def __init__(self):
self.training_items = OrderedDict()
self.training_names = OrderedDict()
self._dataloaders = OrderedDict()
self.num_epochs = 1
self._minibatch_size = 1
self.warmup = 0
self.default_collect_data_inteval = 1
self.print_progress_frequency = 10
self.print_progress_unit = 'batch'
self.print_progress_on_epoch_end = False
self.out_sample_evaluation_frequency = 1
self.out_sample_evaluation_unit = 'epoch'
self.out_sample_evaluation_on_epoch_end = True
self.save_model_frequency = -1
self.save_model_unit = 'batch'
self.execution_id = None
self._is_optimizer_warmup = False
self.callbacks = [
] # if self.callbacks is None: # self.callbacks = [ #
# NumberOfEpochsStoppingCriterionCallback(1)] # elif not any([issubclass(type(cb),
# StoppingCriterionCallback) for cb in self.callbacks]): # # self.callbacks.append( #
# NumberOfEpochsStoppingCriterionCallback(1))
self.is_terminate = False
def __init__(self,
data_source,
batch_size=1,
is_shuffle=True,
drop_last=False,
sample_filter=None):
super().__init__(data_source)
if not isinstance(batch_size, int) or isinstance(
batch_size, bool) or batch_size <= 0:
raise ValueError(
"batch_size should be a positive integeral value, "
"but got batch_size={}".format(batch_size))
if not isinstance(drop_last, bool):
raise ValueError("drop_last should be a boolean value, but got "
"drop_last={}".format(drop_last))
self.data_source = data_source
self.batch_size = batch_size
self.drop_last = drop_last
self.is_shuffle = is_shuffle
self._batch_transform_funcs = OrderedDict()
idxes = np.arange(len(self.data_source.data))
if len(self.data_source) % self.batch_size > 0:
idxes = idxes[:-(len(self.data_source.data) % self.batch_size)]
if self.is_shuffle:
n = len(self.data_source)
random_range = np.arange(n)
np.random.shuffle(random_range)
self.sampler = itertools.cycle(iter(random_range.tolist()))
else:
self.sampler = itertools.cycle(iter(range(len(self.data_source))))
self.sample_filter = None
if inspect.isfunction(sample_filter) or callable(sample_filter):
self.sample_filter = sample_filter
def apply_batch(self,
inputs: Sequence[Tuple],
spec: Optional[TensorSpec] = None):
r"""Apply transform on batch input data."""
if not isinstance(inputs, OrderedDict):
if spec is None and self.is_spatial == True:
self._shape_info = None
spec = TensorSpec(shape=tensor_to_shape(
inputs, need_exclude_batch_axis=True, is_singleton=True),
object_type=ObjectType.rgb)
return self.apply(inputs, spec)
else:
results = OrderedDict()
sampledata = list(inputs.values())[0]
spec = inputs.key_list[0]
if (isinstance(sampledata, Iterable)
and not isinstance(sampledata, np.ndarray)) or (
is_tensor_like(sampledata)
and spec.ndim == sampledata.ndim):
for i in range(len(sampledata)):
self._shape_info = None
for spec, data in inputs.items():
if spec not in results:
results[spec] = []
results[spec].append(self.apply(data[i], spec))
else:
self._shape_info = None
for spec, data in inputs.items():
results[spec] = self.apply(data, spec)
return results
def apply_batch(self,
inputs: Sequence[Tuple],
spec: Optional[TensorSpec] = None):
r"""Apply transform on batch input data."""
if not isinstance(inputs, OrderedDict):
self._text_info = None
if spec is None and self.is_spatial == True:
spec = TensorSpec(shape=tensor_to_shape(
inputs, need_exclude_batch_axis=True, is_singleton=True),
object_type=ObjectType.corpus)
self._precalculate(inputs)
return self.apply(inputs, spec)
else:
results = OrderedDict()
self._text_info = None
is_precalculate = False
for k, v in inputs.items():
if k.object_type is None:
k.object_type = object_type_inference(v)
if isinstance(
k, TensorSpec) and k.object_type == ObjectType.corpus:
self._precalculate(v)
is_precalculate = True
if not is_precalculate:
self._precalculate(inputs.value_list[0])
for spec, data in inputs.items():
results[spec] = self.apply(data, spec)
return results
def get_samples(self, batch_size=8):
if self.enable:
can_sample = len(list(set([len(v) for v in self.value_list]))) == 1
indexs = random_choice(
list(range(len(self.value_list[0]))),
builtins.min(batch_size, len(self.value_list[0])))
return_data = OrderedDict()
if can_sample:
for k, v in self.items():
return_data[k] = [v[idx] for idx in indexs]
return return_data
else:
raise ValueError(
'get_sample need all collection have same length.')
else:
return OrderedDict()
def __init__(self, convert_ratio=0.5, name='random_homophonic_typo', **kwargs):
super().__init__()
self.convert_ratio = convert_ratio
self.text2pronounce = OrderedDict()
self.pronounce2text = OrderedDict()
self.char_freq = OrderedDict()
if not get_session().get_resources('char_freq'):
char_freq = get_session().regist_resources('char_freq', OrderedDict())
with open(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'char_freq.txt'), 'r', encoding='utf-8-sig') as f:
for line in f.readlines():
cols = line.strip().split('\t')
char_freq[cols[0]] = float(cols[1])
self.char_freq = char_freq
else:
self.char_freq = get_session().get_resources('char_freq')
with open(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'text2pronounce.txt'), 'r', encoding='utf-8-sig') as f:
lines = f.readlines()
for line in lines:
words = line.strip().split('\t')
if len(words) == 3:
prons = words[2].split(' ')
if 0 < len(prons) <= 2:
prons = prons[:1]
elif 5 > len(prons) > 2:
prons = prons[:2]
elif len(prons) >= 5:
prons = prons[:3]
self.text2pronounce[words[0]] = prons
if words[0] in self.char_freq and self.char_freq[words[0]] > -15.5:
for p in prons:
if p not in self.pronounce2text:
self.pronounce2text[p] = []
self.pronounce2text[p].append(words[0])
self.name = name
def apply_batch(self,
inputs: Sequence[Tuple],
spec: Optional[TensorSpec] = None):
r"""Apply transform on batch input data."""
if not isinstance(inputs, OrderedDict):
if spec is None and self.is_spatial == True:
self._shape_info = None
spec = TensorSpec(shape=tensor_to_shape(
inputs, need_exclude_batch_axis=True, is_singleton=True),
object_type=ObjectType.rgb)
return self.apply(inputs, spec)
else:
results = OrderedDict()
self._shape_info = None
for spec, data in inputs.items():
results[spec] = self.apply(data, spec)
return results
def __init__(self, convert_ratio=0.5, name='random_homomorphic_typo', **kwargs):
super().__init__()
self.convert_ratio = convert_ratio
download_file_from_google_drive('1MDk7eH7nORa16SyzNzqv7fYzBofzxGRI', dirname=os.path.join(get_trident_dir(), 'download'), filename='chardict.pkl')
self.chardict = unpickle(os.path.join(get_trident_dir(), 'download', 'chardict.pkl'))
self.all_embedding = to_tensor(np.stack(self.chardict.value_list, 0)).to('cpu')
self.name = name
if not get_session().get_resources('char_freq'):
char_freq = get_session().regist_resources('char_freq', OrderedDict())
with open(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'char_freq.txt'), 'r', encoding='utf-8-sig') as f:
for line in f.readlines():
cols = line.strip().split('\t')
char_freq[cols[0]] = float(cols[1])
self.char_freq = char_freq
else:
self.char_freq = get_session().get_resources('char_freq')
def load_lfw(format='aligned_face', is_paired=False):
"""
Args:
format (str):
is_paired (bool): if True, will return anchor-positive-negative pair, or return image-classification label.
Returns:
Exsamples:
>>> dataloader=load_lfw(only_face_area=True, only_aligned_face=True)
>>> len(dataloader)
10
"""
dataset_name = 'lfw'
dirname = os.path.join(_trident_dir, dataset_name)
if not os.path.exists(dirname):
try:
os.makedirs(dirname)
except OSError:
# Except permission denied and potential race conditions
# in multi-threaded environments.
pass
tar_file_path = os.path.join(dirname, 'lfw.tgz')
if format == 'aligned_face':
download_file_from_google_drive('1sVbU8NHC7kzDkqByTtPiib9_EY23eeW7',
dirname, 'lfw-crop.tar')
elif format == 'raw':
download_file('http://vis-www.cs.umass.edu/lfw/lfw.tgz', dirname,
'lfw.tgz')
download_file('http://vis-www.cs.umass.edu/lfw/pairsDevTrain.txt', dirname,
'pairsDevTrain.txt')
download_file('http://vis-www.cs.umass.edu/lfw/pairsDevTest.txt', dirname,
'pairsDevTest.txt')
# if only_face_area:
# if only_aligned_face:
tar_file_path = os.path.join(dirname, 'lfw.tgz')
extract_archive(tar_file_path, dirname, archive_format='tar')
extract_path = os.path.join(
dirname, 'lfw-crop' if format == 'aligned_face' else 'lfw')
# if _backend=='tensorflow':
# import trident.models.tensorflow_mtcnn as mtcnn
# else:
# import trident.models.pytorch_mtcnn as mtcnn
#
# detector = mtcnn.Mtcnn(pretrained=True,verbose=False)
# detector.minsize = 70
#
# detector.detection_threshould = [0.8, 0.8, 0.9]
# detector.nms_threshould = [0.5, 0.5, 0.3]
# faces = glob.glob(os.path.join(dirname, 'lfw-deepfunneled' if only_aligned_face else 'lfw') + '/*/*.*g')
# faces_crop=glob.glob(os.path.join(dirname, 'lfw-crop')+ '/*/*.*g')
# if len(faces_crop)==0 or len(faces_crop)!=len(faces):
# for m in range(len(faces)):
# img_path=faces[m]
# new_path = img_path.replace(os.path.join(dirname, 'lfw-deepfunneled' if only_aligned_face else 'lfw'), os.path.join(dirname, 'lfw-crop'))
#
# if os.path.exists(new_path) and image2array(new_path) is not None:
# pass
# else:
# make_dir_if_need(new_path)
# img = read_image(img_path)
# detector.detection_threshould = [0.8, 0.8, 0.9]
# results = detector.infer_single_image(img_path)
# if len(results)==0 or results is None:
# detector.detection_threshould = [0.5, 0.7, 0.8]
# results = detector.infer_single_image(img_path)
#
# results = detector.rerec(to_tensor(results), img.shape)
# results[:, 0] = clip(results[:, 0], 0, img.shape[1])
# results[:, 1] = clip(results[:, 1], 0, img.shape[0])
# results[:, 2] = clip(results[:, 2], 0, img.shape[1])
# results[:, 3] = clip(results[:, 3], 0, img.shape[0])
# area=detector.area_of(results[:,:2],results[:,2:4])
# area_mask=area>500
# results=to_numpy(results[area_mask,:])
#
# if len(results)>1:
# area=area[area_mask]
# idx=argmax(area).item()
# results=results[idx:idx+1]
# if results is not None and len(results)>0:
# for k in range(len(results)):
# result =np.round(results[k][:4]).astype(np.uint8)
# x1, y1, x2, y2 = result[0],result[1],result[2],result[3]
# crop_img =np.clip( img.copy()[y1:y2,x1:x2, :],0,255).astype(np.uint8)
#
# array2image(crop_img).save(new_path)
# else:
#
# print(img_path+' get {0} results!'.format(results))
#
# sys.stdout.write('\r {0}/{1} image processed!'.format(m,len(faces)))
# sys.stdout.flush()
#
#
data_provider = load_folder_images(dataset_name,
os.path.join(dirname, 'lfw-crop'))
if is_paired:
storage = OrderedDict()
data = list(data_provider.traindata.data)
label = list(data_provider.traindata.label)
class_names = data_provider._class_names
for i in range(len(data)):
class_label = class_names[label[i]]
if class_label not in storage:
storage[class_label] = []
storage[class_label].append(data[i])
metric = MetricIterator(storage.item_list)
data_provider = DataProvider(dataset_name,
data=metric,
scenario='train')
# extract_archive(tar_file_path, dirname, archive_format='tar')
# dataset = load_folder_images(dataset_name, dirname)
return data_provider
class TrainingPlan(object):
def __init__(self):
self.training_items = OrderedDict()
self.training_names = OrderedDict()
self._dataloaders = OrderedDict()
self.num_epochs = 1
self._minibatch_size = 1
self.warmup = 0
self.default_collect_data_inteval = 1
self.print_progress_frequency = 10
self.print_progress_unit = 'batch'
self.print_progress_on_epoch_end = False
self.out_sample_evaluation_frequency = 1
self.out_sample_evaluation_unit = 'epoch'
self.out_sample_evaluation_on_epoch_end = True
self.save_model_frequency = -1
self.save_model_unit = 'batch'
self.execution_id = None
self._is_optimizer_warmup = False
self.callbacks = [
] # if self.callbacks is None: # self.callbacks = [ #
# NumberOfEpochsStoppingCriterionCallback(1)] # elif not any([issubclass(type(cb),
# StoppingCriterionCallback) for cb in self.callbacks]): # # self.callbacks.append( #
# NumberOfEpochsStoppingCriterionCallback(1))
self.is_terminate = False
@property
def minibatch_size(self):
return self._minibatch_size
@minibatch_size.setter
def minibatch_size(self, value):
self._minibatch_size = value
for i, (k, v) in enumerate(self._dataloaders.items()):
v.minibatch_size = value
self._dataloaders[k] = v
def with_callbacks(self, *callbacks):
if len(self.callbacks) == 0:
self.callbacks = to_list(callbacks)
else:
self.callbacks.extend(callbacks)
return self
def __getattr__(self, name):
if name == 'self':
return self
if '_training_items' in self.__dict__:
_training_items = self.__dict__['_training_items']
if name in _training_items:
return _training_items[name]
if '_dataloaders' in self.__dict__:
_dataloaders = self.__dict__['_dataloaders']
if name in _dataloaders:
return _dataloaders[name]
raise AttributeError("'{}' object has no attribute '{}'".format(
type(self).__name__, name))
def extra_repr(self):
r"""Set the extra representation of the module
To print customized extra information, you should reimplement
this method in your own modules. Both single-line and multi-line
strings are acceptable.
"""
return ''
def __str__(self):
self.__repr__()
def _get_name(self):
return self.__class__.__name__
def __repr__(self):
# We treat the extra repr like the sub-module, one item per line
extra_lines = []
extra_repr = self.extra_repr()
# empty string will be split into list ['']
if extra_repr:
extra_lines = extra_repr.split('\n')
child_lines = []
for key, value in self.__dict__.items():
if isinstance(value, OrderedDict):
for subkey, subvalue in value.items():
mod_str = repr(subvalue)
mod_str = addindent(mod_str, 2)
child_lines.append('(' + key + '): ' + mod_str)
else:
mod_str = repr(value)
mod_str = addindent(mod_str, 2)
child_lines.append('(' + key + '): ' + mod_str)
lines = extra_lines + child_lines
main_str = self._get_name() + '('
if lines:
# simple one-liner info, which most builtin Modules will use
if len(extra_lines) == 1 and not child_lines:
main_str += extra_lines[0]
else:
main_str += '\n ' + '\n '.join(lines) + '\n'
main_str += ')'
return main_str
def __dir__(self):
module_attrs = dir(self.__class__)
# training_items = list(self.training_items.keys())
keys = module_attrs
# Eliminate attrs that are not legal Python variable names
keys = [key for key in keys if not key[0].isdigit()]
return sorted(keys)
@classmethod
def create(cls):
plan = cls()
return plan
def add_training_item(self, training_item, name=None, start_epoch=0):
n = len(self.training_items)
if name is not None and len(name) > 0:
self.training_names[n] = name
training_item.name = name
elif training_item.name is not None and len(training_item.name) > 0:
self.training_names[n] = training_item.name
else:
training_item.name = 'model {0}'.format(n)
self.training_names[n] = 'model {0}'.format(n)
self.training_items[n] = training_item
self.training_items[n].start_epoch = start_epoch
# backward compatibility
for k, v in training_item.inputs.items():
if isinstance(v, tuple) and all(
[isinstance(item, numbers.Integral) for item in v]):
training_item.inputs[k] = TensorSpec(shape=to_tensor(v),
name=training_item.name)
training_item.signature.inputs[k] = TensorSpec(
shape=to_tensor(v), name=training_item.name)
elif isinstance(v, TensorSpec):
training_item.signature.inputs[k] = v
for k, v in training_item.outputs.items():
if isinstance(v, tuple) and all(
[isinstance(item, numbers.Integral) for item in v]):
training_item.outputs[k] = TensorSpec(shape=to_tensor(v),
name=training_item.name)
training_item.signature.outputs[k] = TensorSpec(
shape=to_tensor(v), name=training_item.name)
elif isinstance(v, TensorSpec):
training_item.signature.outputs[k] = v
if isinstance(
training_item.model, Layer
) and training_item.signature != training_item.model.signature:
training_item.model.signature = None
training_item.signature = training_item.model.signature
return self
def with_data_loader(self, data_loader, **kwargs):
self._dataloaders[data_loader.__class__.__name__] = data_loader
return self
def repeat_epochs(self, num_epochs: int):
self.num_epochs = num_epochs
return self
def within_minibatch_size(self, minibatch_size: int):
self.minibatch_size = minibatch_size
return self
def out_sample_evaluation_scheduling(self,
frequency: int,
unit='batch',
on_epoch_end=True):
self.out_sample_evaluation_on_epoch_end = on_epoch_end
self.out_sample_evaluation_frequency = frequency
if unit not in ['batch', 'epoch']:
raise ValueError('unit should be batch or epoch')
else:
self.out_sample_evaluation_unit = unit
return self
def print_progress_scheduling(self,
frequency: int,
unit='batch',
on_epoch_end=True,
show_loss_metric_curve=True):
self.print_progress_on_epoch_end = on_epoch_end
self.print_progress_frequency = frequency
self.default_collect_data_inteval = frequency
if unit not in ['batch', 'epoch']:
raise ValueError('unit should be batch or epoch')
else:
self.print_progress_unit = unit
for i in range(len(self.training_items)):
self.training_items[i].training_context[
'print_progress_frequency'] = frequency
self.training_items[i].training_context[
'print_progress_unit'] = self.print_progress_unit
return self
def print_gradients_scheduling(self, frequency: int, unit='batch'):
pg = PrintGradientsCallback(batch_inteval=frequency)
pg.is_shared = False
self.callbacks.append(pg)
return self
def save_model_scheduling(self, frequency: int, unit='batch'):
self.save_model_frequency = frequency
if unit not in ['batch', 'epoch']:
raise ValueError('unit should be batch or epoch')
else:
self.save_model_unit = unit
return self
def display_tile_image_scheduling(self,
frequency: int,
unit='batch',
save_path: str = None,
name_prefix: str = 'tile_image_{0}.png',
include_input=True,
include_output=True,
include_target=True,
include_mask=None,
imshow=None):
if (is_in_ipython() or is_in_colab()) and imshow is None:
imshow = True
elif not is_in_ipython() and not is_in_colab() and imshow is None:
imshow = False
if unit not in ['batch', 'epoch']:
raise ValueError('unit should be batch or epoch')
tile = TileImageCallback(frequency if unit == 'epoch' else -1,
frequency if unit == 'batch' else -1,
save_path=save_path,
name_prefix=name_prefix,
include_input=include_input,
include_output=include_output,
include_target=include_target,
include_mask=include_mask,
imshow=imshow)
self.callbacks.append(tile)
return self
def display_loss_metric_curve_scheduling(
self,
frequency: int,
unit='batch',
save_path: str = None,
name_prefix: str = 'loss_metric_curve_{0}.png',
clean_ipython_output_frequency=5,
imshow=None):
if (is_in_ipython() or is_in_colab()) and imshow is None:
imshow = True
elif not is_in_ipython() and not is_in_colab() and imshow is None:
imshow = False
if save_path is not None:
folder, _, _ = split_path(save_path)
if not os.path.exists(folder):
try:
os.makedirs(folder)
except Exception as e:
PrintException()
raise ValueError(
'save_path:{0} is not valid path'.format(folder))
plot = PlotLossMetricsCallback(
frequency if unit == 'epoch' else -1,
frequency if unit == 'batch' else -1,
save_path=save_path,
name_prefix=name_prefix.format(get_time_suffix()),
clean_ipython_output_frequency=clean_ipython_output_frequency,
imshow=imshow)
self.callbacks.append(plot)
return self
def generate_datafeed(self, data_loader):
if data_loader.signature is None:
_ = data_loader.next()
# data_input=data_loader.traindata.data.symbol
# if len(data_loader.traindata.unpair)>0:
# data_unpair = data_loader.traindata.unpair
for trainingitem in self.training_items.value_list:
existing_data_feed = None
if 'data_feed' in trainingitem.training_context:
existing_data_feed = trainingitem.training_context['data_feed']
data_feed = OrderedDict()
datasets = data_loader.traindata.get_datasets()
available_items = data_loader.signature.outputs.key_list
available_items.extend(trainingitem.signature.outputs.key_list)
available_items = list(set(available_items))
for inp in trainingitem.signature.inputs.key_list:
data_feed[inp] = None
for k, v in trainingitem._losses.items():
for inp in v.signature.inputs.key_list:
data_feed[inp] = None
for k, v in trainingitem._metrics.items():
for inp in v.signature.inputs.key_list:
data_feed[inp] = None
if 'x' in data_feed and 'input' in available_items:
data_feed['x'] = 'input'
available_items.remove('input')
data_symbols = iteration_tools.flatten(
[data_loader.traindata.data.symbol],
iterable_types=(list, tuple))
label_symbols = iteration_tools.flatten(
[data_loader.traindata.label.symbol],
iterable_types=(list, tuple))
unpair_symbols = iteration_tools.flatten(
[data_loader.traindata.unpair.symbol],
iterable_types=(list, tuple))
if "" in label_symbols:
label_symbols.remove("")
if "" in unpair_symbols:
unpair_symbols.remove("")
if len(trainingitem.signature.inputs) == len(data_symbols) == 1:
if assert_spec_compatibility(
trainingitem.signature.inputs.value_list[0],
data_loader.traindata.data.element_spec):
data_feed[trainingitem.signature.inputs.
key_list[0]] = data_loader.traindata.data.symbol
available_items.remove(data_loader.traindata.data.symbol)
if len(trainingitem.signature.outputs) == len(label_symbols) == 1:
data_feed[trainingitem.signature.outputs.key_list[0].replace(
"output", "target").replace(
"student",
"teacher")] = data_loader.traindata.label.symbol
available_items.remove(data_loader.traindata.label.symbol)
elif len(trainingitem.signature.outputs) == 1 and len(
data_symbols) == 1 and len(label_symbols) == 0:
# autoencoder
data_feed[trainingitem.signature.outputs.key_list[0].replace(
"output", "target").replace(
"student",
"teacher")] = data_loader.traindata.data.symbol
for out in trainingitem.signature.outputs.key_list: # fill the data_feed by key
if out in available_items: # output=putput
data_feed[out] = out
available_items.remove(out)
for key in data_feed.keys():
if data_feed[key] == None and key in available_items:
data_feed[key] = key
available_items.remove(key)
elif data_feed[key] == None:
data_feed[key] = key
# elif out.replace("output","target").replace("student","teacher") in available_items:
# data_feed[out] =out.replace("output","target").replace("student","teacher")
# available_items.remove(out.replace("output","target").replace("student","teacher"))
trainingitem.training_context['data_feed'] = data_feed
print('data_feed for {0} :{1}'.format(trainingitem.name,
data_feed))
def start_now(self,
collect_data_inteval=1,
is_resume=False,
only_steps=False,
max_batches=np.inf,
keep_weights_history=False,
keep_gradient_history=False):
try:
self.execution_id = get_time_suffix()
exception_cnt = 0
abnormal_num_count = 0
# update callback
if not is_resume or only_steps == True:
max_name_length = builtins.max(
[len(name) for name in self.training_names.value_list])
for item in self.training_items.values():
item.training_context['execution_id'] = self.execution_id
item.training_context['max_name_length'] = max_name_length
for callback in self.callbacks:
if callback not in item.callbacks:
# private callback
if callback.is_shared == False:
item.with_callbacks(copy.deepcopy(callback))
else:
# shared callback
item.with_callbacks(callback)
# shared callbacks will access training plan dict instead of training_context
for callback in self.callbacks:
if callback.is_shared == True:
callback.on_training_start(self.__dict__)
data_loader = self._dataloaders.value_list[0]
data_loader.minibatch_size = self.minibatch_size
# generate data feed
if not is_resume or only_steps == True:
self.generate_datafeed(data_loader)
if collect_data_inteval == 1 and len(
data_loader.batch_sampler) * self.num_epochs > 1000:
collect_data_inteval = self.default_collect_data_inteval
if only_steps == True:
self.num_epochs = (max_batches //
len(data_loader.batch_sampler)) + 2
for epoch in range(self.num_epochs):
try:
for mbs, return_data in enumerate(data_loader):
if self.is_terminate:
for callback in self.callbacks:
if callback.is_shared == True:
callback.on_training_terminated(
self.__dict__)
for k, trainitem in self.training_items.items():
for callback in trainitem.training_context[
'callbacks']:
if callback.is_shared == False:
callback.on_training_terminated(
trainitem.training_context)
else:
num_batches = len(
data_loader.batch_sampler) * epoch + mbs
iter_data = OrderedDict()
if isinstance(return_data, OrderedDict):
for spec, data in return_data.item_list:
iter_data[spec.name] = data
elif isinstance(return_data, tuple):
for i in range(len(return_data)):
iter_data[
data_loader.traindata.data_template.
key_list[i].name] = return_data[i]
# check weather need out-of-sample evaluation
need_out_sample_evaluation = False
if only_steps == False and self.out_sample_evaluation_on_epoch_end == True and mbs == len(
data_loader.batch_sampler) - 1:
need_out_sample_evaluation = True
elif only_steps == True and self.out_sample_evaluation_on_epoch_end == True and num_batches \
== max_batches - 1:
need_out_sample_evaluation = True
elif only_steps == False and self.out_sample_evaluation_unit == 'batch' and mbs > 0 and mbs % \
self.out_sample_evaluation_frequency == 0:
need_out_sample_evaluation = True
elif only_steps == True and self.out_sample_evaluation_unit == 'batch' and num_batches > 0 \
and num_batches % self.out_sample_evaluation_frequency == 0:
need_out_sample_evaluation = True
elif only_steps == False and self.out_sample_evaluation_unit == 'epoch' and mbs == len(
data_loader.batch_sampler
) - 1 and epoch % self.out_sample_evaluation_frequency == 0:
need_out_sample_evaluation = True
elif only_steps == True and self.out_sample_evaluation_unit == 'epoch' and num_batches == \
max_batches - 1:
need_out_sample_evaluation = True
iter_testdata = None
if isinstance(
data_loader, DataProvider
) and data_loader.testdata is not None and need_out_sample_evaluation:
return_test = data_loader.next_test()
if return_test is not None:
iter_testdata = OrderedDict()
if isinstance(return_test, OrderedDict):
for spec, data in return_test.item_list:
iter_testdata[spec.name] = data
elif isinstance(return_test, tuple):
for i in range(len(return_test)):
iter_testdata[
data_loader.traindata.
data_template.key_list[i].
name] = return_test[i]
# input, target = Variable(input).to(self.device), Variable(target).to(self.device)
for trainitem_name, trainitem in zip(
self.training_names.value_list,
self.training_items.value_list):
train_data = copy.deepcopy(iter_data)
test_data = copy.deepcopy(iter_testdata)
trainitem.training_context[
'model_name'] = trainitem_name
if epoch < int(trainitem.start_epoch):
trainitem.training_context[
'stop_update'] = 1
trainitem.train_model(
train_data,
test_data,
epoch if only_steps == False else 0,
mbs
if only_steps == False else num_batches,
self.num_epochs
if only_steps == False else 1,
len(data_loader.batch_sampler)
if only_steps == False else max_batches,
is_collect_data=mbs %
collect_data_inteval == 0,
is_print_batch_progress=self.
print_progress_unit == 'batch' and
mbs % self.print_progress_frequency == 0,
is_print_epoch_progress=self.
print_progress_unit == 'epoch' and
(epoch + 1) % self.print_progress_frequency
== 0,
log_gradients=keep_gradient_history,
log_weights=keep_weights_history,
accumulate_grads=False,
is_out_sample_evaluation=
need_out_sample_evaluation)
if self.print_progress_unit == 'batch' and mbs % self.print_progress_frequency == 0 and \
self.print_progress_unit == 'epoch' and (epoch + 1) % self.print_progress_frequency == 0:
print('\n', flush=True)
for k, trainitem in self.training_items.items():
for callback in trainitem.training_context[
'callbacks']:
if callback.is_shared == False:
callback.on_overall_batch_end(
trainitem.training_context)
for callback in self.callbacks:
if callback.is_shared == True:
callback.on_overall_batch_end(
self.__dict__)
if self.save_model_frequency > 0 and self.save_model_unit == 'batch' and (num_batches + 1) % \
self.save_model_frequency == 0:
for k, trainitem in self.training_items.items(
):
trainitem.save_model(
trainitem.training_context['save_path']
)
if only_steps == True and num_batches >= max_batches - 1:
for k, trainitem in self.training_items.items(
):
try:
trainitem.save_model()
except Exception as e:
print(e)
return True
if only_steps == False and (mbs + 1) % len(
data_loader.batch_sampler) == 0:
break
except StopIteration:
for k, trainitem in self.training_items.items():
trainitem.do_on_epoch_end()
pass
except ValueError as ve:
print(ve)
PrintException()
for k, trainitem in self.training_items.items():
trainitem.do_on_training_end()
except Exception as e:
print(e)
PrintException()
for k, trainitem in self.training_items.items():
trainitem.do_on_training_end()
if self.save_model_frequency > 0 and self.save_model_unit == 'epoch' and (
epoch + 1) % self.save_model_frequency == 0:
for k, trainitem in self.training_items.items():
trainitem.save_model()
except KeyboardInterrupt:
for k, trainitem in self.training_items.items():
trainitem.save_model()
except Exception as e:
print(e)
PrintException()
for k, trainitem in self.training_items.items():
trainitem.save_model()
def resume(self):
self.start_now(is_resume=True)
def only_steps(self,
num_steps,
collect_data_inteval=1,
keep_weights_history=False,
keep_gradient_history=False):
return self.start_now(collect_data_inteval=collect_data_inteval,
is_resume=False,
only_steps=True,
max_batches=num_steps,
keep_weights_history=keep_weights_history,
keep_gradient_history=keep_gradient_history)
def start_now(self,
collect_data_inteval=1,
is_resume=False,
only_steps=False,
max_batches=np.inf,
keep_weights_history=False,
keep_gradient_history=False):
try:
self.execution_id = get_time_suffix()
exception_cnt = 0
abnormal_num_count = 0
# update callback
if not is_resume or only_steps == True:
max_name_length = builtins.max(
[len(name) for name in self.training_names.value_list])
for item in self.training_items.values():
item.training_context['execution_id'] = self.execution_id
item.training_context['max_name_length'] = max_name_length
for callback in self.callbacks:
if callback not in item.callbacks:
# private callback
if callback.is_shared == False:
item.with_callbacks(copy.deepcopy(callback))
else:
# shared callback
item.with_callbacks(callback)
# shared callbacks will access training plan dict instead of training_context
for callback in self.callbacks:
if callback.is_shared == True:
callback.on_training_start(self.__dict__)
data_loader = self._dataloaders.value_list[0]
data_loader.minibatch_size = self.minibatch_size
# generate data feed
if not is_resume or only_steps == True:
self.generate_datafeed(data_loader)
if collect_data_inteval == 1 and len(
data_loader.batch_sampler) * self.num_epochs > 1000:
collect_data_inteval = self.default_collect_data_inteval
if only_steps == True:
self.num_epochs = (max_batches //
len(data_loader.batch_sampler)) + 2
for epoch in range(self.num_epochs):
try:
for mbs, return_data in enumerate(data_loader):
if self.is_terminate:
for callback in self.callbacks:
if callback.is_shared == True:
callback.on_training_terminated(
self.__dict__)
for k, trainitem in self.training_items.items():
for callback in trainitem.training_context[
'callbacks']:
if callback.is_shared == False:
callback.on_training_terminated(
trainitem.training_context)
else:
num_batches = len(
data_loader.batch_sampler) * epoch + mbs
iter_data = OrderedDict()
if isinstance(return_data, OrderedDict):
for spec, data in return_data.item_list:
iter_data[spec.name] = data
elif isinstance(return_data, tuple):
for i in range(len(return_data)):
iter_data[
data_loader.traindata.data_template.
key_list[i].name] = return_data[i]
# check weather need out-of-sample evaluation
need_out_sample_evaluation = False
if only_steps == False and self.out_sample_evaluation_on_epoch_end == True and mbs == len(
data_loader.batch_sampler) - 1:
need_out_sample_evaluation = True
elif only_steps == True and self.out_sample_evaluation_on_epoch_end == True and num_batches \
== max_batches - 1:
need_out_sample_evaluation = True
elif only_steps == False and self.out_sample_evaluation_unit == 'batch' and mbs > 0 and mbs % \
self.out_sample_evaluation_frequency == 0:
need_out_sample_evaluation = True
elif only_steps == True and self.out_sample_evaluation_unit == 'batch' and num_batches > 0 \
and num_batches % self.out_sample_evaluation_frequency == 0:
need_out_sample_evaluation = True
elif only_steps == False and self.out_sample_evaluation_unit == 'epoch' and mbs == len(
data_loader.batch_sampler
) - 1 and epoch % self.out_sample_evaluation_frequency == 0:
need_out_sample_evaluation = True
elif only_steps == True and self.out_sample_evaluation_unit == 'epoch' and num_batches == \
max_batches - 1:
need_out_sample_evaluation = True
iter_testdata = None
if isinstance(
data_loader, DataProvider
) and data_loader.testdata is not None and need_out_sample_evaluation:
return_test = data_loader.next_test()
if return_test is not None:
iter_testdata = OrderedDict()
if isinstance(return_test, OrderedDict):
for spec, data in return_test.item_list:
iter_testdata[spec.name] = data
elif isinstance(return_test, tuple):
for i in range(len(return_test)):
iter_testdata[
data_loader.traindata.
data_template.key_list[i].
name] = return_test[i]
# input, target = Variable(input).to(self.device), Variable(target).to(self.device)
for trainitem_name, trainitem in zip(
self.training_names.value_list,
self.training_items.value_list):
train_data = copy.deepcopy(iter_data)
test_data = copy.deepcopy(iter_testdata)
trainitem.training_context[
'model_name'] = trainitem_name
if epoch < int(trainitem.start_epoch):
trainitem.training_context[
'stop_update'] = 1
trainitem.train_model(
train_data,
test_data,
epoch if only_steps == False else 0,
mbs
if only_steps == False else num_batches,
self.num_epochs
if only_steps == False else 1,
len(data_loader.batch_sampler)
if only_steps == False else max_batches,
is_collect_data=mbs %
collect_data_inteval == 0,
is_print_batch_progress=self.
print_progress_unit == 'batch' and
mbs % self.print_progress_frequency == 0,
is_print_epoch_progress=self.
print_progress_unit == 'epoch' and
(epoch + 1) % self.print_progress_frequency
== 0,
log_gradients=keep_gradient_history,
log_weights=keep_weights_history,
accumulate_grads=False,
is_out_sample_evaluation=
need_out_sample_evaluation)
if self.print_progress_unit == 'batch' and mbs % self.print_progress_frequency == 0 and \
self.print_progress_unit == 'epoch' and (epoch + 1) % self.print_progress_frequency == 0:
print('\n', flush=True)
for k, trainitem in self.training_items.items():
for callback in trainitem.training_context[
'callbacks']:
if callback.is_shared == False:
callback.on_overall_batch_end(
trainitem.training_context)
for callback in self.callbacks:
if callback.is_shared == True:
callback.on_overall_batch_end(
self.__dict__)
if self.save_model_frequency > 0 and self.save_model_unit == 'batch' and (num_batches + 1) % \
self.save_model_frequency == 0:
for k, trainitem in self.training_items.items(
):
trainitem.save_model(
trainitem.training_context['save_path']
)
if only_steps == True and num_batches >= max_batches - 1:
for k, trainitem in self.training_items.items(
):
try:
trainitem.save_model()
except Exception as e:
print(e)
return True
if only_steps == False and (mbs + 1) % len(
data_loader.batch_sampler) == 0:
break
except StopIteration:
for k, trainitem in self.training_items.items():
trainitem.do_on_epoch_end()
pass
except ValueError as ve:
print(ve)
PrintException()
for k, trainitem in self.training_items.items():
trainitem.do_on_training_end()
except Exception as e:
print(e)
PrintException()
for k, trainitem in self.training_items.items():
trainitem.do_on_training_end()
if self.save_model_frequency > 0 and self.save_model_unit == 'epoch' and (
epoch + 1) % self.save_model_frequency == 0:
for k, trainitem in self.training_items.items():
trainitem.save_model()
except KeyboardInterrupt:
for k, trainitem in self.training_items.items():
trainitem.save_model()
except Exception as e:
print(e)
PrintException()
for k, trainitem in self.training_items.items():
trainitem.save_model()
def generate_datafeed(self, data_loader):
if data_loader.signature is None:
_ = data_loader.next()
# data_input=data_loader.traindata.data.symbol
# if len(data_loader.traindata.unpair)>0:
# data_unpair = data_loader.traindata.unpair
for trainingitem in self.training_items.value_list:
existing_data_feed = None
if 'data_feed' in trainingitem.training_context:
existing_data_feed = trainingitem.training_context['data_feed']
data_feed = OrderedDict()
datasets = data_loader.traindata.get_datasets()
available_items = data_loader.signature.outputs.key_list
available_items.extend(trainingitem.signature.outputs.key_list)
available_items = list(set(available_items))
for inp in trainingitem.signature.inputs.key_list:
data_feed[inp] = None
for k, v in trainingitem._losses.items():
for inp in v.signature.inputs.key_list:
data_feed[inp] = None
for k, v in trainingitem._metrics.items():
for inp in v.signature.inputs.key_list:
data_feed[inp] = None
if 'x' in data_feed and 'input' in available_items:
data_feed['x'] = 'input'
available_items.remove('input')
data_symbols = iteration_tools.flatten(
[data_loader.traindata.data.symbol],
iterable_types=(list, tuple))
label_symbols = iteration_tools.flatten(
[data_loader.traindata.label.symbol],
iterable_types=(list, tuple))
unpair_symbols = iteration_tools.flatten(
[data_loader.traindata.unpair.symbol],
iterable_types=(list, tuple))
if "" in label_symbols:
label_symbols.remove("")
if "" in unpair_symbols:
unpair_symbols.remove("")
if len(trainingitem.signature.inputs) == len(data_symbols) == 1:
if assert_spec_compatibility(
trainingitem.signature.inputs.value_list[0],
data_loader.traindata.data.element_spec):
data_feed[trainingitem.signature.inputs.
key_list[0]] = data_loader.traindata.data.symbol
available_items.remove(data_loader.traindata.data.symbol)
if len(trainingitem.signature.outputs) == len(label_symbols) == 1:
data_feed[trainingitem.signature.outputs.key_list[0].replace(
"output", "target").replace(
"student",
"teacher")] = data_loader.traindata.label.symbol
available_items.remove(data_loader.traindata.label.symbol)
elif len(trainingitem.signature.outputs) == 1 and len(
data_symbols) == 1 and len(label_symbols) == 0:
# autoencoder
data_feed[trainingitem.signature.outputs.key_list[0].replace(
"output", "target").replace(
"student",
"teacher")] = data_loader.traindata.data.symbol
for out in trainingitem.signature.outputs.key_list: # fill the data_feed by key
if out in available_items: # output=putput
data_feed[out] = out
available_items.remove(out)
for key in data_feed.keys():
if data_feed[key] == None and key in available_items:
data_feed[key] = key
available_items.remove(key)
elif data_feed[key] == None:
data_feed[key] = key
# elif out.replace("output","target").replace("student","teacher") in available_items:
# data_feed[out] =out.replace("output","target").replace("student","teacher")
# available_items.remove(out.replace("output","target").replace("student","teacher"))
trainingitem.training_context['data_feed'] = data_feed
print('data_feed for {0} :{1}'.format(trainingitem.name,
data_feed))
