def get_feed_dicts(outputs):
"""Get the training and evaluation dictionaries that map placeholders to the
values that they should take during training and evaluation, respectively
(e.g., used for dropout or batch normalization).
Args:
outputs (dict[str, deep_architect.core.Output]): Dictionary of named
outputs of the model (i.e., with no unspecified hyperparameters
available) sampled from the search space.
Returns:
(dict, dict):
Training and evaluation dictionaries where the keys are placeholders
and the values are the values the placeholders should take during
each of these stages.
"""
train_feed = {}
eval_feed = {}
def fn(x):
if hasattr(x, 'train_feed'):
train_feed.update(x.train_feed)
if hasattr(x, 'eval_feed'):
eval_feed.update(x.eval_feed)
return False
co.traverse_backward(outputs, fn)
return (train_feed, eval_feed)
def setTraining(output_lst, isTraining):
def fn(mx):
if hasattr(mx, 'isTraining'):
mx.isTraining = isTraining
co.traverse_backward(output_lst, fn)
def setRecompile(output_lst, recompile):
def fn(mx):
mx._is_compiled = not recompile
co.traverse_backward(output_lst, fn)
logger.debug('set_recompile')
def extract_features(inputs, outputs):
"""Extract a feature representation of a model represented through inputs and
outputs.
This function has been mostly used for performance prediction on fully
specified models, i.e., after all the hyperparameters in the search space
have specified. After this, there is a single model for which we can compute
an appropriate feature representation containing information about the
connections that the model makes and the values of the hyperparameters.
Args:
inputs (dict[str, deep_architect.core.Input]): Dictionary mapping names
to inputs of the architecture.
outputs (dict[str, deep_architect.core.Output]): Dictionary mapping names to outputs
of the architecture.
Returns:
dict[str, list[str]]:
Representation of the architecture as a dictionary where each
key is associated to a list with different types of features.
"""
module_memo = co.OrderedSet()
module_feats = []
connection_feats = []
module_hyperp_feats = []
# getting all the modules
module_memo = []
def fn(m):
module_memo.append(m)
co.traverse_backward(outputs, fn)
for m in module_memo:
# module features
m_feats = m.get_name()
module_feats.append(m_feats)
for ox_localname, ox in iteritems(m.outputs):
if ox.is_connected():
ix_lst = ox.get_connected_inputs()
for ix in ix_lst:
# connection features
c_feats = "%s |-> %s" % (ox.get_name(), ix.get_name())
connection_feats.append(c_feats)
# module hyperparameters
for h_localname, h in iteritems(m.hyperps):
mh_feats = "%s/%s : %s = %s" % (m.get_name(), h_localname,
h.get_name(), h.get_value())
module_hyperp_feats.append(mh_feats)
return {
'module_feats': module_feats,
'connection_feats': connection_feats,
'module_hyperp_feats': module_hyperp_feats,
}
def _call_fn_on_pytorch_module(outputs, fn):
def fn_iter(mx):
if hasattr(mx, 'pyth_modules'):
for pyth_m in mx.pyth_modules:
fn(pyth_m)
return False
co.traverse_backward(outputs, fn_iter)
def draw_graph(outputs,
draw_hyperparameters=True,
draw_io_labels=True,
draw_module_hyperparameter_info=True,
out_folderpath=None,
graph_name='graph',
print_to_screen=True):
"""Draws a graph representation of the current state of the search space.
All edges are directed. An edge between two modules represents the output of
the first module going into the input of the second module. An edge between
a module and an hyperparameter represents the dependency of that module on
that hyperparameter.
This visualization functionality is useful to verify that the description of
the search space done through the domain-specific language encodes the
intended search space.
.. note::
All drawing options are set to `True` to give a sense of all the
information that can be displayed simultaneously. This can lead to
cluttered graphs and slow rendering. We recommend changing the defaults
according to the desired behavior.
For generating a representation for a fully specified model,
we recommend setting `draw_hyperparameters` to `False`, as if we are
only concerned with the resulting model, the sharing structure of
hyperparameters does not really matter. We recommend using
`draw_hyperparameters` set to `True` when the user wishes to visualize
the hyperparameter sharing pattern, e.g., to verify that it has been
implemented correctly.
Args:
outputs (dict[str, deep_architect.core.Output]): Dictionary of named
outputs from which we can reach all the modules in the search space
by backwards traversal.
draw_hyperparameters (bool, optional): Draw hyperparameter nodes in the
graph, representing the dependencies between hyperparameters and
modules.
draw_io_labels (bool, optional): If `True`,
draw edge labels connecting different modules
with the local names of the input and output that are being connected.
draw_module_hyperparameter_info (bool, optional): If `True`,
draw the hyperparameters of a module alongside the module.
graph_name (str, optional): Name of the file used to store the
rendered graph. Only needs to be provided if we desire to output
the graph to a file, rather than just show it.
out_folderpath (str, optional): Folder to which to store the PDF file with
the rendered graph. If no path is provided, no file is created.
print_to_screen (bool, optional): Shows the result of rendering the
graph directly to screen.
"""
assert print_to_screen or out_folderpath is not None
g = graphviz.Digraph()
edge_fs = '10'
h_fs = '10'
penwidth = '1'
def _draw_connected_input(ix_localname, ix):
ox = ix.get_connected_output()
if not draw_io_labels:
label = ''
else:
ox_localname = None
for ox_iter_localname, ox_iter in ox.get_module().outputs.items():
if ox_iter == ox:
ox_localname = ox_iter_localname
break
assert ox_localname is not None
label = ix_localname + ':' + ox_localname
g.edge(ox.get_module().get_name(),
ix.get_module().get_name(),
label=label,
fontsize=edge_fs)
def _draw_unconnected_input(ix_localname, ix):
g.node(ix.get_name(),
shape='invhouse',
penwidth=penwidth,
fillcolor='firebrick',
style='filled')
g.edge(ix.get_name(), ix.get_module().get_name())
def _draw_module_hyperparameter(m, h_localname, h):
if h.has_value_assigned():
label = h_localname + '=' + str(h.get_value())
else:
label = h_localname
g.edge(h.get_name(), m.get_name(), label=label, fontsize=edge_fs)
def _draw_dependent_hyperparameter_relations(h_dep):
for h_localname, h in h_dep._hyperps.items():
if h.has_value_assigned():
label = h_localname + '=' + str(h.get_value())
else:
label = h_localname
g.edge(h.get_name(),
h_dep.get_name(),
label=label,
fontsize=edge_fs)
def _draw_module_hyperparameter_info(m):
g.node(
m.get_name(),
xlabel="<" + '<br align="right"/>'.join([
'<FONT POINT-SIZE="%s">' % h_fs + h_localname +
('=' + str(h.get_value()) if h.has_value_assigned() else '') +
"</FONT>" for h_localname, h in m.hyperps.items()
]) + ">")
def _draw_output_terminal(ox_localname, ox):
g.node(ox.get_name(),
shape='house',
penwidth=penwidth,
fillcolor='deepskyblue',
style='filled')
g.edge(ox.get_module().get_name(), ox.get_name())
nodes = set()
def fn(m):
"""Adds the module information to the graph that is local to the module.
"""
nodes.add(m.get_name())
for ix_localname, ix in m.inputs.items():
if ix.is_connected():
_draw_connected_input(ix_localname, ix)
else:
_draw_unconnected_input(ix_localname, ix)
if draw_hyperparameters:
for h_localname, h in m.hyperps.items():
_draw_module_hyperparameter(m, h_localname, h)
if draw_module_hyperparameter_info:
_draw_module_hyperparameter_info(m)
return False
# generate most of the graph.
co.traverse_backward(outputs, fn)
# drawing the hyperparameter graph.
if draw_hyperparameters:
hs = co.get_all_hyperparameters(outputs)
for h in hs:
if isinstance(h, co.DependentHyperparameter):
_draw_dependent_hyperparameter_relations(h)
g.node(h.get_name(),
fontsize=h_fs,
fillcolor='darkseagreen',
style='filled')
# add the output terminals.
for m in co.extract_unique_modules(list(outputs.values())):
for ox_localname, ox in m.outputs.items():
_draw_output_terminal(ox_localname, ox)
# minor adjustments to attributes.
for s in nodes:
g.node(s,
shape='rectangle',
penwidth=penwidth,
fillcolor='goldenrod',
style='filled')
if print_to_screen or out_folderpath is not None:
g.render(graph_name,
out_folderpath,
view=print_to_screen,
cleanup=True)
def set_is_training(outputs, is_training):
def fn(mx):
if hasattr(mx, 'is_training'):
mx.is_training = is_training
co.traverse_backward(outputs, fn)
def set_recompile(outputs, recompile):
def fn(mx):
mx._is_compiled = not recompile
co.traverse_backward(outputs, fn)