def main(exp, frame_sizes, generate_from, **params):
params = dict(default_params,
exp=exp,
frame_sizes=frame_sizes,
generate_from=generate_from,
**params)
model = SampleRNN(
frame_sizes=params['frame_sizes'],
n_rnn=params['n_rnn'],
dim=params['dim'],
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
nb_classes=params['nb_classes'],
weight_norm=params['weight_norm'],
)
# model = SampleRNN([16, 4], 2, 1024, True, 256, True)
print('Loading saved model' + params['generate_from'])
checkpoint = torch.load(params['generate_from'])
temporary_dict = {}
for k, v in checkpoint.items():
temporary_dict[k[6:]] = v
checkpoint = temporary_dict
model.load_state_dict(checkpoint)
if not os.path.exists(params['generate_to']):
os.mkdir(params['generate_to'])
print(params['cond'])
generator = GeneratorPlugin(params['generate_to'], params['n_samples'],
params['sample_length'], params['sample_rate'],
params['nb_classes'], params['cond'])
generator.register_generate(model.cuda(), params['cuda'])
generator.epoch(exp)
def main(exp, frame_sizes, dataset, **params):
params = dict(default_params,
exp=exp,
frame_sizes=frame_sizes,
dataset=dataset,
**params)
results_path = setup_results_dir(params)
tee_stdout(os.path.join(results_path, 'log'))
# Save samplernn parameters in .json for future audio generation
import json
with open(os.path.join(results_path, 'sample_rnn_params.json'), 'w') as fp:
json.dump(params, fp, sort_keys=True, indent=4)
model = SampleRNN(frame_sizes=params['frame_sizes'],
n_rnn=params['n_rnn'],
dim=params['dim'],
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
weight_norm=params['weight_norm'])
predictor = Predictor(model)
if params['cuda']:
model = model.cuda()
predictor = predictor.cuda()
optimizer = gradient_clipping(torch.optim.Adam(predictor.parameters()))
data_loader = make_data_loader(model.lookback, params)
test_split = 1 - params['test_frac']
val_split = test_split - params['val_frac']
trainer = Trainer(predictor,
sequence_nll_loss_bits,
optimizer,
data_loader(0, val_split, eval=False),
cuda=params['cuda'])
checkpoints_path = os.path.join(results_path, 'checkpoints')
checkpoint_data = load_last_checkpoint(checkpoints_path)
if checkpoint_data is not None:
(state_dict, epoch, iteration) = checkpoint_data
trainer.epochs = epoch
trainer.iterations = iteration
predictor.load_state_dict(state_dict)
else:
trainer.epochs = 0
trainer.iterations = 0
torch.save(predictor,
os.path.join(checkpoints_path, "pytorch_model.bin"))
# else:
# print("***** Saving fine-tuned model *****")
# output_model_file = os.path.join(params['results_path'], "pytorch_model.bin")
# if params['cuda']:
# torch.save(predictor, output_model_file)
# else:
# torch.save(predictor, output_model_file)
trainer.register_plugin(
TrainingLossMonitor(smoothing=params['loss_smoothing']))
trainer.register_plugin(
ValidationPlugin(data_loader(val_split, test_split, eval=True),
data_loader(test_split, 1, eval=True)))
trainer.register_plugin(AbsoluteTimeMonitor())
trainer.register_plugin(
SaverPlugin(checkpoints_path, params['keep_old_checkpoints']))
trainer.register_plugin(
GeneratorPlugin(os.path.join(results_path, 'samples'),
params['n_samples'], params['sample_length'],
params['sample_rate'], params['sampling_temperature']))
trainer.register_plugin(
Logger(['training_loss', 'validation_loss', 'test_loss', 'time']))
trainer.register_plugin(
StatsPlugin(results_path,
iteration_fields=[
'training_loss', ('training_loss', 'running_avg'),
'time'
],
epoch_fields=['validation_loss', 'test_loss', 'time'],
plots={
'loss': {
'x':
'iteration',
'ys': [
'training_loss',
('training_loss', 'running_avg'),
'validation_loss',
'test_loss',
],
'log_y':
True
}
}))
init_comet(params, trainer)
trainer.run(params['epoch_limit'])
def main(exp, frame_sizes, dataset, **params):
params = dict(default_params,
exp=exp,
frame_sizes=frame_sizes,
dataset=dataset,
**params)
results_path = setup_results_dir(params)
tee_stdout(os.path.join(results_path, 'log'))
model = SampleRNN(frame_sizes=params['frame_sizes'],
n_rnn=params['n_rnn'],
dim=params['dim'],
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
weight_norm=params['weight_norm'])
predictor = Predictor(model)
if params['cuda']:
model = model.cuda()
predictor = predictor.cuda()
optimizer = gradient_clipping(torch.optim.Adam(predictor.parameters()))
data_loader = make_data_loader(model.lookback, params)
test_split = 1 - params['test_frac']
val_split = test_split - params['val_frac']
trainer = Trainer(predictor,
sequence_nll_loss_bits,
optimizer,
data_loader(0, val_split, eval=False),
cuda=params['cuda'])
checkpoints_path = os.path.join(results_path, 'checkpoints')
checkpoint_data = load_last_checkpoint(checkpoints_path)
if checkpoint_data is not None:
(state_dict, epoch, iteration) = checkpoint_data
trainer.epochs = epoch
trainer.iterations = iteration
predictor.load_state_dict(state_dict)
trainer.register_plugin(
TrainingLossMonitor(smoothing=params['loss_smoothing']))
trainer.register_plugin(
ValidationPlugin(data_loader(val_split, test_split, eval=True),
data_loader(test_split, 1, eval=True)))
trainer.register_plugin(AbsoluteTimeMonitor())
trainer.register_plugin(
SaverPlugin(checkpoints_path, params['keep_old_checkpoints']))
trainer.register_plugin(
GeneratorPlugin(os.path.join(results_path,
'samples'), params['n_samples'],
params['sample_length'], params['sample_rate']))
trainer.register_plugin(
Logger(['training_loss', 'validation_loss', 'test_loss', 'time']))
trainer.register_plugin(
StatsPlugin(results_path,
iteration_fields=[
'training_loss', ('training_loss', 'running_avg'),
'time'
],
epoch_fields=['validation_loss', 'test_loss', 'time'],
plots={
'loss': {
'x':
'iteration',
'ys': [
'training_loss',
('training_loss', 'running_avg'),
'validation_loss',
'test_loss',
],
'log_y':
True
}
}))
init_comet(params, trainer)
trainer.run(params['epoch_limit'])
def main(exp, dataset, **params):
params = dict(default_params, exp=exp, dataset=dataset, **params)
print(params)
storage_client = None
bucket = None
path = os.path.join(params['datasets_path'], params['dataset'])
if params['bucket']:
storage_client = storage.Client()
bucket = Bucket(storage_client, params['bucket'])
preload_dataset(path, storage_client, bucket)
results_path = setup_results_dir(params)
tee_stdout(os.path.join(results_path, 'log'))
(quantize, dequantize) = quantizer(params['q_method'])
model = SampleRNN(frame_sizes=params['frame_sizes'],
n_rnn=params['n_rnn'],
dim=params['dim'],
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
weight_norm=params['weight_norm'])
predictor = Predictor(model, dequantize)
if params['cuda'] is not False:
print(params['cuda'])
model = model.cuda()
predictor = predictor.cuda()
optimizer = gradient_clipping(
torch.optim.Adam(predictor.parameters(), lr=params['learning_rate']))
data_loader = make_data_loader(path, model.lookback, quantize, params)
test_split = 1 - params['test_frac']
val_split = test_split - params['val_frac']
trainer = Trainer(predictor,
sequence_nll_loss_bits,
optimizer,
data_loader(0, val_split, eval=False),
cuda=params['cuda'])
checkpoints_path = os.path.join(results_path, 'checkpoints')
checkpoint_data = load_last_checkpoint(checkpoints_path, storage_client,
bucket)
if checkpoint_data is not None:
(state_dict, epoch, iteration) = checkpoint_data
trainer.epochs = epoch
trainer.iterations = iteration
predictor.load_state_dict(state_dict)
trainer.register_plugin(
TrainingLossMonitor(smoothing=params['loss_smoothing']))
trainer.register_plugin(
ValidationPlugin(data_loader(val_split, test_split, eval=True),
data_loader(test_split, 1, eval=True)))
trainer.register_plugin(SchedulerPlugin(params['lr_scheduler_step']))
def upload(file_path):
if bucket is None:
return
name = file_path.replace(os.path.abspath(os.curdir) + '/', '')
blob = Blob(name, bucket)
try:
blob.upload_from_filename(file_path, timeout=300)
except Exception as e:
print(str(e))
trainer.register_plugin(AbsoluteTimeMonitor())
samples_path = os.path.join(results_path, 'samples')
trainer.register_plugin(
SaverPlugin(checkpoints_path, params['keep_old_checkpoints'], upload))
trainer.register_plugin(
GeneratorPlugin(samples_path,
params['n_samples'],
params['sample_length'],
params['sample_rate'],
params['q_levels'],
dequantize,
params['sampling_temperature'],
upload=upload))
trainer.register_plugin(
Logger(['training_loss', 'validation_loss', 'test_loss', 'time']))
trainer.register_plugin(
StatsPlugin(
results_path,
iteration_fields=[
'training_loss',
#('training_loss', 'running_avg'),
'time'
],
epoch_fields=[
'training_loss', ('training_loss', 'running_avg'),
'validation_loss', 'test_loss', 'time'
],
plots={
'loss': {
'x':
'iteration',
'ys': [
'training_loss',
# ('training_loss', 'running_avg'),
'validation_loss',
'test_loss'
],
'log_y':
True
}
}))
init_comet(params, trainer, samples_path, params['n_samples'],
params['sample_rate'])
trainer.run(params['epoch_limit'])
def main(checkpoint, **args):
task_id = setup_logging(
'gen', logging.NOTSET if args.get('debug', False) else logging.INFO)
params = dict(
{
'n_rnn': 3,
'dim': 1024,
'learn_h0': False,
'q_levels': 256,
'weight_norm': True,
'frame_sizes': [16, 16, 4],
'sample_rate': 16000,
'n_samples': 1,
'sample_length': 16000 * 60 * 4,
'sampling_temperature': 1,
'q_method': QMethod.LINEAR,
},
exp=checkpoint,
**args)
logging.info(str(params))
logging.info('booting')
# dataset = storage_client.list_blobs(bucket, prefix=path)
# for blob in dataset:
# blob.download_to_filename(blob.name)
bucket = None
if args['bucket']:
logging.debug('setup google storage bucket {}'.format(args['bucket']))
storage_client = storage.Client()
bucket = Bucket(storage_client, args['bucket'])
preload_checkpoint(checkpoint, storage_client, bucket)
results_path = os.path.abspath(
os.path.join(checkpoint, os.pardir, os.pardir, task_id))
ensure_dir_exists(results_path)
checkpoint = os.path.abspath(checkpoint)
tmp_pretrained_state = torch.load(
checkpoint,
map_location=lambda storage, loc: storage.cuda(0)
if args['cuda'] else storage)
# Load all tensors onto GPU 1
# torch.load('tensors.pt', map_location=lambda storage, loc: storage.cuda(1))
pretrained_state = OrderedDict()
for k, v in tmp_pretrained_state.items():
# Delete "model." from key names since loading the checkpoint automatically attaches it
layer_name = k.replace("model.", "")
pretrained_state[layer_name] = v
# print("k: {}, layer_name: {}, v: {}".format(k, layer_name, np.shape(v)))
# Create model with same parameters as used in training
model = SampleRNN(frame_sizes=params['frame_sizes'],
n_rnn=params['n_rnn'],
dim=params['dim'],
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
weight_norm=params['weight_norm'])
if params['cuda']:
model = model.cuda()
# Load pretrained model
model.load_state_dict(pretrained_state)
def upload(file_path):
if bucket is None:
return
# remove prefix /app
name = file_path.replace(os.path.abspath(os.curdir) + '/', '')
blob = Blob(name, bucket)
logging.info('uploading {}'.format(name))
blob.upload_from_filename(file_path)
(_, dequantize) = quantizer(params['q_method'])
gen = Gen(Runner(model), params['cuda'])
gen.register_plugin(
GeneratorPlugin(results_path, params['n_samples'],
params['sample_length'], params['sample_rate'],
params['q_levels'], dequantize,
params['sampling_temperature'], upload))
gen.run()
# Load pretrained model
model.load_state_dict(new_pretrained_state)
# Generate Plugin
num_samples = 1 # params['n_samples']
sample_length = params['sample_length']
sample_rate = params['sample_rate']
sampling_temperature = params['sampling_temperature']
# Override from our options
sample_length = sample_rate * int(options.length)
print("Number samples: {}, sample_length: {}, sample_rate: {}".format(num_samples, sample_length, sample_rate))
print("Generating %d seconds of audio" % (sample_length / sample_rate))
generator = GeneratorPlugin(GENERATED_PATH, num_samples, sample_length, sample_rate, sampling_temperature)
# Call new register function to accept the trained model and the cuda setting
generator.register_generate(model.cuda(), params['cuda'])
# Generate new audio
# $$$ check if we already have generated audio and increment the file name
generator.epoch(OUTPUT_NAME)
GENERATED_FILEPATH = GENERATED_PATH + "ep" + OUTPUT_NAME + "-s1.wav"
print("Saved audio to %s " % GENERATED_FILEPATH)
if options.output:
print("Moving to %s" % options.output)
os.rename(GENERATED_FILEPATH, options.output)
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
weight_norm=params['weight_norm'])
# Delete "model." from key names since loading the checkpoint automatically attaches it to the key names
pretrained_state = torch.load(PRETRAINED_PATH)
new_pretrained_state = OrderedDict()
for k, v in pretrained_state.items():
layer_name = k.replace("model.", "")
new_pretrained_state[layer_name] = v
# print("k: {}, layer_name: {}, v: {}".format(k, layer_name, np.shape(v)))
# Load pretrained model
model.load_state_dict(new_pretrained_state)
# Generate Plugin
num_samples = 2 # params['n_samples']
sample_length = params['sample_length']
sample_rate = params['sample_rate']
print("Number samples: {}, sample_length: {}, sample_rate: {}".format(
num_samples, sample_length, sample_rate))
generator = GeneratorPlugin(GENERATED_PATH, num_samples, sample_length,
sample_rate)
# Call new register function to accept the trained model and the cuda setting
generator.register_generate(model.cuda(), params['cuda'])
# Generate new audio
generator.epoch('Test2')
# Gets initial samples form 1 test sample and check if it re-generates it
audio_filename = dataset_filenames[0]
from librosa.core import load
sr = params['sample_rate']
seq, sr = load(audio_filename, sr=sr, mono=True)
print("Sample rate: {}".format(sr))
# Generate Plugin
num_samples = 6 # params['n_samples']
initial_seq_size = 64 * 100 # has to be multiple of rnn.n_frame_samples ???
initial_seq = None
if initial_seq_size > 1:
init = utils.linear_quantize(torch.from_numpy(seq[0:initial_seq_size]), params['q_levels'])
# init = seq[0:initial_seed_size]
init = np.tile(init, (num_samples, 1))
initial_seq = torch.LongTensor(init)
# initial_seed = utils.linear_quantize(initial_seed, params['q_levels'])
sample_length = params['sample_length']
sample_rate = params['sample_rate']
print("Number samples: {}, sample_length: {}, sample_rate: {}".format(num_samples, sample_length, sample_rate))
generator = GeneratorPlugin(GENERATED_PATH, num_samples, sample_length, sample_rate)
# Overloads register function to accept the trained model and the cuda setting
generator.register_generate(model.cuda(), params['cuda'])
# Generate new audio
generator.epoch('Test19_{}'.format(initial_seq_size), initial_seed=initial_seq)
def main(exp, frame_sizes, dataset, **params):
params = dict(
default_params,
exp=exp, frame_sizes=frame_sizes, dataset=dataset,
**params
)
results_path = setup_results_dir(params)
tee_stdout(os.path.join(results_path, 'log'))
# Save samplernn parameters in .json for future audio generation
import json
with open(os.path.join(results_path, 'sample_rnn_params.json'), 'w') as fp:
json.dump(params, fp, sort_keys=True, indent=4)
# Model
model = SampleRNN(
frame_sizes=params['frame_sizes'],
n_rnn=params['n_rnn'],
dim=params['dim'],
learn_h0=params['learn_h0'],
q_levels=params['q_levels'],
weight_norm=params['weight_norm'],
batch_size=params['batch_size']
)
print("CUDA num: {}".format(torch.cuda.device_count()))
predictor = Predictor(model)
if params['cuda']:
model = model.cuda()
predictor = predictor.cuda()
model_cnnseq2sample = CNNSeq2SampleRNN(params).cuda()
optimizer = gradient_clipping(torch.optim.Adam(predictor.parameters()))
data_loader = make_data_loader(model.lookback, params)
data_loader_test = make_data_loader(model.lookback, params, npz_filename=params['npz_filename_test'])
# test_split = 1 - params['test_frac']
# val_split = test_split - params['val_frac']
trainer = Trainer(
predictor, model_cnnseq2sample, sequence_nll_loss_bits, optimizer,
# data_loader(0, val_split, eval=False),
data_loader(0, 1, eval=False),
cuda=params['cuda']
)
checkpoints_path = os.path.join(results_path, 'checkpoints')
checkpoint_data = load_last_checkpoint(checkpoints_path)
checkpoint_data_cnnseq2sample = load_last_checkpoint(checkpoints_path, model_type='cnnseq2sample')
if checkpoint_data is not None:
(state_dict, epoch, iteration) = checkpoint_data
(state_dict_cnnseq2sample, epoch, iteration) = checkpoint_data_cnnseq2sample
trainer.epochs = epoch
trainer.iterations = iteration
predictor.load_state_dict(state_dict)
model_cnnseq2sample.load_state_dict(state_dict_cnnseq2sample)
trainer.register_plugin(TrainingLossMonitor(
smoothing=params['loss_smoothing']
))
trainer.register_plugin(ValidationPlugin(
# data_loader(val_split, test_split, eval=True),
# data_loader_test(0, 1, eval=True)
data_loader_test(0, params['val_frac'], eval=True),
data_loader_test(params['val_frac'], 1, eval=True)
# data_loader(test_split, 1, eval=True)
))
trainer.register_plugin(AbsoluteTimeMonitor())
trainer.register_plugin(SaverPlugin(
checkpoints_path, params['keep_old_checkpoints']
))
trainer.register_plugin(GeneratorPlugin(
os.path.join(results_path, 'samples'), params['n_samples'],
params['sample_length'], params['sample_rate']
))
trainer.register_plugin(
Logger([
'training_loss',
'validation_loss',
'test_loss',
'time'
])
)
trainer.register_plugin(StatsPlugin(
results_path,
iteration_fields=[
'training_loss',
('training_loss', 'running_avg'),
'time'
],
epoch_fields=[
'validation_loss',
'test_loss',
'time'
],
plots={
'loss': {
'x': 'iteration',
'ys': [
'training_loss',
('training_loss', 'running_avg'),
'validation_loss',
'test_loss',
],
'log_y': True
}
}
))
trainer.run(params['epoch_limit'])