.regularization_loss
if reg_loss is not None: loss_tensor += reg_loss
# Plug in
self.loss.plug(loss_tensor)
# Define metric
if metric is not None:
metric_function = metrics.get(metric)
with tf.name_scope('Metric'):
metric_tensor = metric_function(self._targets.tensor,
self._outputs.tensor)
self._metric.plug(metric_tensor,
as_loss=metric_is_like_loss,
symbol=metric_name)
if hub.summary:
tf.add_to_collection(
pedia.validation_summaries,
tf.summary.scalar('metric_sum', self._metric.tensor))
# Merge summaries
self._merge_summaries()
# Define train step
self._define_train_step(optimizer)
def _plug_target_in(self, shape):
target_tensor = tf.placeholder(hub.dtype, shape, name='targets')
self._targets.plug(target_tensor, collection=pedia.default_feed_dict)
# endregion : Build
# region : Train
# TODO
# def begin_round(self, **kwargs):
# if self.master is Recurrent:
# th = kwargs.get('th')
# assert isinstance(th, TrainerHub)
# self.reset_state(th.batch_size)
def update_model(self, data_batch, **kwargs):
if self.master is Feedforward:
return Feedforward.update_model(self, data_batch, **kwargs)
# Update recurrent model
feed_dict = self._get_default_feed_dict(data_batch, is_training=True)
results = self._update_group.run(feed_dict)
self._state_array = results.pop(self._state)
return results
# endregion : Train
# region : Public Methods
def predict(self, data, batch_size=None, extractor=None, **kwargs):
outputs = []
for batch in self.get_data_batches(data, batch_size):
batch = self._sanity_check_before_use(batch)
feed_dict = self._get_default_feed_dict(batch, is_training=False)
output = self._outputs.run(feed_dict)
if extractor is not None: output = extractor(output)
outputs.append(output)
return np.concatenate(outputs)
def evaluate_model(self, data, batch_size=None, **kwargs):
# Check metric
if not self.metric.activated:
raise AssertionError('!! Metric not defined')
# Show status
console.show_status('Evaluating {} ...'.format(data.name))
result = self.validate_model(data, batch_size,
allow_sum=False)[self.metric]
console.supplement('{} = {:.3f}'.format(self.metric.symbol, result))
# endregion : Public Methods
# region : Private Methods
def _get_default_feed_dict(self, batch, is_training):
feed_dict = Feedforward._get_default_feed_dict(self, batch,
is_training)
if self.master is Recurrent:
assert isinstance(batch, DataSet)
# If a new sequence begin while training, reset state
if is_training:
if batch.should_reset_state:
if hub.notify_when_reset: console.write_line('- ' * 40)
self.reset_state(batch.size)
if batch.should_partially_reset_state:
if hub.notify_when_reset:
if batch.reset_values is not None:
info = [(i, v) for i, v in zip(
batch.reset_batch_indices, batch.reset_values)]
else:
info = batch.reset_batch_indices
console.write_line('{}'.format(info))
self.reset_part_state(batch.reset_batch_indices,
batch.reset_values)
batch_size = None if is_training else batch.size
# If is not training, always set a zero state to model
feed_dict.update(self._get_state_dict(batch_size=batch_size))
return feed_dict