def wrapper(self, *args, **kwargs):
result = getattr(self, new_method_name)(*args, **kwargs)
# pickle biflow
filename = (self._snapshot_name_ + "_%d" % self._snapshot_counter_ +
PICKLE_EXT)
robust_pickle(self._snapshot_path_, filename, self)
self._snapshot_counter_ += 1
return result
def wrapper(self, *args, **kwargs):
result = getattr(self, new_method_name)(*args, **kwargs)
# pickle biflow
filename = (self._snapshot_name_ + "_%d" % self._snapshot_counter_ +
PICKLE_EXT)
robust_pickle(self._snapshot_path_, filename, self)
self._snapshot_counter_ += 1
return result
def show_training(flow, data_iterables, msg_iterables=None, stop_messages=None,
path=None, tracer=None,
debug=False, show_size=False, open_browser=True,
**kwargs):
"""Perform both the flow training and the training inspection.
The return value is the filename of the slideshow HTML file.
This function must be used with the untrained flow (no previous call
of Flow.train is required, the training happens here).
This function is more convenient than inspect_training since it includes
all required steps, but it is also less customizable. After everything
is complete the inspection slideshow is opened in the browser.
flow -- The untrained Flow or BiFlow. After this function has been called
the flow will be fully trained.
data_iterables, msg_iterables, stop_messages -- Same as for calling train
on a flow.
path -- Path were both the training snapshots and the inspection slides
will be stored. If None (default value) a temporary directory will be
used.
tracer -- Instance of InspectionHTMLTracer, can be None for
default class.
debug -- Ignore exception during training and try to complete the slideshow
(default value is False).
show_size -- Show the approximate memory footprint of all nodes.
open_browser -- If True (default value) then the slideshow file is
automatically opened in a webbrowser. One can also use string value
with the browser name (for webbrowser.get) to request a specific
browser.
**kwargs -- Additional arguments for flow.train can be specified
as keyword arguments.
"""
if path is None:
path = tempfile.mkdtemp(prefix='MDP_')
# get first part of data iterators as sample data for inspection
# if data_iterables is an array, wrap it up in a list
if isinstance(data_iterables, numx.ndarray):
data_iterables = [[data_iterables]] * len(flow)
x_samples = []
for i, data_iterable in enumerate(data_iterables):
if data_iterable is None:
x_sample, new_data_iterable = None, None
else:
x_sample, new_data_iterable = first_iterable_elem(data_iterable)
x_samples.append(x_sample)
data_iterables[i] = new_data_iterable
del x_sample
if msg_iterables:
msg_samples = []
for i, msg_iterable in enumerate(msg_iterables):
if msg_iterable is None:
msg_sample, new_msg_iterable = None, None
else:
msg_sample, new_msg_iterable = first_iterable_elem(msg_iterable)
msg_samples.append(msg_sample)
msg_iterables[i] = new_msg_iterable
del msg_sample
else:
msg_samples = None
# store the data to disk to disk to save memory and safeguard against
# any change made to the data during the training
robust_pickle(path, "training_data_samples.pckl",
(x_samples, msg_samples, stop_messages))
del x_samples
del msg_samples
# perform the training and gather snapshots
prepare_training_inspection(flow=flow, path=path)
try:
if isinstance(flow, BiFlow):
flow.train(data_iterables, msg_iterables, stop_messages, **kwargs)
else:
flow.train(data_iterables, **kwargs)
except Exception:
if debug:
traceback.print_exc()
print ("exception during training, " +
"inspecting up to failure point...")
# create the last snapshot manually
try:
# if a normal mdp.Flow instance was given then this fails
flow._bi_reset()
except Exception:
pass
filename = (flow._snapshot_name_ + "_%d" % flow._snapshot_counter_
+ PICKLE_EXT)
robust_pickle(flow._snapshot_path_, filename, flow)
else:
raise
remove_inspection_residues(flow)
# reload data samples
with open(os.path.join(path, "training_data_samples.pckl"), "rb") as sample_file:
x_samples, msg_samples, stop_messages = pickle.load(sample_file)
# create slideshow
slideshow = inspect_training(snapshot_path=path,
inspection_path=path,
x_samples=x_samples,
msg_samples=msg_samples,
stop_messages=stop_messages,
tracer=tracer,
debug=debug, show_size=show_size,
verbose=False)
filename = os.path.join(path, "training_inspection.html")
title = "Training Inspection"
with open(filename, 'w') as html_file:
html_file.write('<html>\n<head>\n<title>%s</title>\n' % title)
html_file.write('<style type="text/css" media="screen">')
html_file.write(standard_css())
html_file.write('</style>\n</head>\n<body>\n')
html_file.write('<h3>%s</h3>\n' % title)
html_file.write(slideshow)
html_file.write('</body>\n</html>')
if open_browser:
_open_custom_brower(open_browser, os.path.abspath(filename))
return filename
def show_training(flow,
data_iterables,
msg_iterables=None,
stop_messages=None,
path=None,
tracer=None,
debug=False,
show_size=False,
open_browser=True,
**kwargs):
"""Perform both the flow training and the training inspection.
The return value is the filename of the slideshow HTML file.
This function must be used with the untrained flow (no previous call
of Flow.train is required, the training happens here).
This function is more convenient than inspect_training since it includes
all required steps, but it is also less customizable. After everything
is complete the inspection slideshow is opened in the browser.
flow -- The untrained Flow or BiFlow. After this function has been called
the flow will be fully trained.
data_iterables, msg_iterables, stop_messages -- Same as for calling train
on a flow.
path -- Path were both the training snapshots and the inspection slides
will be stored. If None (default value) a temporary directory will be
used.
tracer -- Instance of InspectionHTMLTracer, can be None for
default class.
debug -- Ignore exception during training and try to complete the slideshow
(default value is False).
show_size -- Show the approximate memory footprint of all nodes.
open_browser -- If True (default value) then the slideshow file is
automatically opened in a webbrowser. One can also use string value
with the browser name (for webbrowser.get) to request a specific
browser.
**kwargs -- Additional arguments for flow.train can be specified
as keyword arguments.
"""
if path is None:
path = tempfile.mkdtemp(prefix='MDP_')
# get first part of data iterators as sample data for inspection
# if data_iterables is an array, wrap it up in a list
if isinstance(data_iterables, numx.ndarray):
data_iterables = [[data_iterables]] * len(flow)
x_samples = []
for i, data_iterable in enumerate(data_iterables):
if data_iterable is None:
x_sample, new_data_iterable = None, None
else:
x_sample, new_data_iterable = first_iterable_elem(data_iterable)
x_samples.append(x_sample)
data_iterables[i] = new_data_iterable
del x_sample
if msg_iterables:
msg_samples = []
for i, msg_iterable in enumerate(msg_iterables):
if msg_iterable is None:
msg_sample, new_msg_iterable = None, None
else:
msg_sample, new_msg_iterable = first_iterable_elem(
msg_iterable)
msg_samples.append(msg_sample)
msg_iterables[i] = new_msg_iterable
del msg_sample
else:
msg_samples = None
# store the data to disk to disk to save memory and safeguard against
# any change made to the data during the training
robust_pickle(path, "training_data_samples.pckl",
(x_samples, msg_samples, stop_messages))
del x_samples
del msg_samples
# perform the training and gather snapshots
prepare_training_inspection(flow=flow, path=path)
try:
if isinstance(flow, BiFlow):
flow.train(data_iterables, msg_iterables, stop_messages, **kwargs)
else:
flow.train(data_iterables, **kwargs)
except Exception:
if debug:
traceback.print_exc()
print("exception during training, " +
"inspecting up to failure point...")
# create the last snapshot manually
try:
# if a normal mdp.Flow instance was given then this fails
flow._bi_reset()
except Exception:
pass
filename = (flow._snapshot_name_ +
"_%d" % flow._snapshot_counter_ + PICKLE_EXT)
robust_pickle(flow._snapshot_path_, filename, flow)
else:
raise
remove_inspection_residues(flow)
# reload data samples
with open(os.path.join(path, "training_data_samples.pckl"),
"rb") as sample_file:
x_samples, msg_samples, stop_messages = pickle.load(sample_file)
# create slideshow
slideshow = inspect_training(snapshot_path=path,
inspection_path=path,
x_samples=x_samples,
msg_samples=msg_samples,
stop_messages=stop_messages,
tracer=tracer,
debug=debug,
show_size=show_size,
verbose=False)
filename = os.path.join(path, "training_inspection.html")
title = "Training Inspection"
with open(filename, 'w') as html_file:
html_file.write('<html>\n<head>\n<title>%s</title>\n' % title)
html_file.write('<style type="text/css" media="screen">')
html_file.write(standard_css())
html_file.write('</style>\n</head>\n<body>\n')
html_file.write('<h3>%s</h3>\n' % title)
html_file.write(slideshow)
html_file.write('</body>\n</html>')
if open_browser:
_open_custom_brower(open_browser, os.path.abspath(filename))
return filename
