def main(hparams, cluster=None, results_dict=None):
"""
Main training routine specific for this project
:param hparams:
:return:
"""
# init experiment
log_dir = os.path.dirname(os.path.realpath(__file__))
exp = Experiment(
name='test_tube_exp',
debug=True,
save_dir=log_dir,
version=0,
autosave=False,
description='test demo'
)
hparams.training_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/training'
hparams.validation_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/validation'
hparams.test_task_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/test_task'
hparams.batch_size = 4
# set the hparams for the experiment
exp.argparse(hparams)
exp.save()
# build model
model = ContrastivePredictiveSystem(hparams)
# callbacks
early_stop = EarlyStopping(
monitor=hparams.early_stop_metric,
patience=hparams.early_stop_patience,
verbose=True,
mode=hparams.early_stop_mode
)
model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
checkpoint = ModelCheckpoint(
filepath=model_save_path,
save_best_only=True,
verbose=True,
monitor=hparams.model_save_monitor_value,
mode=hparams.model_save_monitor_mode
)
# configure trainer
trainer = Trainer(
experiment=exp,
checkpoint_callback=checkpoint,
early_stop_callback=early_stop,
# distributed_backend='dp',
#gpus=[0],
nb_sanity_val_steps=2
)
# train model
trainer.fit(model)
def test_classification_task_remote(self):
weights_path = '/home/idivinci3005/experiments/checkpoints/_ckpt_epoch_20.ckpt'
tags_path = '/home/idivinci3005/experiments/logs/immersions_scalogram_resnet/version_7/meta_tags.csv'
task_dataset_path = '/home/idivinci3005/data/immersions/small_test_task'
cpc_system = ContrastivePredictiveSystem.load_from_metrics(
weights_path, tags_path, on_gpu=False)
cpc_system.hparams.batch_size = 4
task_system = ClassificationTaskModel(cpc_system, task_dataset_path)
task_system.calculate_data()
trainer = pl.Trainer()
trainer.fit(task_system)
def test_preprocessing_noise(self):
dataset = MaestroDataset('/Volumes/Elements/Datasets/maestro-v2.0.0',
item_length=176400,
sampling_rate=44100,
mode='validation',
max_file_count=20,
shuffle_with_seed=123)
parser = HyperOptArgumentParser(strategy='random_search',
add_help=False)
#parser = ArgumentParser()
parser = ContrastivePredictiveSystem.add_model_specific_args(
parser, root_dir='../')
hparams = parser.parse_args()
preprocessing = PreprocessingModule(hparams)
pass
def load_model(self, weights_path, tags_path, model_shapes_path, ranges_path):
if self.dev != 'cpu' and torch.cuda.is_available():
self.model = ContrastivePredictiveSystem.load_from_metrics(weights_path, tags_path)
self.model.model = self.model.model.module
self.model.preprocessing = self.model.preprocessing.module
self.model.to(self.dev)
else:
self.model = ContrastivePredictiveSystem.load_from_metrics(weights_path, tags_path)
self.model.model = self.model.model.module
self.model.preprocessing = self.model.preprocessing.module
self.model.cpu()
self.model.freeze()
self.model.eval()
self.register = self.model.activation_register
self.register.active = True
print("prediction_model weight", self.model.model.prediction_model.weight)
pprint.pprint(self.model)
self.activation_normalization = ActivationNormalization(self.mean_statistics, self.std_statistics,
var_epsilon=1e-1, norm_epsilon=1e-2)
self.model_activation = ModelActivations(shapes_path=model_shapes_path,
ignore_time_dimension=False,
remove_results=True)
with open(ranges_path, 'rb') as handle:
self.activation_ranges = pickle.load(handle)
self.high_freq_filter = torch.linspace(-1., 1., steps=self.model_activation.shapes['scalogram'][1])
self.high_freq_filter = torch.clamp(self.high_freq_filter, 0., 1.)[None, :, None]
if self.dev != 'cpu' and torch.cuda.is_available():
self.activation_normalization.cuda()
self.high_freq_filter = self.high_freq_filter.cuda()
else:
self.activation_normalization.cpu()
self.high_freq_filter.cpu()
def main(hparams, cluster=None, results_dict=None):
"""
Main training routine specific for this project
:param hparams:
:return:
"""
# init experiment
log_dir = os.path.dirname(os.path.realpath(__file__))
exp = Experiment(name='test_tube_exp',
debug=True,
save_dir=log_dir,
version=0,
autosave=False,
description='test demo')
hparams.training_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/training'
hparams.validation_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/validation'
hparams.test_task_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/test_task'
hparams.dummy_datasets = False
hparams.audio_noise = 3e-3
hparams.cqt_fmin = 40.
hparams.cqt_bins_per_octave = 24
hparams.cqt_n_bins = 216
hparams.cqt_hop_length = 512
hparams.cqt_filter_scale = 0.43
hparams.enc_channels = (1, 8, 16, 32, 64, 128, 256, 512, 512)
hparams.enc_kernel_1_w = (3, 3, 3, 3, 3, 3, 3, 3)
hparams.enc_kernel_1_h = (3, 3, 3, 3, 3, 3, 3, 3)
hparams.enc_kernel_2_w = (1, 3, 1, 3, 1, 3, 1, 3)
hparams.enc_kernel_2_h = (25, 3, 25, 3, 25, 3, 4, 3)
hparams.enc_padding_1 = (1, 1, 1, 1, 1, 1, 1, 1)
hparams.enc_padding_2 = (0, 1, 0, 1, 0, 1, 0, 0)
hparams.enc_stride_1 = (1, 1, 1, 1, 1, 1, 1, 1)
hparams.enc_stride_2 = (1, 1, 1, 1, 1, 1, 1, 1)
hparams.enc_pooling_1 = (2, 1, 1, 1, 2, 1, 1, 1)
hparams.ar_kernel_sizes = (5, 4, 1, 3, 3, 1, 3, 1, 6)
hparams.ar_self_attention = (False, False, False, False, False, False,
False, False, False)
hparams.batch_size = 4
hparams.learning_rate = 3e-4
hparams.warmup_steps = 1000
hparams.annealing_steps = 100000
hparams.score_over_all_timesteps = False
hparams.visible_steps = 60
# set the hparams for the experiment
exp.argparse(hparams)
exp.save()
# build model
model = ContrastivePredictiveSystem(hparams)
# callbacks
early_stop = EarlyStopping(monitor=hparams.early_stop_metric,
patience=hparams.early_stop_patience,
verbose=True,
mode=hparams.early_stop_mode)
model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name,
exp.version)
checkpoint = ModelCheckpoint(filepath=model_save_path,
save_best_only=True,
verbose=True,
monitor=hparams.model_save_monitor_value,
mode=hparams.model_save_monitor_mode)
# configure trainer
trainer = Trainer(
experiment=exp,
checkpoint_callback=checkpoint,
#early_stop_callback=early_stop,
# distributed_backend='dp',
#gpus=[0],
nb_sanity_val_steps=2,
gradient_clip=0.5)
# train model
trainer.fit(model)
# configure trainer
trainer = Trainer(
experiment=exp,
checkpoint_callback=checkpoint,
#early_stop_callback=early_stop,
# distributed_backend='dp',
#gpus=[0],
nb_sanity_val_steps=2,
gradient_clip=0.5)
# train model
trainer.fit(model)
if __name__ == '__main__':
# use default args given by lightning
root_dir = os.path.split(os.path.dirname(
sys.modules['__main__'].__file__))[0]
parent_parser = HyperOptArgumentParser(strategy='random_search',
add_help=False)
add_default_args(parent_parser, root_dir)
# allow model to overwrite or extend args
parser = ContrastivePredictiveSystem.add_model_specific_args(
parent_parser, root_dir)
hyperparams = parser.parse_args()
# train model
main(hyperparams)
import torch
import pickle
from immersions.cpc_system import ContrastivePredictiveSystem
from immersions.input_optimization.activation_utilities import ActivationStatistics
weights_path = '/Volumes/Elements/Projekte/Immersions/checkpoints/immersions_scalogram_resnet/_ckpt_epoch_20.ckpt'
tags_path = '/Volumes/Elements/Projekte/Immersions/logs/immersions_scalogram_resnet/version_7/meta_tags.csv'
ranges_path = '/Users/vincentherrmann/Documents/Projekte/Immersions/immersions/immersions/misc/immersions_scalogram_resnet_ranges.p'
training_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/training'
validation_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/validation'
model = ContrastivePredictiveSystem.load_from_metrics(
weights_path, tags_path, on_gpu=torch.cuda.is_available())
model.freeze()
model.eval()
model.hparams.training_set_path = training_set_path
model.hparams.validation_set_path = validation_set_path
model.batch_size = 8
model.setup_datasets()
register = model.activation_register
item_length = model.preprocessing_receptive_field + 63 * 4096
with open(ranges_path, 'rb') as handle:
activation_ranges = pickle.load(handle)
activation_statistics = ActivationStatistics()
dataloader = model.tng_dataloader
def main(hparams, cluster=None, results_dict=None):
"""
Main training routine specific for this project
:param hparams:
:return:
"""
name = 'immersions_scalogram_resnet_house_smaller'
version = 1
hparams.log_dir = '/home/idivinci3005/experiments/logs'
hparams.checkpoint_dir = '/home/idivinci3005/experiments/checkpoints/' + name + '/' + str(
version)
hparams.training_set_path = '/home/idivinci3005/data/immersions/training'
hparams.validation_set_path = '/home/idivinci3005/data/immersions/validation'
hparams.test_task_set_path = '/home/idivinci3005/data/immersions/test_task'
hparams.dummy_datasets = False
hparams.audio_noise = 3e-3
hparams.cqt_fmin = 40.
hparams.cqt_bins_per_octave = 24
hparams.cqt_n_bins = 216
hparams.cqt_hop_length = 512
hparams.cqt_filter_scale = 0.43
hparams.enc_channels = (1, 8, 16, 32, 64, 128, 256, 512, 512)
hparams.enc_kernel_1_w = (3, 3, 3, 3, 3, 3, 3, 3)
hparams.enc_kernel_1_h = (3, 3, 3, 3, 3, 3, 3, 3)
hparams.enc_kernel_2_w = (1, 3, 1, 3, 1, 3, 1, 3)
hparams.enc_kernel_2_h = (25, 3, 25, 3, 25, 3, 4, 3)
hparams.enc_padding_1 = (1, 1, 1, 1, 1, 1, 1, 1)
hparams.enc_padding_2 = (0, 1, 0, 1, 0, 1, 0, 0)
hparams.enc_stride_1 = (1, 1, 1, 1, 1, 1, 1, 1)
hparams.enc_stride_2 = (1, 1, 1, 1, 1, 1, 1, 1)
hparams.enc_pooling_1 = (2, 1, 1, 1, 2, 1, 1, 1)
hparams.ar_kernel_sizes = (5, 4, 1, 3, 3, 1, 3, 1, 6)
hparams.ar_self_attention = (False, False, False, False, False, False,
False, False, False)
hparams.batch_size = 4
hparams.learning_rate = 3e-4
hparams.warmup_steps = 1000
hparams.annealing_steps = 100000
hparams.score_over_all_timesteps = False
hparams.visible_steps = 60
hparams.batch_size = 32
hparams.learning_rate = 3e-4
hparams.warmup_steps = 1000
hparams.annealing_steps = 100000
hparams.score_over_all_timesteps = False
hparams.visible_steps = 60
# init experiment
exp = Experiment(name=name,
debug=False,
save_dir=hparams.log_dir,
version=version,
autosave=False,
description='test demo')
# set the hparams for the experiment
exp.argparse(hparams)
exp.save()
# build model
model = ContrastivePredictiveSystem(hparams)
task_model = ClassificationTaskModel(
model, task_dataset_path=hparams.test_task_set_path)
# callbacks
early_stop = EarlyStopping(monitor=hparams.early_stop_metric,
patience=hparams.early_stop_patience,
verbose=True,
mode=hparams.early_stop_mode)
checkpoint = ModelCheckpoint(filepath=hparams.checkpoint_dir,
save_best_only=False,
verbose=True,
monitor=hparams.model_save_monitor_value,
mode=hparams.model_save_monitor_mode)
# configure trainer
trainer = Trainer(
experiment=exp,
checkpoint_callback=checkpoint,
#early_stop_callback=early_stop,
#distributed_backend='dp',
gpus=[0],
nb_sanity_val_steps=5,
val_check_interval=0.2,
gradient_clip=0.5,
track_grad_norm=2)
# train model
trainer.fit(model)
def main(hparams, cluster=None, results_dict=None):
"""
Main training routine specific for this project
:param hparams:
:return:
"""
name = 'immersions_scalogram_resnet_house'
version = 0
hparams.log_dir = '/home/idivinci3005/experiments/logs'
hparams.checkpoint_dir = '/home/idivinci3005/experiments/checkpoints/' + name + '/' + str(
version)
hparams.training_set_path = '/home/idivinci3005/data/immersions/training'
hparams.validation_set_path = '/home/idivinci3005/data/immersions/validation'
hparams.test_task_set_path = '/home/idivinci3005/data/immersions/test_task'
hparams.dummy_datasets = False
hparams.audio_noise = 3e-3
hparams.ar_kernel_sizes = (5, 4, 1, 3, 3, 1, 3, 1, 6)
hparams.ar_self_attention = (False, False, False, False, False, False,
False, False, False)
hparams.batch_size = 32
hparams.learning_rate = 3e-4
hparams.warmup_steps = 1000
hparams.annealing_steps = 100000
hparams.score_over_all_timesteps = False
hparams.visible_steps = 62
# init experiment
exp = Experiment(name=name,
debug=False,
save_dir=hparams.log_dir,
version=version,
autosave=False,
description='test demo')
# set the hparams for the experiment
exp.argparse(hparams)
exp.save()
# build model
model = ContrastivePredictiveSystem(hparams)
task_model = ClassificationTaskModel(
model, task_dataset_path=hparams.test_task_set_path)
# callbacks
early_stop = EarlyStopping(monitor=hparams.early_stop_metric,
patience=hparams.early_stop_patience,
verbose=True,
mode=hparams.early_stop_mode)
checkpoint = ModelCheckpoint(filepath=hparams.checkpoint_dir,
save_best_only=False,
verbose=True,
monitor=hparams.model_save_monitor_value,
mode=hparams.model_save_monitor_mode)
# configure trainer
trainer = Trainer(
experiment=exp,
checkpoint_callback=checkpoint,
early_stop_callback=early_stop,
#distributed_backend='dp',
gpus=[0],
nb_sanity_val_steps=5,
val_check_interval=0.2)
# train model
trainer.fit(model)
def test_training(self):
# use default args given by lightning
root_dir = '/Volumes/Elements/Projekte/Immersions'
parent_parser = HyperOptArgumentParser(strategy='random_search', add_help=False)
add_default_args(parent_parser, root_dir)
# allow model to overwrite or extend args
parser = ContrastivePredictiveSystem.add_model_specific_args(parent_parser, root_dir)
hparams = parser.parse_args()
name = 'immersions_scalogram_resnet_test'
version = 0
hparams.log_dir = '/Volumes/Elements/Projekte/Immersions/logs'
hparams.checkpoint_dir = '/Volumes/Elements/Projekte/Immersions/checkpoints'
hparams.training_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/training'
hparams.validation_set_path = '/Volumes/Elements/Datasets/Immersions/house_data_mp3/validation'
hparams.dummy_datasets = False
hparams.batch_size = 64
hparams.learning_rate = 2e-4
hparams.warmup_steps = 1000
hparams.annealing_steps = 100000
# init experiment
exp = Experiment(
name=name,
debug=False,
save_dir=hparams.log_dir,
version=version,
autosave=False,
description='test demo'
)
# set the hparams for the experiment
exp.argparse(hparams)
exp.save()
# build model
model = ContrastivePredictiveSystem(hparams)
# callbacks
early_stop = EarlyStopping(
monitor=hparams.early_stop_metric,
patience=hparams.early_stop_patience,
verbose=True,
mode=hparams.early_stop_mode
)
checkpoint = ModelCheckpoint(
filepath=hparams.checkpoint_dir,
save_best_only=False,
verbose=True,
monitor=hparams.model_save_monitor_value,
mode=hparams.model_save_monitor_mode
)
# configure trainer
trainer = Trainer(
experiment=exp,
checkpoint_callback=checkpoint,
# early_stop_callback=early_stop,
# distributed_backend='dp',
gpus=[0],
nb_sanity_val_steps=5,
val_check_interval=0.2,
train_percent_check=0.01,
max_nb_epochs=1
)
# train model
trainer.fit(model)