def stochastic_weight_average(self):
experiment_set = self.dataset.experiments[self.train_set_name]
set_name = self.train_set_name.lower()
has_run_average = self.graphs[
set_name].update_stochastic_weighted_average_parameters(
) # noqa: F841
# Think this through - we can probably skip this step but it doesn't harm anything
# if not has_run_average:
# return False
self.graphs[set_name].prepare_for_batchnorm_update()
self.graphs[set_name].train()
experiment_set.reset_epoch()
while True:
Console_UI().inform_user('==>updating batchnorm')
i = 0
for batch in experiment_set:
i += 1
if i % 100 == 1:
Console_UI().inform_user(
f'Updating batchnorm for {self.get_name()}, doing {self.train_set_name} on step {i}'
)
if batch is None:
self.graphs[set_name].finish_batchnorm_update()
return
self.graphs[set_name].update_batchnorm(batch)
def load_pytorch_neural_net(self, neural_net_name: str):
current_run_last_save = self.get_network_full_path(
neural_net_name=neural_net_name, scene_name='last')
if os.path.exists(current_run_last_save):
Console_UI().inform_user(
f'Resuming current runs {neural_net_name:>90}::last network')
return torch.load(current_run_last_save)['state_dict']
if self.resume_prefix is not None:
scene_name = self.global_cfgs.get('resume_scene')
network_filename = self.get_network_filename(
neural_net_name=neural_net_name, scene_name=scene_name)
neural_net_path = os.path.join(self.scenario_log_root,
self.resume_prefix, 'neural_nets',
network_filename)
if os.path.exists(neural_net_path):
Console_UI().inform_user(
f'Resuming from {self.resume_prefix} the network {network_filename}'
)
return torch.load(neural_net_path)['state_dict']
if self.model_zoo_root is not None:
model_zoo_neural_net_path = os.path.join(
self.model_zoo_root, '%s.t7' % (neural_net_name))
if (os.path.exists(model_zoo_neural_net_path)):
Console_UI().inform_user(
f'loading from model_zoo {model_zoo_neural_net_path}')
return torch.load(model_zoo_neural_net_path)['state_dict']
if not self.global_cfgs.get('silent_init_info'):
Console_UI().inform_user(
f'{neural_net_name} does not exist, Initializing from scratch')
return None
def __init__(
self,
neural_net_name,
neural_net_cfgs,
layers,
optimizer_type: str = 'sgd',
input_name: str = '',
output_name: str = '',
input_shape: list = [],
output_shape: list = [],
load_from_batch=True,
add_noise=False,
):
super().__init__()
self.neural_net_name = neural_net_name
self.neural_net_cfgs = neural_net_cfgs
self.input_name = input_name
self.output_name = output_name
self.input_shape = input_shape
self.output_shape = output_shape
self.layers = layers
self.optimizer_type = optimizer_type
self.add_noise = add_noise
self.load_from_batch = load_from_batch
self.weighted_average_parameters = None
self.weighted_average_parameters_counter = 0
self.batch_norm_update_counter = 0
self.momenta = {}
if self.load_from_batch:
self.forward = self.forward_from_batch
else:
self.forward = self.forward_data
if Global_Cfgs().get('DEVICE_BACKEND') == 'cuda':
self.layers.cuda()
self.layers = nn.DataParallel(layers)
self.network_memory_usage = None
try:
self.network_memory_usage = summarizeModelSize(
model=layers,
input_size=(*self.input_shape, ),
device=Global_Cfgs().get('DEVICE_BACKEND'),
)
except Exception as e:
Console_UI().warn_user(
f'Failed to get size for {neural_net_name}: {e}')
pass
Console_UI().debug(self.layers)
self.optimizer = None
self.optimizer = self.get_optimizer()
self.load()
def validate(
self,
iteration_counter,
scene_name: str,
set_name=None,
):
if set_name is None:
set_name = self.val_set_name
if set_name not in self.dataset.experiments:
raise ValueError(f'The set "{set_name}" cannot be found in data')
experiment_set = self.dataset.experiments[set_name]
Console_UI().inform_user(f'Validating {self.get_name()}: {set_name}')
bar = ProgressBar(total=len(experiment_set))
for batch in experiment_set:
if batch is None:
bar.done()
break
bar.current += 1
bar()
batch.update({
'epoch': self.epoch,
# 'graph_name': self.graphs[self.train_set_name.lower()].get_name(),
'graph_name': self.graphs[set_name.lower()].get_name(),
'task_name': self.get_name(),
'iteration_counter': iteration_counter,
})
self.graphs[set_name].eval(batch)
self.end_epoch(set_name, scene_name=scene_name)
def setup_annotations(self):
rel_path = self.get_cfgs('annotations_path')
fm = File_Manager()
self.annotations = fm.read_csv_annotations(
dataset_name=self.dataset_name,
annotations_rel_path=rel_path,
multi_view_per_sample=self.multi_view_per_sample,
)
if self.annotations is None:
annotations_url = self.get_cfgs('annotations_url')
available_csvs_str = '\', \''.join(
fm.get_available_csvs(self.dataset_name))
Console_UI().inform_user(
'"%s" does not exist among the available datasets: \'%s\'.\nDownloading from:\n %s'
% (rel_path, available_csvs_str, annotations_url))
fm.download_zip_file(
url=annotations_url,
dataset_name=self.get_dataset_name(),
)
self.annotations = fm.read_csv_annotations(
dataset_name=self.dataset_name,
annotations_rel_path=rel_path,
multi_view_per_sample=self.multi_view_per_sample,
)
if self.get_cfgs('test_run'):
self.annotations = self.annotations[
self.annotations.index.get_level_values(0) < 100]
def __init__(
self,
annotations_root,
scenario_log_root,
tmp_root,
model_zoo_root,
resume_prefix,
resume_scene,
log_folder,
global_cfgs,
*args,
**kwargs,
):
super().__init__(*args, **kwargs)
self.annotations_root = annotations_root
self.scenario_log_root = scenario_log_root
self.tmp_root = tmp_root
self.model_zoo_root = model_zoo_root
self.resume_prefix = resume_prefix
self.resume_scene = resume_scene
self.log_folder = log_folder
Console_UI().set_log_folder(self.log_folder)
self.log_configs = True
# TODO - set iteration counter on init to last value (this should be saved in a iteration counter txt file)
# We can't use the Singleton pattern here as the Global_Cfgs() imports and initiates File_Manager
self.global_cfgs = global_cfgs
self.__cache = {}
def check_suggested_dictionary(self, modality_name: str,
dictionary: pd.DataFrame,
action_on_missing: str):
modality_name = modality_name.lower()
if modality_name not in self.__values:
return True
if 'name' not in dictionary:
# Multi-bipolar dictionaries should not be checked in this manner
return False
values = pd.Series(self.__values[modality_name])
not_in_dictionary = ~values.isin(dictionary['name'])
if any(not_in_dictionary):
msg = f'Missing values "{values[not_in_dictionary].tolist()}" from the suggested dictionary' +\
f' for {modality_name}'
if action_on_missing == 'exception':
raise IndexError(msg)
elif action_on_missing != 'silent':
Console_UI().inform_user(msg)
return False
return True
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.time_frame = None
self.time_frame_counter = 0
self.travers_order = None
self.init_explicit_modalities()
self.graph = self.init_graph()
self.graph_travers_order = self.get_graph_traverse_order()
self.models = {}
self.init_models_and_adjust_sizes()
self.init_remaining_modalities()
self.losses = {}
self.init_losses()
# Save mermaidjs description of graph to log folder
if not os.path.exists(Global_Cfgs().log_folder):
os.makedirs(Global_Cfgs().log_folder, exist_ok=True)
fn = 'mermaidjs_{ds_name}_{graph_name}_{exp_name}_{scene_name}.txt'\
.format(ds_name=self.get_cfgs('dataset_name'),
graph_name=self.get_name(),
exp_name=self.experiment_name,
scene_name=self.scene_cfgs['name'])
mermain_fn = os.path.join(Global_Cfgs().log_folder, fn)
with open(mermain_fn, 'w') as mermain_file:
mermain_file.write(self.convert_to_mermaidjs())
Console_UI().inform_user(f'Wrote mermaidjs config to {mermain_fn}')
self.exception_counter = 0
def update_learning_rate(self, learning_rate):
if self.optimizer:
if (self.optimizer_type.lower() == 'Adam'.lower()):
if (learning_rate > 1e-3):
Console_UI().warn_user(
f'learning rate {learning_rate:.2e} for Adam is too big.'
+ ' We recommend a learning rate of less than 1e-3')
elif (self.optimizer_type.lower() == 'SGD'.lower()):
if (learning_rate > 1e-1):
Console_UI().warn_user(
f'learning rate {learning_rate:.2e} for Adam is too big.'
+ ' We recommend a learning rate of less than 1e-1')
for param_group in self.optimizer.param_groups:
param_group['lr'] = learning_rate
param_group['weight_decay'] = learning_rate / 200
def analyze_results(self, batch, loss):
self.batch_loss.append(loss)
try:
self.analyze_modality_specific_results(batch)
except Exception as e:
traceback.print_exception(type(e), e, e.__traceback__)
Console_UI().warn_user(f'Failed to get results for {self.modality_name}: {e}')
pass
def get_graph_traverse_order(self):
ordered_nodes = list(nx.topological_sort(self.graph))
try:
ordered_models = [
m for m in ordered_nodes
if self.graph.nodes[m]['node_type'] == 'model'
]
return ordered_models
except KeyError as e:
Console_UI().warn_user(
"You have probably missed a key with node_type:")
Console_UI().warn_user([
m for m in ordered_nodes
if 'node_type' not in self.graph.nodes[m]
])
raise KeyError(f'Key not found: {e}')
return None
def save(self, scene_name):
Console_UI().inform_user(
'\n*****************************************' +
f'\nSave network and losses for {scene_name}')
no_networks = 0
no_losses = 0
for model_name in self.graph_travers_order:
self.models[model_name].save(scene_name)
no_networks += 1
for _, loss in self.losses.items():
loss.save(scene_name)
no_losses += 1
Console_UI().inform_user(
f'Saved {no_networks} networks and {no_losses} losses to disk' +
f' check out: {File_Manager().get_network_dir_path()}' +
'\n*****************************************\n')
def main():
cfgs = Global_Cfgs()
scenario = Scenario(scenario_name=cfgs.get('scenario'))
start_scene = cfgs.get('start_scene')
try:
for scene in iter(scenario):
if start_scene is None or scene.scene_name.strip().lower(
) == start_scene.strip().lower():
start_scene = None
scene.run_scene()
else:
Console_UI().inform_user(
f'Skip \'{scene.scene_name}\' - waiting for \'{start_scene}\''
)
except RuntimeError as error:
Console_UI().warn_user(error)
Console_UI().inform_user("\n\n Traceback: \n")
traceback.print_exc()
except KeyboardInterrupt:
Console_UI().inform_user(
f'\nInterrupted by ctrl+c - stopped ad "{scene.scene_name}"')
else:
Console_UI().inform_user("Done with all scenarios!")
Console_UI().inform_user('To view results, checkout the tensorboard:')
Console_UI().inform_user(
f'tensorboard --logdir /media/max/HD_1_3TB/log/{cfgs.sub_log_path}/tensorboard'
)
def step(self):
if self.optimizer:
try:
self.optimizer.step()
except RuntimeError as e:
Console_UI().warn_user(
f'Failed to optimize {self.get_name()}- a masking issue? {e}'
)
pass
def compute_auc(outputs, targets):
mask = np.logical_or(targets == 1, targets == -1)
if len(mask) < 2 or mask.sum() < 2:
return np.nan
try:
auc = roc_auc_score(y_true=targets[mask] == 1, y_score=outputs[mask])
except IndexError as error:
# TODO: Why is this throwing?
msg = f'IndexError in AUC calculation: {error}'
Console_UI().warn_user(msg)
return np.nan
except ValueError as error:
msg = f'ValueError in AUC calculation: {error}'
Console_UI().warn_user(msg)
return np.nan
return auc
def __init__(self, test_mode=False, test_init_values={}):
if test_mode:
self.cfgs = test_init_values
Console_UI(self.get('log_level', 'warning'), globalConfigs=self)
self.sub_log_path = self.get('sub_log_path', 'sub_log_not_set')
File_Manager(annotations_root=self.get('annotation_root'),
log_folder=self.get('log_folder'),
scenario_log_root=self.get('scenario_log_root'),
resume_prefix=self.get('resume'),
resume_scene=self.get('resume_scene'),
tmp_root=self.get('tmp_root'),
model_zoo_root=self.get('model_zoo_root'),
global_cfgs=self)
return
args = self.parse_argument()
self.cfgs = args.__dict__
Console_UI(self.get('log_level', 'info'), globalConfigs=self)
self.read_environment_variables()
self.start_time = datetime.now(
pytz.timezone('Europe/Stockholm')).strftime('%Y%m%d/%H.%M')
self.sub_log_path = os.path.join(self.get('scenario'), self.start_time)
if self.get('resume') is not None:
self.prep_resume()
fm = File_Manager(scenario_log_root=self.scenario_log_root,
log_folder=self.log_folder,
annotations_root=self.get('annotation_root'),
resume_prefix=self.get('resume'),
resume_scene=self.get('resume_scene'),
tmp_root=self.get('tmp_root'),
model_zoo_root=self.get('model_zoo_root'),
global_cfgs=self)
setup_data = {'call': ' '.join(sys.argv), 'setup': self.cfgs}
fm.log_setup(data=setup_data, name='base_setup.yaml')
self.__forward_noise = 0
def get_cfgs(self, name, default=None):
try:
if name in self.graph_cfgs:
return self.graph_cfgs[name]
if name in self.task_cfgs:
return self.task_cfgs[name]
if name in self.task_cfgs['apply']:
return self.task_cfgs['apply'][name]
if name in self.scene_cfgs:
return self.scene_cfgs[name]
if name in self.scenario_cfgs:
return self.scenario_cfgs[name]
except TypeError as e:
Console_UI().inform_user(self.graph_cfgs)
Console_UI().inform_user(self.task_cfgs)
Console_UI().inform_user(self.scene_cfgs)
Console_UI().inform_user(self.scenario_cfgs)
raise TypeError(f'Error during {self.get_name()}: {e}')
return Global_Cfgs().get(name)
def eval(self, batch=None):
for model_name in self.graph_travers_order:
self.models[model_name].eval()
for _, loss in self.losses.items():
loss.eval()
if batch is not None:
self.encode(batch)
if self.reconstruction:
self.decode(batch)
# Only in evaluation phase we visualize the reconstructed image (if exist)
Console_UI().add_last_reconstructed_input(batch)
self.compute_loss(batch)
def download_zip_file(self, url, dataset_name):
"""
annotations has to be zipped with the following command:
7z a -ppassword -mem=ZipCrypto imagenet.zip imagenet
"""
url_content = urlopen(url)
zipfile = ZipFile(BytesIO(url_content.read()))
pswd = Console_UI().receive_password(
'Password for unzipping annotations of %s dataset:' %
(dataset_name))
zipfile.extractall(self.annotations_root, pwd=bytes(pswd, 'utf-8'))
def read_csv_annotations(
self,
dataset_name: str,
annotations_rel_path: str,
multi_view_per_sample: bool = False,
):
annotations_path = os.path.join(
self.get_annotations_path(dataset_name), annotations_rel_path)
if os.path.exists(annotations_path):
cache_path = f'csv:{annotations_path}'
if cache_path in self.__cache:
annotation = self.__cache[cache_path]
else:
annotation = pd.read_csv(annotations_path, low_memory=False)
self.__cache[cache_path] = annotation
if multi_view_per_sample and not isinstance(
annotation.index, pd.MultiIndex):
if 'index' not in annotation:
annotation['index'] = np.arange(len(annotation), dtype=int)
else:
assert np.issubdtype(
annotation['index'],
np.dtype(int)), 'Index should be integers'
assert annotation['index'].min(
) == 0, 'The index has to be indexed from 0'
if 'sub_index' not in annotation:
annotation['sub_index'] = np.zeros(len(annotation),
dtype=int)
else:
assert np.issubdtype(
annotation['sub_index'],
np.dtype(int)), 'Sub index should be integers'
assert annotation['sub_index'].max(
) > 0, 'You have provided a sub_index without purpose (max 0)'
assert annotation['sub_index'].min(
) == 0, 'The sub_index has to start from 0'
annotation.set_index(['index', 'sub_index'], inplace=True)
if 'num_views' not in annotation:
annotation['num_views'] =\
[a for b in [[i] * i for i in annotation.groupby(level=0).size()] for a in b]
return annotation
Console_UI().warn_user(
f'Failed to load file from disk: \'{annotations_path}\'')
return None
def prep_resume(self):
ui = Console_UI()
resume_prefix = self.get('resume')
resume_scene = self.get('resume_scene')
if resume_scene is not None and resume_prefix is None:
raise ValueError(
'You must provide resume prefix if you have set a resume scene'
)
# for debug mode uncomment:
# scenario_log_root = "/media/max/SSD_1TB/log/"
if resume_prefix.lower() == 'last':
dirs = sorted([
d for d in iglob(f'{self.scenario_log_root}/*/*/neural_nets')
])
dirs = [
d for d in dirs
if len([f for f in iglob(f'{d}/*{resume_scene}.t7')]) > 0
]
if len(dirs) == 0:
raise Exception(
f'No previous runs found in \'{self.scenario_log_root}\' with *{resume_scene}.t7'
)
resume_prefix = dirs[-1].lstrip(
self.scenario_log_root).rstrip('/neural_nets')
ui.inform_user(f'Resuming run from {resume_prefix}')
elif resume_prefix is not None:
resume_prefix = retrieve_dir(path=resume_prefix,
base_path=self.scenario_log_root,
expected_depth=1)
ui.inform_user(f'Resuming run from {resume_prefix}')
self.cfgs['resume'] = resume_prefix
# for debug mode uncomment:
# self.cfgs['resume'] = "../%s" % self.cfgs['resume']
if not self.cfgs['skip_tensorboard']:
dst_tensorboard_path = os.path.join(self.log_folder, 'tensorboard')
if os.path.exists(dst_tensorboard_path):
ui.inform_user(
f'Removing previous tensorboard catalogue: {dst_tensorboard_path}'
)
shutil.rmtree(dst_tensorboard_path)
ui.inform_user('Copying the previous tensorboard data')
shutil.copytree(
src=os.path.join(self.scenario_log_root, resume_prefix,
'tensorboard'),
dst=dst_tensorboard_path,
)
def init_modality(self, modality_name: str, modality_cfgs: dict = None):
modality_name = modality_name.lower()
assert (modality_cfgs is not None
), 'modality_cfgs should not be None in %s' % (modality_name)
start_time = time.time()
Modality, content, dictionary = get_modality_and_content(
annotations=self.annotations,
modality_name=modality_name,
modality_cfgs=modality_cfgs,
ignore_index=
-100 # The -100 is defined in the loss_cfgs and not available here :-(
)
modality = Modality(
dataset_name=self.dataset_name,
dataset_cfgs=self.dataset_cfgs,
experiment_name=self.experiment_name,
experiment_cfgs=self.experiment_cfgs,
modality_name=modality_name,
modality_cfgs=modality_cfgs,
content=content,
dictionary=dictionary,
)
if modality.is_explicit_modality():
self.explicit_modalities[modality_name] = modality
if modality.is_input_modality():
self.explicit_input_modalities[modality_name] = modality
elif modality.is_output_modality():
self.explicit_output_modalities[modality_name] = modality
else:
raise BaseException(
'Explicit Modalities should either be input or output')
elif modality.is_implicit_modality():
self.implicit_modalities[modality_name] = modality
# Add explicit and implicit modalities
# Todo - Ali: why do we need to have this split? When do we have the case were a modality is neither
self.modalities.update(self.explicit_modalities)
self.modalities.update(self.implicit_modalities)
if not Global_Cfgs().get('silent_init_info'):
Console_UI().inform_user(
info='Initializing %s modality in %s in %d milliseconds' %
(modality_name, self.get_name(), 1000 *
(time.time() - start_time)),
debug=(modality_cfgs),
)
def end_epoch(self, experiment_name: str, scene_name: str):
experiment_set = self.dataset.experiments[experiment_name]
summary = {
'epoch': self.epoch,
'graph_name': self.graph_name,
'task_name': self.get_name()
}
if experiment_name == self.train_set_name:
summary['epoch_size'] = len(experiment_set)
self.save(scene_name='last')
self.epoch += 1
experiment_set.end_epoch(summary=summary, scene_name=scene_name)
Console_UI().add_epoch_results(summary)
def __next__(self):
batch = None
i = 0
iterator = self.get_iterator()
# start_time = time.time()
while (batch is None and i < 5):
try:
batch = next(iterator)
if batch['encoder_image'].max() == 0:
raise ValueError(
'No non-zero images in batch - check file folder')
# Convert all tensors to cuda if environment calls for it
for key in batch:
batch[key] = convert_2_cuda(batch[key])
batch_size_info = f'{self.bin_weights[self.batch_index]} ({len(self.bins[self.batch_index])} bins)'
batch.update({
'batch_index': self.batch_index,
'epoch_size': len(self),
'batch_size': batch_size_info,
})
except StopIteration:
self.reset_epoch()
return None
except Exception as ex:
batch = None
Console_UI().warn_user(f'Failed to load batch: "{ex}"')
traceback.print_exception(type(ex), ex, ex.__traceback__)
self.batch_index += 1
i += 1
if i >= 5:
raise Exception(
'Failed multiple times when trying to retrieve batch')
# Check images - seem ok
# import cv2
# for i in range(batch['encoder_image'].shape[0]):
# for ii in range(batch['encoder_image'].shape[1]):
# img = batch['encoder_image'][i, ii, 0] * 255
# cv2.imwrite(f'/home/max/tmp/test{i}_{ii}.png', img.reshape(256, 256, 1).cpu().numpy())
# Profile time for loading batch - not much is gained by more than two workers (time is 0.01 to 0.1 seconds)
# print("Time spent time retrieving batch: %0.2f" % (time.time() - start_time))
return batch
def train(self, batch=None):
for model_name in self.graph_travers_order:
self.models[model_name].train()
for _, loss in self.losses.items():
loss.train()
if batch is not None:
try:
self.zero_grad()
self.encode(batch)
if self.reconstruction:
self.decode(batch)
loss = self.compute_loss(batch)
start_time = time.time()
if loss > 0:
loss.backward()
self.step()
batch['time']['backward'][self.get_name()] = {
'start': start_time,
'end': time.time()
}
self.collect_runtime_stats(batch)
self.exception_counter = 0
return True
except KeyError as e:
Console_UI().warn_user(f'Could not find {e} in:')
Console_UI().warn_user(sorted(batch.keys()))
Console_UI().inform_user("\n\n Traceback: \n")
traceback.print_exc()
raise e
except Exception as e:
Console_UI().warn_user(
f'** Error while training batch in {self.get_name()} **')
Console_UI().warn_user(
f'Indices: {batch["indices"]} and encoder image shape {batch["encoder_image"].shape}'
)
Console_UI().warn_user(f'Error message: {e}')
Console_UI().inform_user("\n\n Traceback: \n")
traceback.print_exc()
self.exception_counter += 1
if self.exception_counter > 5:
raise RuntimeError(f'Error during training: {e}')
return False
def step(
self,
iteration_counter: int,
scene_name: str,
):
# Due to the parallel nature of loading the data we need to reset
# the start time for the batch in order to get the true processing
start_time = time.time()
experiment_set = self.dataset.experiments[self.train_set_name]
batch = next(experiment_set)
if batch is None:
self.end_epoch(experiment_name=self.train_set_name,
scene_name=scene_name)
if ((self.epoch < 5 and iteration_counter < 1e3)
or self.epoch % self.validate_when_epoch_is_devisable_by
== 0):
self.validate(iteration_counter, scene_name=scene_name)
batch = next(experiment_set)
if batch is None:
raise Exception('The next batch after resetting was empty!?')
batch['time']['start'] = start_time
batch.update({
'epoch':
self.epoch,
'graph_name':
self.graphs[self.train_set_name.lower()].get_name(),
'task_name':
self.get_name(),
'iteration_counter':
iteration_counter,
})
success = self.graphs[self.train_set_name.lower()].train(batch)
if not success:
return False
Console_UI().add_batch_results(batch)
return True
def read_dictionary(
self,
dataset_name: str,
modality_name: str,
):
"""
If we have a dictionary associated with the current weights we should use those.
The fallback is the resume weight's dictionary and lastly the annotation's dictionary.
"""
filename = f'{modality_name.lower()}_dictionary.csv'
cachename = f'dictionary:{dataset_name}->{filename}'
if cachename in self.__cache:
return self.__cache[cachename]
dictionary_path = os.path.join(self.log_folder, 'neural_nets',
filename)
if (not os.path.exists(dictionary_path)
and self.resume_prefix is not None):
dictionary_path = os.path.join(self.scenario_log_root,
self.resume_prefix, 'neural_nets',
filename)
if not os.path.exists(dictionary_path):
dictionary_path = os.path.join(
self.get_annotations_path(dataset_name), filename)
if os.path.exists(dictionary_path):
try:
dictionary = pd.read_csv(dictionary_path)
self.__cache[cachename] = dictionary
return dictionary
except pd.errors.EmptyDataError:
Console_UI().warn_user(
f'The dictionary for {modality_name} is corrupt - see file {dictionary_path}'
)
return None
def collect_dictionaries(self):
"""
Check all the Datasets for common items, e.g. body part and then create
a general dictionary for all of them.
"""
datasets = []
for scene in self.scenario_cfgs['scenes']:
for task in scene['tasks'].values():
if task['dataset_name'] not in datasets:
datasets.append(task['dataset_name'])
configs = {}
for dataset_name in datasets:
configs[dataset_name] = File_Manager().read_dataset_config(
dataset_name)
modalities_with_dictionaries = [
'one_vs_rest',
'bipolar',
'multi_bipolar',
] # TODO: add 'hierarchical_label' but this has some fancy logic :-S
dictionary_candidates = []
for dataset_name in datasets:
config = configs[dataset_name]
try:
for experiment in config['experiments'].values():
if isinstance(experiment['modalities'], dict):
[
dictionary_candidates.append(name)
for name, cfg in experiment['modalities'].items()
if
cfg['type'].lower() in modalities_with_dictionaries
and name not in dictionary_candidates
]
except Exception as e:
raise Exception(
f'Failed to get dictionary for {dataset_name}: {e}')
# Store all the different values available for this modality into the dictionary singleton that
# keeps track of the unique values
dg = Dictionary_Generator()
for modality_name in dictionary_candidates:
for dataset_name in datasets:
dg.append_suggested_dictionary(dataset_name=dataset_name,
modality_name=modality_name)
config = configs[dataset_name]
for experiment in config['experiments'].values():
annotations = File_Manager().read_csv_annotations(
dataset_name,
annotations_rel_path=experiment['annotations_path'],
# Multi-view argument should be irrelevant for this
)
if annotations is None:
raise ValueError(
f'Could not find the dataset: {dataset_name} in {experiment["annotations_path"]}'
)
modalities = experiment['modalities']
if modalities == 'same_as_train_set':
modalities = config['experiments']['train_set'][
'modalities']
if modality_name in modalities:
if 'column_name' in modalities[modality_name]:
try:
colname = modalities[modality_name][
'column_name']
dg.append_values(modality_name=modality_name,
values=annotations[colname])
except KeyError as e:
Console_UI().warn_user(
f'Got a key annotation exception for {colname}'
)
Console_UI().warn_user(
modalities[modality_name])
Console_UI().warn_user(annotations.columns)
raise e
except Exception as e:
Console_UI().warn_user(
f'Got an annotation exception for {colname}'
)
Console_UI().warn_user(
modalities[modality_name])
Console_UI().warn_user(annotations)
raise e
elif 'columns' in modalities[modality_name]:
for column_name in modalities[modality_name][
'columns']:
if isinstance(column_name, dict):
assert 'csv_name' in column_name, \
f'The column doesn\'t have the expected csv_name element, got: {column_name}'
column_name = column_name['csv_name']
if column_name not in annotations:
n = 3 if len(
annotations.columns) < 10 else ceil(
len(annotations.columns) / 3)
closest = get_close_matches(
word=column_name,
possibilities=annotations.columns,
n=n,
)
closest = ', '.join(closest)
raise IndexError(
f'The {column_name} from {modality_name} doesn\'t exist.'
+ f' Closest matching are: {closest}')
dg.append_values(
modality_name=modality_name,
values=annotations[column_name])
else:
raise IndexError(
f'Expected {modality_name} to have either columns or column_name defined'
)
def closing_credits(self):
Console_UI().inform_user("That's it folks")
def __init__(
self,
graph_name,
experiment_set,
task_cfgs,
scene_cfgs,
scenario_cfgs,
):
self.graph_name = graph_name
self.task_cfgs = task_cfgs
self.scene_cfgs = scene_cfgs
self.scenario_cfgs = scenario_cfgs
self.experiment_set = experiment_set
self.experiment_name = self.experiment_set.get_name()
self.graph_cfgs = self.get_graph_cfgs(self.graph_name)
self.classification = self.get_cfgs('classification', default=False)
self.reconstruction = self.get_cfgs('reconstruction', default=False)
self.identification = self.get_cfgs('identification', default=False)
self.regression = self.get_cfgs('regression', default=False)
self.pi_model = self.get_cfgs('pi_model', default=False)
self.real_fake = self.get_cfgs('real_fake', default=False)
self.optimizer_type = self.get_cfgs('optimizer_type')
if not Global_Cfgs().get('silent_init_info'):
UI = Console_UI()
UI.inform_user(
info=[
'explicit experiment modalities',
list(self.get_experiment_explicit_modalities().keys())
],
debug=self.get_experiment_explicit_modalities(),
)
UI.inform_user(
info=[
'implicit experiment modalities',
list(self.get_experiment_implicit_modalities().keys())
],
debug=self.get_experiment_implicit_modalities(),
)
UI.inform_user(
info=[
'explicit graph modalities',
list(self.get_graph_specific_explicit_modalities().keys())
],
debug=self.get_graph_specific_explicit_modalities(),
)
UI.inform_user(
info=[
'implicit graph modalities',
list(self.get_graph_specific_implicit_modalities().keys())
],
debug=self.get_graph_specific_implicit_modalities(),
)
UI.inform_user(
info=[
'explicit models',
list(self.get_explicit_models().keys())
],
debug=self.get_explicit_models(),
)
UI.inform_user(
info=[
'implicit models',
list(self.get_implicit_models().keys())
],
debug=self.get_implicit_models(),
)