def create_forward(workflow, normalizer, labels_mapping, loader_config):
# Disable plotters:
workflow.plotters_are_enabled = False
# Link downloader
workflow.start_point.unlink_after()
workflow.downloader = Downloader(
workflow,
url="https://s3-eu-west-1.amazonaws.com/veles.forge/MNIST/"
"mnist_test.tar",
directory=root.common.dirs.datasets,
files=["mnist_test"])
workflow.downloader.link_from(workflow.start_point)
workflow.repeater.link_from(workflow.downloader)
# Cnanging MnistLoader for another Loader:
new_loader = workflow.change_unit(
workflow.loader.name,
FileListImageLoader(workflow, **loader_config))
# Link attributes:
# TODO: remove link attributes after adding in change_unit() function
# TODO: data links transmission
workflow.forwards[0].link_attrs(
new_loader, ("input", "minibatch_data"))
workflow.evaluator.link_attrs(
new_loader,
"class_keys",
("batch_size", "minibatch_size"),
("labels", "minibatch_labels"),
("max_samples_per_epoch", "total_samples"),
"class_lengths", ("offset", "minibatch_offset"))
workflow.decision.link_attrs(
new_loader, "minibatch_class", "last_minibatch",
"minibatch_size", "class_lengths", "epoch_ended", "epoch_number")
# Set normalizer from previous Loader to new one:
new_loader._normalizer = normalizer
# Set labels_mapping and class_keys in Evaluator to correct writting the
# results:
workflow.evaluator.labels_mapping = labels_mapping
workflow.evaluator.class_keys = new_loader.class_keys
def __init__(self, workflow, **kwargs):
super(WineWorkflow, self).__init__(workflow, **kwargs)
layers = kwargs["layers"]
self.downloader = Downloader(
self,
url=root.wine.downloader.url,
directory=root.wine.downloader.directory,
files=root.wine.downloader.files,
)
self.downloader.link_from(self.start_point)
self.repeater.link_from(self.downloader)
self.loader = WineLoader(
self,
minibatch_size=root.wine.loader.minibatch_size,
force_numpy=root.wine.loader.force_numpy,
dataset_file=root.wine.loader.dataset_file,
normalization_type=root.wine.loader.normalization_type,
)
self.loader.link_from(self.repeater)
# Add fwds units
del self.forwards[:]
for i, layer in enumerate(layers):
if i < len(layers) - 1:
aa = all2all.All2AllTanh(self, output_sample_shape=(layer,))
else:
aa = all2all.All2AllSoftmax(self, output_sample_shape=(layer,))
self.forwards.append(aa)
if i:
self.forwards[-1].link_from(self.forwards[-2])
self.forwards[-1].link_attrs(self.forwards[-2], ("input", "output"))
else:
self.forwards[-1].link_from(self.loader)
self.forwards[-1].link_attrs(self.loader, ("input", "minibatch_data"))
# Add evaluator for single minibatch
self.evaluator = evaluator.EvaluatorSoftmax(self)
self.evaluator.link_from(self.forwards[-1])
self.evaluator.link_attrs(self.forwards[-1], "output", "max_idx")
self.evaluator.link_attrs(
self.loader,
("batch_size", "minibatch_size"),
("max_samples_per_epoch", "total_samples"),
("labels", "minibatch_labels"),
("offset", "minibatch_offset"),
"class_lengths",
)
# Add decision unit
self.decision = decision.DecisionGD(
self, fail_iterations=root.wine.decision.fail_iterations, max_epochs=root.wine.decision.max_epochs
)
self.decision.link_from(self.evaluator)
self.decision.link_attrs(
self.loader,
"minibatch_class",
"minibatch_size",
"last_minibatch",
"class_lengths",
"epoch_ended",
"epoch_number",
)
self.decision.link_attrs(
self.evaluator,
("minibatch_n_err", "n_err"),
("minibatch_confusion_matrix", "confusion_matrix"),
("minibatch_max_err_y_sum", "max_err_output_sum"),
)
self.snapshotter = NNSnapshotterToFile(
self,
prefix=root.wine.snapshotter.prefix,
directory=root.common.dirs.snapshots,
compression="",
interval=root.wine.snapshotter.interval,
time_interval=root.wine.snapshotter.time_interval,
)
self.snapshotter.link_from(self.decision)
self.snapshotter.link_attrs(self.decision, ("suffix", "snapshot_suffix"))
self.snapshotter.gate_skip = ~self.loader.epoch_ended
self.snapshotter.skip = ~self.decision.improved
self.end_point.link_from(self.snapshotter)
self.end_point.gate_block = ~self.decision.complete
# Add gradient descent units
self.gds[:] = (None,) * len(self.forwards)
self.gds[-1] = (
gd.GDSoftmax(self)
.link_from(self.snapshotter)
.link_attrs(self.evaluator, "err_output")
.link_attrs(self.forwards[-1], "output", "input", "weights", "bias")
.link_attrs(self.loader, ("batch_size", "minibatch_size"))
)
self.gds[-1].gate_skip = self.decision.gd_skip
self.gds[-1].gate_block = self.decision.complete
for i in range(len(self.forwards) - 2, -1, -1):
self.gds[i] = (
gd.GDTanh(self)
.link_from(self.gds[i + 1])
.link_attrs(self.gds[i + 1], ("err_output", "err_input"))
.link_attrs(self.forwards[i], "output", "input", "weights", "bias")
.link_attrs(self.loader, ("batch_size", "minibatch_size"))
)
self.gds[i].gate_skip = self.decision.gd_skip
self.gds[0].need_err_input = False
self.repeater.link_from(self.gds[0])
self.loader.gate_block = self.decision.complete
def __init__(self, workflow, **kwargs):
kwargs["name"] = kwargs.get("name", "Kohonen")
super(KohonenWorkflow, self).__init__(workflow, **kwargs)
self.downloader = Downloader(
self, url=root.kohonen.downloader.url,
directory=root.kohonen.downloader.directory,
files=root.kohonen.downloader.files)
self.downloader.link_from(self.start_point)
self.repeater.link_from(self.downloader)
self.loader = KohonenLoader(
self, name="Kohonen fullbatch loader",
minibatch_size=root.kohonen.loader.minibatch_size,
force_numpy=root.kohonen.loader.force_numpy)
self.loader.link_from(self.repeater)
# Kohonen training layer
self.trainer = kohonen.KohonenTrainer(
self, shape=root.kohonen.forward.shape,
weights_filling=root.kohonen.forward.weights_filling,
weights_stddev=root.kohonen.forward.weights_stddev,
gradient_decay=root.kohonen.train.gradient_decay,
radius_decay=root.kohonen.train.radius_decay)
self.trainer.link_from(self.loader)
self.trainer.link_attrs(self.loader, ("input", "minibatch_data"))
# Loop decision
self.decision = kohonen.KohonenDecision(
self, max_epochs=root.kohonen.decision.epochs)
self.decision.link_from(self.trainer)
self.decision.link_attrs(self.loader,
"minibatch_class",
"last_minibatch",
"class_lengths",
"epoch_ended",
"epoch_number")
self.decision.link_attrs(self.trainer, "weights", "winners")
self.ipython = Shell(self)
self.ipython.link_from(self.decision)
self.ipython.gate_skip = ~self.decision.epoch_ended
self.repeater.link_from(self.ipython)
self.ipython.gate_block = self.decision.complete
self.end_point.link_from(self.decision)
self.end_point.gate_block = ~self.decision.complete
self.loader.gate_block = self.decision.complete
# Error plotter
self.plotters = [nn_plotting_units.KohonenHits(self),
nn_plotting_units.KohonenInputMaps(self),
nn_plotting_units.KohonenNeighborMap(self)]
self.plotters[0].link_attrs(self.trainer, "shape") \
.link_from(self.ipython)
self.plotters[0].input = self.decision.winners_mem
self.plotters[0].gate_block = ~self.decision.epoch_ended
self.plotters[1].link_attrs(self.trainer, "shape") \
.link_from(self.ipython)
self.plotters[1].input = self.decision.weights_mem
self.plotters[1].gate_block = ~self.decision.epoch_ended
self.plotters[2].link_attrs(self.trainer, "shape") \
.link_from(self.ipython)
self.plotters[2].input = self.decision.weights_mem
self.plotters[2].gate_block = ~self.decision.epoch_ended
def __init__(self, workflow, **kwargs):
kwargs["name"] = kwargs.get("name", "Kohonen")
super(KohonenWorkflow, self).__init__(workflow, **kwargs)
self.downloader = Downloader(
self,
url=root.kohonen.downloader.url,
directory=root.kohonen.downloader.directory,
files=root.kohonen.downloader.files)
self.downloader.link_from(self.start_point)
self.repeater.link_from(self.downloader)
self.loader = KohonenLoader(
self,
name="Kohonen fullbatch loader",
minibatch_size=root.kohonen.loader.minibatch_size,
force_numpy=root.kohonen.loader.force_numpy)
self.loader.link_from(self.repeater)
# Kohonen training layer
self.trainer = kohonen.KohonenTrainer(
self,
shape=root.kohonen.forward.shape,
weights_filling=root.kohonen.forward.weights_filling,
weights_stddev=root.kohonen.forward.weights_stddev,
gradient_decay=root.kohonen.train.gradient_decay,
radius_decay=root.kohonen.train.radius_decay)
self.trainer.link_from(self.loader)
self.trainer.link_attrs(self.loader, ("input", "minibatch_data"))
# Loop decision
self.decision = kohonen.KohonenDecision(
self, max_epochs=root.kohonen.decision.epochs)
self.decision.link_from(self.trainer)
self.decision.link_attrs(self.loader, "minibatch_class",
"last_minibatch", "class_lengths",
"epoch_ended", "epoch_number")
self.decision.link_attrs(self.trainer, "weights", "winners")
self.ipython = Shell(self)
self.ipython.link_from(self.decision)
self.ipython.gate_skip = ~self.decision.epoch_ended
self.repeater.link_from(self.ipython)
self.ipython.gate_block = self.decision.complete
self.end_point.link_from(self.decision)
self.end_point.gate_block = ~self.decision.complete
self.loader.gate_block = self.decision.complete
# Error plotter
self.plotters = [
nn_plotting_units.KohonenHits(self),
nn_plotting_units.KohonenInputMaps(self),
nn_plotting_units.KohonenNeighborMap(self)
]
self.plotters[0].link_attrs(self.trainer, "shape") \
.link_from(self.ipython)
self.plotters[0].input = self.decision.winners_mem
self.plotters[0].gate_block = ~self.decision.epoch_ended
self.plotters[1].link_attrs(self.trainer, "shape") \
.link_from(self.ipython)
self.plotters[1].input = self.decision.weights_mem
self.plotters[1].gate_block = ~self.decision.epoch_ended
self.plotters[2].link_attrs(self.trainer, "shape") \
.link_from(self.ipython)
self.plotters[2].input = self.decision.weights_mem
self.plotters[2].gate_block = ~self.decision.epoch_ended
def __init__(self, workflow, **kwargs):
super(WineWorkflow, self).__init__(workflow, **kwargs)
layers = kwargs["layers"]
self.downloader = Downloader(
self, url=root.wine.downloader.url,
directory=root.wine.downloader.directory,
files=root.wine.downloader.files)
self.downloader.link_from(self.start_point)
self.repeater.link_from(self.downloader)
self.loader = WineLoader(
self, minibatch_size=root.wine.loader.minibatch_size,
force_numpy=root.wine.loader.force_numpy,
dataset_file=root.wine.loader.dataset_file,
normalization_type=root.wine.loader.normalization_type)
self.loader.link_from(self.repeater)
# Add fwds units
del self.forwards[:]
for i, layer in enumerate(layers):
if i < len(layers) - 1:
aa = all2all.All2AllTanh(
self, output_sample_shape=(layer,),
weights_stddev=0.05, bias_stddev=0.05)
else:
aa = all2all.All2AllSoftmax(
self, output_sample_shape=(layer,),
weights_stddev=0.05, bias_stddev=0.05)
self.forwards.append(aa)
if i:
self.forwards[-1].link_from(self.forwards[-2])
self.forwards[-1].link_attrs(
self.forwards[-2], ("input", "output"))
else:
self.forwards[-1].link_from(self.loader)
self.forwards[-1].link_attrs(
self.loader, ("input", "minibatch_data"))
# Add evaluator for single minibatch
self.evaluator = evaluator.EvaluatorSoftmax(self)
self.evaluator.link_from(self.forwards[-1])
self.evaluator.link_attrs(self.forwards[-1], "output", "max_idx")
self.evaluator.link_attrs(self.loader,
("batch_size", "minibatch_size"),
("max_samples_per_epoch", "total_samples"),
("labels", "minibatch_labels"),
("offset", "minibatch_offset"),
"class_lengths")
# Add decision unit
self.decision = decision.DecisionGD(
self, fail_iterations=root.wine.decision.fail_iterations,
max_epochs=root.wine.decision.max_epochs)
self.decision.link_from(self.evaluator)
self.decision.link_attrs(self.loader,
"minibatch_class", "minibatch_size",
"last_minibatch", "class_lengths",
"epoch_ended", "epoch_number")
self.decision.link_attrs(
self.evaluator,
("minibatch_n_err", "n_err"),
("minibatch_confusion_matrix", "confusion_matrix"),
("minibatch_max_err_y_sum", "max_err_output_sum"))
self.snapshotter = NNSnapshotterToFile(
self, prefix=root.wine.snapshotter.prefix,
directory=root.common.dirs.snapshots, compression="",
interval=root.wine.snapshotter.interval,
time_interval=root.wine.snapshotter.time_interval)
self.snapshotter.link_from(self.decision)
self.snapshotter.link_attrs(self.decision,
("suffix", "snapshot_suffix"))
self.snapshotter.gate_skip = ~self.loader.epoch_ended
self.snapshotter.skip = ~self.decision.improved
self.end_point.link_from(self.snapshotter)
self.end_point.gate_block = ~self.decision.complete
# Add gradient descent units
self.gds[:] = (None,) * len(self.forwards)
self.gds[-1] = gd.GDSoftmax(self) \
.link_from(self.snapshotter) \
.link_attrs(self.evaluator, "err_output") \
.link_attrs(self.forwards[-1], "output", "input",
"weights", "bias") \
.link_attrs(self.loader, ("batch_size", "minibatch_size"))
self.gds[-1].gate_skip = self.decision.gd_skip
self.gds[-1].gate_block = self.decision.complete
for i in range(len(self.forwards) - 2, -1, -1):
self.gds[i] = gd.GDTanh(self) \
.link_from(self.gds[i + 1]) \
.link_attrs(self.gds[i + 1], ("err_output", "err_input")) \
.link_attrs(self.forwards[i], "output", "input",
"weights", "bias") \
.link_attrs(self.loader, ("batch_size", "minibatch_size"))
self.gds[i].gate_skip = self.decision.gd_skip
self.gds[0].need_err_input = False
self.repeater.link_from(self.gds[0])
self.loader.gate_block = self.decision.complete
