def __init__(self,
model,
period,
time_reference,
dirName,
sample_indices_by_dataset={VALIDATION: []},
hop_legth_cqt=128,
dataset_keys=[TRAIN, VALIDATION],
target_dim=1):
self._dir_name = dirName + '/pca_latent'
super().__init__(model,
period,
time_reference,
dataset_keys=dataset_keys,
hop_legth_cqt=hop_legth_cqt,
dirName=self._dir_name)
self.model_class_name = type(model).__name__
self._sample_indices_by_dataset = sample_indices_by_dataset
check_dataset_keys_not_loop(list(sample_indices_by_dataset.keys()))
tf_logging.info("Create WavLatentPCAHook for: \n" + \
"\n".join([ds_key + ": " + ", ".join(map(str, idxs)) \
for ds_key, idxs in sample_indices_by_dataset.items()]))
def load_images(self, session):
check_dataset_keys_not_loop(list(self._images_indexes.keys()))
images = {ds_key: (index_list, self._model.dataset.get_elements(self._model.x, self._ds_handle,
self._ds_handles[ds_key],
self._ds_initializers[ds_key], session,
index_list)) \
for (ds_key, index_list) in self._images_indexes.items()}
self._images = images
def load_labels(self, session):
if self._conditional and self._labels is None:
check_dataset_keys_not_loop(list(self._images_indexes.keys()))
labels = {ds_key: (index_list, self._model.dataset.get_elements(self._model.y, self._ds_handle,
self._ds_handles[ds_key],
self._ds_initializers[ds_key], session,
index_list)) \
for (ds_key, index_list) in self._images_indexes.items()}
self._labels = labels
def __init__(
self,
model,
period,
time_reference,
dirName,
sample_indices_by_dataset={VALIDATION: []},
fast_gen=True, # use fast_generation wavenet for reconstruction without teacher forcing
debug_fast_gen=True,
# use fast_generation wavenet with the true input shifted and quantized to reconstruct with teacher forcing and check the FastGen network
hop_legth_cqt=128,
dataset_keys=[TRAIN, VALIDATION],
save_wav=False,
compute_reconstruction_metrics=True,
_plot=True,
generate_from_mean=True,
spider_plot_time_splits=None,
anomaly_detection_params=None):
super().__init__(model,
period,
time_reference,
dataset_keys=dataset_keys,
hop_legth_cqt=hop_legth_cqt,
dirName=dirName)
self.model_class_name = type(model).__name__
self.save_wav = save_wav
self.compute_reconstruction_metrics = compute_reconstruction_metrics
self._plot = _plot
self.generate_from_mean = generate_from_mean
self.spider_plot_time_splits = spider_plot_time_splits
self.anomaly_detection_params = anomaly_detection_params
if compute_reconstruction_metrics:
self.reconstr_metrics_file_names = {
TRAIN: dirName + '/reconstr_metrics_x_train.txt',
VALIDATION: dirName + '/reconstr_metrics_x_validation.txt',
TEST: dirName + '/reconstr_metrics_x_test.txt',
}
self._sample_indices_by_dataset = sample_indices_by_dataset
self._fast_gen = bool(fast_gen)
self._debug_fast_gen = bool(debug_fast_gen)
check_dataset_keys_not_loop(list(sample_indices_by_dataset.keys()))
tf_logging.info("Create WavGenerateHook for: \n" + \
"\n".join([ds_key + ": " + ", ".join(map(str, idxs or ['all'])) \
for ds_key, idxs in sample_indices_by_dataset.items()]))
def load_masks(self, session):
if self._masks == None and self._model.mask is not None:
check_dataset_keys_not_loop(list(self._images_indexes.keys()))
masks = {ds_key : (index_list, self._model.dataset.get_elements(self._model.mask, self._ds_handle, self._ds_handles[ds_key], self._ds_initializers[ds_key], session, index_list)) \
for (ds_key, index_list) in self._images_indexes.items()}
# I set something like the following structure, e.g.
# images = {TRAIN : ([0,100,200,300,400,500], train_images),
# VALIDATION : ([0,100,200,300], validation_images),
# },
self._masks = masks
def load_images_once(self, session):
if self._images==None:
check_dataset_keys_not_loop(list(self._images_indexes.keys()))
images = {ds_key : (couple_indices_list, [(img1, img2) for img1, img2 in zip(
self._model.dataset.get_elements(self._model.x, self._ds_handle, self._ds_handles[ds_key], self._ds_initializers[ds_key], session, [i[0] for i in couple_indices_list]),
self._model.dataset.get_elements(self._model.x, self._ds_handle, self._ds_handles[ds_key], self._ds_initializers[ds_key], session, [i[1] for i in couple_indices_list]))]) \
for (ds_key, couple_indices_list) in self._images_indexes.items()}
# I set something like the following structure, e.g.
# images = {TRAIN : ([(0,50),(100,230),(200,790),(300,600),(400,1000),(500,10)], list_of_couples_train_images,
# VALIDATION : ([(0,50),(100,230),(200,790),(300,600),(400,1000),(500,10)], list_of_couples_validation_images,
# },
self._images = images
def __init__(self,
model,
period,
time_reference,
datasets_keys,
plot_offset,
tensorboard_dir=None,
trigger_summaries=False,
extra_feed_dict={}):
time_choices = [EPOCHS, STEPS]
if not time_reference in time_choices:
raise ValueError("time reference attribute can be only in %s" %
time_choices)
self._timer = tf.train.SecondOrStepTimer(every_secs=None,
every_steps=period)
self._time_reference_str = time_reference
self._time_ref_shortstr = self._time_reference_str[:2]
self._model = model
self._plot_offset = plot_offset
self._extra_feed_dict = extra_feed_dict
# called in before_run
self._nodes_to_be_computed_by_run = {}
check_dataset_keys_not_loop(datasets_keys)
self._datasets_keys = datasets_keys
self._ds_initializers = model.datasets_initializers
self._ds_handles_nodes = model.datasets_handles_nodes
self._ds_handle = model.ds_handle
# these needs to be defined in the child class
self._tensors_names = None
self._tensors_plots = None
self._tensors_values = None
self._trigger_summaries = trigger_summaries
self._tensorboard_dir = tensorboard_dir
assert not self._trigger_summaries or self._tensorboard_dir is not None, \
"If you specified that you want to Trigger Summeries, you should also specify where to save them."
self.SUMMARIES_KEY = "log_mean_summaries"
self._default_plot_bool = True
def __init__(self,
model,
period,
time_reference,
dirName,
sample_indices_by_dataset={VALIDATION: []},
hop_legth_cqt=128,
dataset_keys=[TRAIN, VALIDATION],
save_wav=False,
reconstruct_from_mean=True,
batch_size=20,
_plot=True,
spider_plot_time_splits=None,
anomaly_detection_params=None,
compute_reconstruction_metrics=True):
super().__init__(model,
period,
time_reference,
dataset_keys=dataset_keys,
hop_legth_cqt=hop_legth_cqt,
dirName=dirName)
self._plot = _plot
self.reconstruct_from_mean = reconstruct_from_mean
self.save_wav = save_wav
self.batch_size = batch_size
self.spider_plot_time_splits = spider_plot_time_splits
self._sample_indices_by_dataset = sample_indices_by_dataset
self.compute_reconstruction_metrics = compute_reconstruction_metrics
self.anomaly_detection_params = anomaly_detection_params
if compute_reconstruction_metrics:
self.reconstr_metrics_file_names = {
TRAIN: dirName + '/reconstr_metrics_tf_train.txt',
VALIDATION: dirName + '/reconstr_metrics_tf_validation.txt',
TEST: dirName + '/reconstr_metrics_tf_test.txt',
}
check_dataset_keys_not_loop(list(sample_indices_by_dataset.keys()))
tf_logging.info("Create WavReconstructHook for: \n" + \
"\n".join([ds_key + ": " + ", ".join(map(str, idxs or ['all'])) \
for ds_key, idxs in sample_indices_by_dataset.items()]))
def __init__(self,
model,
period,
time_reference,
dirName,
sample_indices_by_dataset={
VALIDATION: []},
hop_legth_cqt=128,
dataset_keys=[TRAIN, VALIDATION]
):
_dirName = dirName + '/mean_variance_plots'
super().__init__(model, period, time_reference, dataset_keys=dataset_keys, hop_legth_cqt=hop_legth_cqt,
dirName=_dirName)
self._sample_indices_by_dataset = sample_indices_by_dataset
check_dataset_keys_not_loop(list(sample_indices_by_dataset.keys()))
tf_logging.info("Create WavenetGaussianVisualizationHook for: \n"
+ "\n".join([ds_key + ": " + ", ".join(map(str, idxs))
for ds_key, idxs in sample_indices_by_dataset.items()]))
