def main():
print(torch.cuda.is_available())
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
cfg = {}
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
outfile = options.outfile
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
bin_nyquist = frame_len // 2 + 1
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
print("Preparing data reading...")
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
# - load filelist
with io.open(os.path.join(datadir, 'filelists', 'valid')) as f:
filelist = [l.rstrip() for l in f if l.rstrip()]
with io.open(os.path.join(datadir, 'filelists', 'test')) as f:
filelist += [l.rstrip() for l in f if l.rstrip()]
# - create generator for spectra
spects = (cached(
options.cache_spectra
and os.path.join(options.cache_spectra, fn + '.npy'),
audio.extract_spect, os.path.join(datadir, 'audio',
fn), sample_rate, frame_len, fps)
for fn in filelist)
# - pitch-shift if needed
if options.pitchshift:
import scipy.ndimage
spline_order = 2
spects = (scipy.ndimage.affine_transform(
spect, (1, 1 / (1 + options.pitchshift / 100.)),
output_shape=(len(spect), mel_max),
order=spline_order) for spect in spects)
# - prepare mel filterbank
filterbank = audio.create_mel_filterbank(sample_rate, frame_len, mel_bands,
mel_min, mel_max)
filterbank = filterbank[:bin_mel_max].astype(floatX)
# - define generator for mel spectra
spects = (np.log(
np.maximum(np.dot(spect[:, :bin_mel_max], filterbank), 1e-7))
for spect in spects)
# - load mean/std
meanstd_file = os.path.join(os.path.dirname(__file__),
'%s_meanstd.npz' % options.dataset)
with np.load(meanstd_file) as f:
mean = f['mean']
std = f['std']
mean = mean.astype(floatX)
istd = np.reciprocal(std).astype(floatX)
# - define generator for Z-scoring
spects = ((spect - mean) * istd for spect in spects)
# - define generator for silence-padding
pad = np.tile((np.log(1e-7) - mean) * istd, (blocklen // 2, 1))
spects = (np.concatenate((pad, spect, pad), axis=0) for spect in spects)
# - we start the generator in a background thread (not required)
spects = augment.generate_in_background([spects], num_cached=1)
mdl = model.CNNModel()
mdl.load_state_dict(torch.load(modelfile))
mdl.to(device)
mdl.eval()
# run prediction loop
print("Predicting:")
predictions = []
for spect in progress(spects, total=len(filelist), desc='File '):
# naive way: pass excerpts of the size used during training
# - view spectrogram memory as a 3-tensor of overlapping excerpts
num_excerpts = len(spect) - blocklen + 1
excerpts = np.lib.stride_tricks.as_strided(
spect,
shape=(num_excerpts, blocklen, spect.shape[1]),
strides=(spect.strides[0], spect.strides[0], spect.strides[1]))
# - pass mini-batches through the network and concatenate results
preds = np.vstack(
mdl(
torch.from_numpy(
np.transpose(
excerpts[pos:pos + batchsize, :, :, np.newaxis], (
0, 3, 1, 2))).to(device)).cpu().detach().numpy()
for pos in range(0, num_excerpts, batchsize))
predictions.append(preds)
# save predictions
print("Saving predictions")
np.savez(outfile, **{fn: pred for fn, pred in zip(filelist, predictions)})
def main():
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
outfile = options.outfile
# read configuration files and immediate settings
cfg = {}
if os.path.exists(modelfile + '.vars'):
options.vars.insert(1, modelfile + '.vars')
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
# read some settings into local variables
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
bin_nyquist = frame_len // 2 + 1
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
print("Preparing data reading...")
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
# - load filelist
filelist = []
for d in options.filelists.split(','):
with io.open(os.path.join(datadir, 'filelists', d)) as f:
filelist.extend(l.rstrip() for l in f if l.rstrip())
# - create generator for spectra
spects = (cached(
options.cache_spectra
and os.path.join(options.cache_spectra, fn + '.npy'),
audio.extract_spect, os.path.join(datadir, 'audio',
fn), sample_rate, frame_len, fps)
for fn in filelist)
# - pitch-shift if needed
if options.pitchshift:
import scipy.ndimage
spline_order = 2
spects = (scipy.ndimage.affine_transform(
spect, (1, 1 / (1 + options.pitchshift / 100.)),
output_shape=(len(spect), mel_max),
order=spline_order) for spect in spects)
# - define generator for cropped spectra
spects = (spect[:, :bin_mel_max] for spect in spects)
# - adjust loudness if needed
if options.loudness:
spects = (spect * float(10.**(options.loudness / 10.))
for spect in spects)
# - define generator for silence-padding
pad = np.zeros((blocklen // 2, bin_mel_max), dtype=floatX)
spects = (np.concatenate((pad, spect, pad), axis=0) for spect in spects)
# - we start the generator in a background thread (not required)
spects = augment.generate_in_background([spects], num_cached=1)
print("Preparing prediction function...")
# instantiate neural network
input_var = T.tensor3('input')
inputs = input_var.dimshuffle(0, 'x', 1, 2) # insert "channels" dimension
network = model.architecture(inputs, (None, 1, blocklen, bin_mel_max), cfg)
# load saved weights
with np.load(modelfile) as f:
lasagne.layers.set_all_param_values(
network, [f['param%d' % i] for i in range(len(f.files))])
# performant way: convert to fully-convolutional network
if not options.mem_use == 'low':
import model_to_fcn
network = model_to_fcn.model_to_fcn(network, allow_unlink=True)
# create output expression
outputs = lasagne.layers.get_output(network, deterministic=True)
# prepare and compile prediction function
print("Compiling prediction function...")
test_fn = theano.function([input_var], outputs)
# run prediction loop
print("Predicting:")
predictions = []
for spect in progress(spects, total=len(filelist), desc='File '):
if options.mem_use == 'high':
# fastest way: pass full spectrogram through network at once
preds = test_fn(spect[np.newaxis]) # insert batch dimension
elif options.mem_use == 'mid':
# performant way: pass spectrogram in equal chunks of up to one
# minute, taking care to overlap by `blocklen // 2` frames and to
# not pass a chunk shorter than `blocklen` frames
chunks = np.ceil(len(spect) / (fps * 60.))
hopsize = int(np.ceil(len(spect) / chunks))
chunksize = hopsize + blocklen - 1
preds = np.vstack(
test_fn(spect[np.newaxis, pos:pos + chunksize])
for pos in range(0, len(spect), hopsize))
else:
# naive way: pass excerpts of the size used during training
# - view spectrogram memory as a 3-tensor of overlapping excerpts
num_excerpts = len(spect) - blocklen + 1
excerpts = np.lib.stride_tricks.as_strided(
spect,
shape=(num_excerpts, blocklen, spect.shape[1]),
strides=(spect.strides[0], spect.strides[0], spect.strides[1]))
# - pass mini-batches through the network and concatenate results
preds = np.vstack(
test_fn(excerpts[pos:pos + batchsize])
for pos in range(0, num_excerpts, batchsize))
predictions.append(preds)
if options.plot:
if spect.ndim == 3:
spect = spect[0] # remove channel axis
spect = spect[blocklen // 2:-blocklen // 2] # remove zero padding
import matplotlib.pyplot as plt
fig, (ax1, ax2) = plt.subplots(2, 1, sharex=True)
ax1.imshow(spect.T[::-1],
vmin=-3,
cmap='hot',
aspect='auto',
interpolation='nearest')
ax2.plot(preds)
ax2.set_ylim(0, 1.1)
plt.show()
# save predictions
print("Saving predictions")
data = dict(zip(filelist, predictions))
if outfile.endswith('.pkl'):
try:
import cPickle as pickle
except ImportError:
import pickle
with io.open(outfile, 'wb') as f:
pickle.dump(data, f, protocol=-1)
else:
np.savez(outfile, **data)
def main():
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
# read configuration files and immediate settings
cfg = {}
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
# prepare dataset
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
print("Preparing training data feed...")
with io.open(os.path.join(datadir, 'filelists', 'train')) as f:
filelist = [l.rstrip() for l in f if l.rstrip()]
train_feed, train_formats = data.prepare_datafeed(filelist, datadir,
'train', cfg)
# If told so, we plot some mini-batches on screen.
if cfg.get('plot_datafeed'):
import matplotlib.pyplot as plt
for batch in data.run_datafeed(train_feed, cfg):
plt.matshow(np.log(batch['spect'][0]).T,
aspect='auto',
origin='lower',
cmap='hot',
interpolation='nearest')
plt.colorbar()
plt.title(str(batch['label'][0]))
plt.show()
# We start the mini-batch generator and augmenter in one or more
# background threads or processes (unless disabled).
bg_threads = cfg['bg_threads']
bg_processes = cfg['bg_processes']
if not bg_threads and not bg_processes:
# no background processing: just create a single generator
batches = data.run_datafeed(train_feed, cfg)
elif bg_threads:
# multithreading: create a separate generator per thread
batches = augment.generate_in_background([
data.run_datafeed(feed, cfg)
for feed in data.split_datafeed(train_feed, bg_threads, cfg)
],
num_cached=bg_threads * 2)
elif bg_processes:
# multiprocessing: single generator is forked along with processes
batches = augment.generate_in_background(
[data.run_datafeed(train_feed, cfg)] * bg_processes,
num_cached=bg_processes * 25,
in_processes=True)
# If told so, we benchmark the creation of a given number of mini-batches.
if cfg.get('benchmark_datafeed'):
print("Benchmark: %d mini-batches of %d items " %
(cfg['benchmark_datafeed'], cfg['batchsize']),
end='')
if bg_threads:
print("(in %d threads): " % bg_threads)
elif bg_processes:
print("(in %d processes): " % bg_processes)
else:
print("(in main thread): ")
import time
import itertools
t0 = time.time()
next(
itertools.islice(batches, cfg['benchmark_datafeed'],
cfg['benchmark_datafeed']), None)
t1 = time.time()
print(t1 - t0)
return
# - prepare validation data generator
if options.validate:
print("Preparing validation data feed...")
with io.open(os.path.join(datadir, 'filelists', 'valid')) as f:
filelist_val = [l.rstrip() for l in f if l.rstrip()]
val_feed, val_formats = data.prepare_datafeed(filelist_val, datadir,
'valid', cfg)
if bg_threads or bg_processes:
multi = bg_threads or bg_processes
val_feed = data.split_datafeed(val_feed, multi, cfg)
def run_val_datafeed():
if bg_threads or bg_processes:
return augment.generate_in_background(
[data.run_datafeed(feed, cfg) for feed in val_feed],
num_cached=multi,
in_processes=bool(bg_processes))
else:
return data.run_datafeed(val_feed, cfg)
print("Preparing training function...")
# instantiate neural network
input_vars = {
name: T.TensorType(str(np.dtype(dtype)), (False, ) * len(shape))(name)
for name, (dtype, shape) in train_formats.items()
}
input_shapes = {
name: shape
for name, (dtype, shape) in train_formats.items()
}
network = model.architecture(input_vars, input_shapes, cfg)
print(
"- %d layers (%d with weights), %f mio params" %
(len(lasagne.layers.get_all_layers(network)),
sum(hasattr(l, 'W') for l in lasagne.layers.get_all_layers(network)),
lasagne.layers.count_params(network, trainable=True) / 1e6))
print("- weight shapes: %r" % [
l.W.get_value().shape for l in lasagne.layers.get_all_layers(network)
if hasattr(l, 'W') and hasattr(l.W, 'get_value')
])
cost_vars = dict(input_vars)
# prepare for born-again-network, if needed
if cfg.get('ban'):
network2 = model.architecture(input_vars, input_shapes, cfg)
with np.load(cfg['ban'], encoding='latin1') as f:
lasagne.layers.set_all_param_values(
network2, [f['param%d' % i] for i in range(len(f.files))])
cost_vars['pseudo_label'] = lasagne.layers.get_output(
network2, deterministic=True)
# load pre-trained weights, if needed
if cfg.get('init_from'):
param_values = []
for fn in cfg['init_from'].split(':'):
with np.load(fn, encoding='latin1') as f:
param_values.extend(f['param%d' % i]
for i in range(len(f.files)))
lasagne.layers.set_all_param_values(network, param_values)
del param_values
# create cost expression
outputs = lasagne.layers.get_output(network, deterministic=False)
cost = T.mean(model.cost(outputs, cost_vars, 'train', cfg))
if cfg.get('l2_decay', 0):
cost_l2 = lasagne.regularization.regularize_network_params(
network, lasagne.regularization.l2) * cfg['l2_decay']
else:
cost_l2 = 0
# prepare and compile training function
params = lasagne.layers.get_all_params(network, trainable=True)
initial_eta = cfg['initial_eta']
eta_decay = cfg['eta_decay']
eta_decay_every = cfg.get('eta_decay_every', 1)
eta_cycle = tuple(map(float, str(cfg['eta_cycle']).split(':')))
if eta_cycle == (0, ):
eta_cycle = (1, ) # so eta_cycle=0 equals disabling it
patience = cfg.get('patience', 0)
trials_of_patience = cfg.get('trials_of_patience', 1)
patience_criterion = cfg.get(
'patience_criterion',
'valid_loss' if options.validate else 'train_loss')
momentum = cfg['momentum']
first_params = params[:cfg['first_params']]
first_params_eta_scale = cfg['first_params_eta_scale']
if cfg['learn_scheme'] == 'nesterov':
learn_scheme = lasagne.updates.nesterov_momentum
elif cfg['learn_scheme'] == 'momentum':
learn_scheme = lasagne.update.momentum
elif cfg['learn_scheme'] == 'adam':
learn_scheme = lasagne.updates.adam
else:
raise ValueError('Unknown learn_scheme=%s' % cfg['learn_scheme'])
eta = theano.shared(lasagne.utils.floatX(initial_eta))
if not first_params or first_params_eta_scale == 1:
updates = learn_scheme(cost + cost_l2, params, eta, momentum)
else:
grads = theano.grad(cost + cost_l2, params)
updates = learn_scheme(grads[len(first_params):],
params[len(first_params):], eta, momentum)
if first_params_eta_scale > 0:
updates.update(
learn_scheme(grads[:len(first_params)], first_params,
eta * first_params_eta_scale, momentum))
print("Compiling training function...")
train_fn = theano.function(list(input_vars.values()),
cost,
updates=updates,
on_unused_input='ignore')
# prepare and compile validation function, if requested
if options.validate:
print("Compiling validation function...")
outputs_test = lasagne.layers.get_output(network, deterministic=True)
cost_test = T.mean(model.cost(outputs_test, input_vars, 'valid', cfg))
if isinstance(outputs_test, (list, tuple)):
outputs_test = outputs_test[0]
val_fn = theano.function([input_vars[k] for k in val_formats],
[cost_test, outputs_test],
on_unused_input='ignore')
# restore previous training state, or create fresh training state
state = {}
if options.keep_state:
statefile = modelfile[:-len('.npz')] + '.state'
if os.path.exists(statefile):
print("Restoring training state...")
state = np.load(modelfile[:-len('.npz')] + '.state',
encoding='latin1')
restore_state(network, updates, state['network'])
epochs = cfg['epochs']
epochsize = cfg['epochsize']
batches = iter(batches)
if options.save_errors:
errors = state.get('errors', [])
if first_params and cfg['first_params_log']:
first_params_hist = []
if options.keep_state and os.path.exists(modelfile[:-4] + '.hist.npz'):
with np.load(modelfile[:-4] + '.hist.npz') as f:
first_params_hist = list(
zip(*(f['param%d' % i] for i in range(len(first_params)))))
if patience > 0:
best_error = state.get('best_error', np.inf)
best_state = state.get('best_state') or get_state(network, updates)
patience = state.get('patience', patience)
trials_of_patience = state.get('trials_of_patience',
trials_of_patience)
epoch = state.get('epoch', 0)
del state
# run training loop
print("Training:")
for epoch in range(epoch, epochs):
# actual training
err = 0
for batch in progress(range(epochsize),
min_delay=.5,
desc='Epoch %d/%d: Batch ' %
(epoch + 1, epochs)):
err += train_fn(**next(batches))
if not np.isfinite(err):
print("\nEncountered NaN loss in training. Aborting.")
sys.exit(1)
if first_params and cfg['first_params_log'] and (
batch % cfg['first_params_log'] == 0):
first_params_hist.append(
tuple(param.get_value() for param in first_params))
np.savez(
modelfile[:-4] + '.hist.npz', **{
'param%d' % i: param
for i, param in enumerate(zip(*first_params_hist))
})
# report training loss
print("Train loss: %.3f" % (err / epochsize))
if options.save_errors:
errors.append(err / epochsize)
# compute and report validation loss, if requested
if options.validate:
import time
t0 = time.time()
# predict in mini-batches
val_err = 0
val_batches = 0
preds = []
truth = []
for batch in run_val_datafeed():
e, p = val_fn(**batch)
val_err += np.sum(e)
val_batches += 1
preds.append(p)
truth.append(batch['label'])
t1 = time.time()
# join mini-batches
preds = np.concatenate(preds) if len(preds) > 1 else preds[0]
truth = np.concatenate(truth) if len(truth) > 1 else truth[0]
# show results
print("Validation loss: %.3f" % (val_err / val_batches))
from eval import evaluate
results = evaluate(preds, truth)
print("Validation error: %.3f" % (1 - results['accuracy']))
print("Validation MAP: %.3f" % results['map'])
print("(took %.2f seconds)" % (t1 - t0))
if options.save_errors:
errors.append(val_err / val_batches)
errors.append(1 - results['accuracy'])
errors.append(results['map'])
# update learning rate and/or apply early stopping, if needed
if patience > 0:
if patience_criterion == 'train_loss':
cur_error = err / epochsize
elif patience_criterion == 'valid_loss':
cur_error = val_err / val_batches
elif patience_criterion == 'valid_error':
cur_error = 1 - results['accuracy']
elif patience_criterion == 'valid_map':
cur_error = 1 - results['map']
if cur_error <= best_error:
best_error = cur_error
best_state = get_state(network, updates)
patience = cfg['patience']
else:
patience -= 1
if patience == 0:
if eta_decay_every == 'trial_of_patience' and eta_decay != 1:
eta.set_value(eta.get_value() *
lasagne.utils.floatX(eta_decay))
restore_state(network, updates, best_state)
patience = cfg['patience']
trials_of_patience -= 1
print("Lost patience (%d remaining trials)." %
trials_of_patience)
if trials_of_patience == 0:
break
if eta_decay_every != 'trial_of_patience' and eta_decay != 1 and \
(epoch + 1) % eta_decay_every == 0:
eta.set_value(eta.get_value() * lasagne.utils.floatX(eta_decay))
if eta_cycle[epoch % len(eta_cycle)] != 1:
eta.set_value(
eta.get_value() *
lasagne.utils.floatX(eta_cycle[epoch % len(eta_cycle)]))
# store current training state, if needed
if options.keep_state:
state = {}
state['epoch'] = epoch + 1
state['network'] = get_state(network, updates)
if options.save_errors:
state['errors'] = errors
if patience > 0:
state['best_error'] = best_error
state['best_state'] = best_state
state['patience'] = patience
state['trials_of_patience'] = trials_of_patience
with open(statefile, 'wb') as f:
pickle.dump(state, f, -1)
del state
# for debugging: print memory use and break into debugger
#import resource, psutil
#print("Memory usage: %.3f MiB / %.3f MiB" %
# (resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024.,
# psutil.Process().memory_info()[0] / float(1024**2)))
#import pdb; pdb.set_trace()
# save final network
print("Saving final model")
save_model(modelfile, network, cfg)
if options.save_errors:
np.savez(modelfile[:-len('.npz')] + '.err.npz',
np.asarray(errors).reshape(epoch + 1, -1))
def main():
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
outfile = options.outfile
if options.split_pool and options.saliency:
parser.error("--split-pool and --saliency cannot be combined.")
# read configuration files and immediate settings
cfg = {}
if os.path.exists(modelfile + '.vars'):
options.vars.insert(1, modelfile + '.vars')
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
# read some settings into local variables
fps = cfg['fps']
len_min = cfg['len_min']
len_max = cfg['len_max']
# prepare dataset
print("Preparing data reading...")
datadir = os.path.join(os.path.dirname(__file__),
os.path.pardir, 'datasets', options.dataset)
# - load filelists
filelist = []
for d in options.filelists.split(','):
with io.open(os.path.join(datadir, 'filelists', d)) as f:
filelist.extend(l.rstrip() for l in f if l.rstrip())
# - create data feed
feed, input_formats = data.prepare_datafeed(filelist, datadir, 'test', cfg)
# - we start the generator in a background thread
if not options.plot:
batches = augment.generate_in_background([data.run_datafeed(feed, cfg)],
num_cached=1)
else:
# unless we're plotting; this would mess up the progress counter
batches = data.run_datafeed(feed, cfg)
print("Preparing prediction function...")
# instantiate neural network
input_vars = {name: T.TensorType(str(np.dtype(dtype)),
(False,) * len(shape))(name)
for name, (dtype, shape) in input_formats.items()}
input_shapes = {name: shape
for name, (dtype, shape) in input_formats.items()}
network = model.architecture(input_vars, input_shapes, cfg)
if isinstance(network, list) and not options.include_side_outputs:
network = network[0] # only use the main output
# load saved weights
with np.load(modelfile, encoding='latin1') as f:
lasagne.layers.set_all_param_values(
network, [f['param%d' % i] for i in range(len(f.files))])
# insert guided backprop, if needed for saliency
if options.saliency:
from gbprop import replace_nonlinearities
replace_nonlinearities(network, lasagne.nonlinearities.leaky_rectify)
# create output expression(s)
if options.split_pool:
network_end = network
network = next(l for l in lasagne.layers.get_all_layers(network)[::-1]
if l.name == 'before_pool')
outputs = lasagne.layers.get_output(network, deterministic=True)
if options.split_pool:
split_input_var = T.tensor4('input2')
split_outputs = lasagne.layers.get_output(
network_end, {network: split_input_var}, deterministic=True)
split_input_vars = [v for v in theano.gof.graph.inputs([split_outputs])
if not isinstance(v, theano.compile.SharedVariable)
and not isinstance(v, theano.tensor.Constant)]
# create saliency map expression, if needed
if options.saliency:
saliency = theano.grad(outputs[:, options.saliency].sum(), input_vars['spect'])
outputs = outputs + [saliency] if isinstance(outputs, list) else [outputs, saliency]
# prepare and compile prediction function
print("Compiling prediction function...")
test_fn = theano.function(list(input_vars.values()), outputs,
on_unused_input='ignore')
if options.split_pool:
pool_fn = theano.function(split_input_vars, split_outputs,
on_unused_input='ignore')
# prepare plotting, if needed
if options.plot:
import matplotlib
if os.environ.get('MPLBACKEND'):
matplotlib.use(os.environ['MPLBACKEND']) # for old versions
import matplotlib.pyplot as plt
with open(os.path.join(datadir, 'labels', 'labelset'), 'rb') as f:
labelset = [l.rstrip('\r\n') for l in f]
# run prediction loop
print("Predicting:")
predictions = []
for batch in batches:
spect = batch.pop('spect')
if spect.shape[-2] <= len_max * fps or len_max == 0:
# predict on full spectrogram at once
preds = test_fn(spect=spect, **batch)
else:
# predict in segments of len_max, with overlap len_min
# drop any reminder shorter than len_min (len_max if len_min == 0)
preds = [test_fn(spect=spect[..., pos:pos + len_max * fps, :],
**batch)
for pos in range(0, (spect.shape[-2] + 1 -
(len_min or len_max) * fps),
(len_max - len_min) * fps)]
if isinstance(preds[0], list):
preds = [np.concatenate(p, axis=2 if p[0].ndim > 2 else 0)
for p in zip(*preds)]
else:
preds = np.concatenate(preds,
axis=2 if preds[0].ndim > 2 else 0)
if cfg['arch.pool'] == 'none' or '_nopool' in cfg['arch.pool']:
if isinstance(preds, list):
preds = [p[0, :, :, 0].T if p.ndim == 4 else p for p in preds]
else:
preds = preds[0, :, :, 0].T
elif options.split_pool:
preds = pool_fn(preds, **batch)
predictions.append(preds)
if options.plot:
if spect.ndim == 4:
spect = spect[0] # remove batch axis
if spect.ndim == 3:
spect = spect[0] # remove channel axis
if isinstance(preds, list):
preds, sides = preds[0], preds[1:]
else:
sides = []
fig, axs = plt.subplots(2 + len(sides), 1, sharex=True)
axs[0].imshow(np.log1p(1e-3 * spect).T[::-1], cmap='hot',
aspect='auto', interpolation='nearest')
K = 5
top_k = lme(preds, axis=0).argpartition(preds.shape[1] - 1 -
np.arange(K))[::-1][:K]
#top_k = (preds * softmax(sides[0], axis=0).mean(axis=1, keepdims=True)).sum(axis=0).argpartition(preds.shape[1] - 1 - np.arange(K))[::-1][:K]
#top_k = softmax(preds, axis=-1).max(axis=0).argpartition(preds.shape[1] - 1 - np.arange(K))[::-1][:K]
#top_k[-1] = labelset.index('mphbjm')
preds = softmax(preds, axis=-1)
x = np.arange(len(preds)) * (len(spect) / float(len(preds)))
for k in top_k:
axs[1].plot(x, preds[:, k], label=labelset[k])
#axs[1].set_ylim(0, 1.1)
axs[1].legend(loc='best')
for side, ax in zip(sides, axs[2:]):
side = softmax(side, axis=0)
ax.plot(x, side)
plt.show()
# save predictions
print("Saving predictions")
predictions = dict(zip(filelist, predictions))
if outfile.endswith('.pkl'):
try:
import cPickle as pickle
except ImportError:
import pickle
with io.open(outfile, 'wb') as f:
pickle.dump(predictions, f, protocol=-1)
else:
np.savez(outfile, **predictions)
def main():
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
lossgradient = options.lossgradient
cfg = {}
print(options.vars)
print('Model save file:', modelfile)
print('Lossgrad file:', lossgradient)
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
print('Occluded amount:',cfg['occlude'])
bin_nyquist = frame_len // 2 + 1
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
datadir = os.path.join(os.path.dirname(__file__),
os.path.pardir, 'datasets', options.dataset)
meanstd_file = os.path.join(os.path.dirname(__file__),
'%s_meanstd.npz' % options.dataset)
dataloader = DatasetLoader(options.dataset, options.cache_spectra, datadir, input_type=options.input_type)
batches = dataloader.prepare_batches(sample_rate, frame_len, fps,
mel_bands, mel_min, mel_max, blocklen, batchsize)
validation_data = DatasetLoader(options.dataset, '../ismir2015/experiments/mel_data/', datadir,
dataset_split='valid', input_type='mel_spects')
mel_spects_val, labels_val = validation_data.prepare_batches(sample_rate, frame_len, fps,
mel_bands, mel_min, mel_max, blocklen, batchsize, batch_data=False)
with np.load(meanstd_file) as f:
mean = f['mean']
std = f['std']
mean = mean.astype(floatX)
istd = np.reciprocal(std).astype(floatX)
if(options.input_type=='mel_spects'):
mdl = model.CNNModel(input_type='mel_spects_norm', is_zeromean=False,
sample_rate=sample_rate, frame_len=frame_len, fps=fps,
mel_bands=mel_bands, mel_min=mel_min, mel_max=mel_max,
bin_mel_max=bin_mel_max, meanstd_file=meanstd_file, device=device)
if(lossgradient!='None'):
mdl_lossgrad = model.CNNModel(input_type=options.input_type,
is_zeromean=False, sample_rate=sample_rate, frame_len=frame_len, fps=fps,
mel_bands=mel_bands, mel_min=mel_min, mel_max=mel_max,
bin_mel_max=bin_mel_max, meanstd_file=meanstd_file, device=device)
mdl_lossgrad.load_state_dict(torch.load(lossgradient))
mdl_lossgrad.to(device)
mdl_lossgrad.eval()
mdl = mdl.to(device)
#Setting up learning rate and learning rate parameters
initial_eta = cfg['initial_eta']
eta_decay = cfg['eta_decay']
momentum = cfg['momentum']
eta_decay_every = cfg.get('eta_decay_every', 1)
eta = initial_eta
#set up loss
criterion = torch.nn.BCELoss()
#set up optimizer
optimizer = torch.optim.SGD(mdl.parameters(),lr=eta,momentum=momentum,nesterov=True)
#optimizer = torch.optim.Adam(mdl.parameters(), lr=eta, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, amsgrad=False)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,step_size=eta_decay_every,gamma=eta_decay)
#set up optimizer
writer = SummaryWriter(os.path.join(modelfile,'runs'))
epochs = cfg['epochs']
epochsize = cfg['epochsize']
batches = iter(batches)
#conditions to save model
best_val_loss = 100000.
best_val_error = 1.
#loss gradient values for validation data
loss_grad_val = validation_data.prepare_loss_grad_batches(options.loss_grad_save,
mel_spects_val, labels_val, mdl_lossgrad, criterion, blocklen, batchsize, device)
for epoch in range(epochs):
# - Initialize certain parameters that are used to monitor training
err = 0
total_norm = 0
loss_accum = 0
mdl.train(True)
# - Compute the L-2 norm of the gradients
for p in mdl.parameters():
if p.grad is not None:
param_norm = p.grad.data.norm(2)
total_norm += param_norm.item() ** 2
total_norm = total_norm ** (1. / 2)
# - Start the training for this epoch
for batch in progress(range(epochsize), min_delay=0.5,desc='Epoch %d/%d: Batch ' % (epoch+1, epochs)):
data = next(batches)
if(options.input_type=='audio' or options.input_type=='stft'):
input_data = data[0]
else:
input_data = np.transpose(data[0][:,:,:,np.newaxis],(0,3,1,2))
labels = data[1][:,np.newaxis].astype(np.float32)
input_data_loss = input_data
if lossgradient!='None':
g = loss_grad(mdl_lossgrad, torch.from_numpy(input_data_loss).to(device).requires_grad_(True), torch.from_numpy(labels).to(device), criterion)
g = np.squeeze(g)
input_data = (input_data-mean) * istd
for i in range(batchsize):
if(options.lossgrad_algorithm=='grad'):
rank_matrix = np.abs(g[i])
elif(options.lossgrad_algorithm=='gradxinp'):
rank_matrix = np.squeeze(g[i]*input_data[i,:,:,:])
elif(options.lossgrad_algorithm=='gradorig'):
rank_matrix = g[i]
v = np.argsort(rank_matrix, axis=None)[-cfg['occlude']:]
input_data[i,:,v//80,v%80] = 0
else:
for i in range(batchsize):
#print('random')
v = np.random.choice(115*80, cfg['occlude'], replace=False)
input_data[i,:,v//80,v%80] = 0
input_data = input_data.astype(floatX)
labels = (0.02 + 0.96*labels)
optimizer.zero_grad()
outputs = mdl(torch.from_numpy(input_data).to(device))
loss = criterion(outputs, torch.from_numpy(labels).to(device))
loss.backward()
optimizer.step()
#print(loss.item())
loss_accum += loss.item()
# - Compute validation loss and error if desired
if options.validate:
#mdl.model_type = 'mel_spects'
from eval import evaluate
mdl.train(False)
val_loss = 0
preds = []
labs = []
max_len = fps
num_iter = 0
for spect, label, g in zip(mel_spects_val, labels_val, loss_grad_val):
num_excerpts = len(spect) - blocklen + 1
excerpts = np.lib.stride_tricks.as_strided(
spect, shape=(num_excerpts, blocklen, spect.shape[1]),
strides=(spect.strides[0], spect.strides[0], spect.strides[1]))
# - Pass mini-batches through the network and concatenate results
for pos in range(0, num_excerpts, batchsize):
input_data = np.transpose(excerpts[pos:pos + batchsize,:,:,np.newaxis],(0,3,1,2))
#if (pos+batchsize>num_excerpts):
# label_batch = label[blocklen//2+pos:blocklen//2+num_excerpts,
# np.newaxis].astype(np.float32)
#else:
# label_batch = label[blocklen//2+pos:blocklen//2+pos+batchsize,
# np.newaxis].astype(np.float32)
if (pos+batchsize>num_excerpts):
label_batch = label[pos:num_excerpts,
np.newaxis].astype(np.float32)
else:
label_batch = label[pos:pos+batchsize,
np.newaxis].astype(np.float32)
#input_data_loss = input_data
if lossgradient!='None':
#grads = loss_grad(mdl_lossgrad, torch.from_numpy(input_data_loss).to(device).requires_grad_(True), torch.from_numpy(label_batch).to(device), criterion)
input_data = (input_data-mean) * istd
for i in range(input_data.shape[0]):
if(options.lossgrad_algorithm=='grad'):
rank_matrix = np.abs(g[i])
elif(options.lossgrad_algorithm=='gradxinp'):
rank_matrix = np.squeeze(g[i]*input_data[i,:,:,:])
elif(options.lossgrad_algorithm=='gradorig'):
rank_matrix = g[i]
v = np.argsort(np.abs(rank_matrix), axis=None)[-cfg['occlude']:]
input_data[i,:,v//80,v%80] = 0
else:
for i in range(input_data.shape[0]):
#print('random')
v = np.random.choice(115*80, cfg['occlude'], replace=False)
input_data[i,:,v//80,v%80] = 0
input_data = input_data.astype(floatX)
pred = mdl(torch.from_numpy(input_data).to(device))
e = criterion(pred,torch.from_numpy(label_batch).to(device))
preds = np.append(preds,pred[:,0].cpu().detach().numpy())
labs = np.append(labs,label_batch)
val_loss +=e.item()
num_iter+=1
#mdl.model_type = 'mel_spects_norm'
print("Validation loss: %.3f" % (val_loss / num_iter))
_, results = evaluate(preds,labs)
print("Validation error: %.3f" % (1 - results['accuracy']))
if(1-results['accuracy']<best_val_error):
torch.save(mdl.state_dict(), os.path.join(modelfile, 'model.pth'))
best_val_loss = val_loss/num_iter
best_val_error = 1-results['accuracy']
print('New saved model',best_val_loss, best_val_error)
#Update the learning rate
scheduler.step()
print('Training Loss per epoch', loss_accum/epochsize)
# - Save parameters for examining
writer.add_scalar('Training Loss',loss_accum/epochsize,epoch)
if(options.validate):
writer.add_scalar('Validation loss', val_loss/num_iter,epoch)
writer.add_scalar('Gradient norm', total_norm, epoch)
writer.add_scalar('Validation error', 1-results['accuracy'])
#for param_group in optimizer.param_groups:
#print(param_group['lr'])
if not options.validate:
torch.save(mdl.state_dict(), os.path.join(modelfile, 'model.pth'))
with io.open(os.path.join(modelfile, 'model.vars'), 'w') as f:
f.writelines('%s=%s\n' % kv for kv in cfg.items())
def main():
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
cfg = {}
print(options.vars)
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
bin_nyquist = frame_len // 2 + 1
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
meanstd_file = os.path.join(os.path.dirname(__file__),
'%s_meanstd.npz' % options.dataset)
if (options.input_type == 'audio'):
dataloader = DatasetLoader(options.dataset,
options.cache_spectra,
datadir,
input_type=options.input_type)
batches = dataloader.prepare_audio_batches(sample_rate, frame_len, fps,
blocklen, batchsize)
else:
dataloader = DatasetLoader(options.dataset,
options.cache_spectra,
datadir,
input_type=options.input_type)
batches = dataloader.prepare_batches(sample_rate, frame_len, fps,
mel_bands, mel_min, mel_max,
blocklen, batchsize)
validation_data = DatasetLoader(options.dataset,
'../ismir2015/experiments/mel_data/',
datadir,
dataset_split='valid',
input_type='mel_spects')
mel_spects_val, labels_val = validation_data.prepare_batches(
sample_rate,
frame_len,
fps,
mel_bands,
mel_min,
mel_max,
blocklen,
batchsize,
batch_data=False)
mdl = model.CNNModel(model_type=options.model_type,
input_type=options.input_type,
is_zeromean=False,
sample_rate=sample_rate,
frame_len=frame_len,
fps=fps,
mel_bands=mel_bands,
mel_min=mel_min,
mel_max=mel_max,
bin_mel_max=bin_mel_max,
meanstd_file=meanstd_file,
device=device)
mdl = mdl.to(device)
#Setting up learning rate and learning rate parameters
initial_eta = cfg['initial_eta']
eta_decay = cfg['eta_decay']
momentum = cfg['momentum']
eta_decay_every = cfg.get('eta_decay_every', 1)
eta = initial_eta
#set up loss
criterion = torch.nn.BCELoss()
#set up optimizer
optimizer = torch.optim.SGD(mdl.parameters(),
lr=eta,
momentum=momentum,
nesterov=True)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=eta_decay_every,
gamma=eta_decay)
#set up optimizer
writer = SummaryWriter(os.path.join(modelfile, 'runs'))
epochs = cfg['epochs']
epochsize = cfg['epochsize']
batches = iter(batches)
#conditions to save model
best_val_loss = 100000.
best_val_error = 1.
for epoch in range(epochs):
# - Initialize certain parameters that are used to monitor training
err = 0
total_norm = 0
loss_accum = 0
mdl.train(True)
# - Compute the L-2 norm of the gradients
for p in mdl.parameters():
if p.grad is not None:
param_norm = p.grad.data.norm(2)
total_norm += param_norm.item()**2
total_norm = total_norm**(1. / 2)
# - Start the training for this epoch
for batch in progress(range(epochsize),
min_delay=0.5,
desc='Epoch %d/%d: Batch ' %
(epoch + 1, epochs)):
data = next(batches)
if (options.input_type == 'audio' or options.input_type == 'stft'):
input_data = data[0]
else:
input_data = np.transpose(data[0][:, :, :, np.newaxis],
(0, 3, 1, 2))
labels = data[1][:, np.newaxis].astype(np.float32)
#map labels to make them softer
if not options.adversarial_training:
labels = (0.02 + 0.96 * labels)
optimizer.zero_grad()
if (options.adversarial_training):
mdl.train(False)
if (options.input_type == 'stft'):
input_data_adv = attacks.PGD(
mdl,
torch.from_numpy(input_data).to(device),
target=torch.from_numpy(labels).to(device),
eps=cfg['eps'],
step_size=cfg['eps_iter'],
iterations=cfg['nb_iter'],
use_best=True,
random_start=True,
clip_min=0,
clip_max=1e8).cpu().detach().numpy()
else:
input_data_adv = attacks.PGD(
mdl,
torch.from_numpy(input_data).to(device),
target=torch.from_numpy(labels).to(device),
eps=cfg['eps'],
step_size=cfg['eps_iter'],
iterations=cfg['nb_iter'],
use_best=True,
random_start=True).cpu().detach().numpy()
mdl.train(True)
optimizer.zero_grad()
outputs = mdl(torch.from_numpy(input_data_adv).to(device))
else:
optimizer.zero_grad()
outputs = mdl(torch.from_numpy(input_data).to(device))
#input(outputs.size())
#input(mdl.conv(torch.from_numpy(input_data).to(device)).cpu().detach().numpy().shape)
loss = criterion(outputs, torch.from_numpy(labels).to(device))
loss.backward()
optimizer.step()
print(loss.item())
loss_accum += loss.item()
# - Compute validation loss and error if desired
if options.validate:
mdl.input_type = 'mel_spects'
from eval import evaluate
mdl.train(False)
val_loss = 0
preds = []
labs = []
max_len = fps
num_iter = 0
for spect, label in zip(mel_spects_val, labels_val):
num_excerpts = len(spect) - blocklen + 1
excerpts = np.lib.stride_tricks.as_strided(
spect,
shape=(num_excerpts, blocklen, spect.shape[1]),
strides=(spect.strides[0], spect.strides[0],
spect.strides[1]))
# - Pass mini-batches through the network and concatenate results
for pos in range(0, num_excerpts, batchsize):
input_data = np.transpose(
excerpts[pos:pos + batchsize, :, :, np.newaxis],
(0, 3, 1, 2))
#if (pos+batchsize>num_excerpts):
# label_batch = label[blocklen//2+pos:blocklen//2+num_excerpts,
# np.newaxis].astype(np.float32)
#else:
# label_batch = label[blocklen//2+pos:blocklen//2+pos+batchsize,
# np.newaxis].astype(np.float32)
if (pos + batchsize > num_excerpts):
label_batch = label[pos:num_excerpts,
np.newaxis].astype(np.float32)
else:
label_batch = label[pos:pos + batchsize,
np.newaxis].astype(np.float32)
pred = mdl(torch.from_numpy(input_data).to(device))
e = criterion(pred,
torch.from_numpy(label_batch).to(device))
preds = np.append(preds, pred[:, 0].cpu().detach().numpy())
labs = np.append(labs, label_batch)
val_loss += e.item()
num_iter += 1
mdl.input_type = options.input_type
print("Validation loss: %.3f" % (val_loss / num_iter))
_, results = evaluate(preds, labs)
print("Validation error: %.3f" % (1 - results['accuracy']))
if (1 - results['accuracy'] < best_val_error):
torch.save(mdl.state_dict(),
os.path.join(modelfile, 'model.pth'))
best_val_loss = val_loss / num_iter
best_val_error = 1 - results['accuracy']
print('New saved model', best_val_loss, best_val_error)
#Update the learning rate
scheduler.step()
print('Training Loss per epoch', loss_accum / epochsize)
# - Save parameters for examining
writer.add_scalar('Training Loss', loss_accum / epochsize, epoch)
writer.add_scalar('Validation loss', val_loss / num_iter, epoch)
writer.add_scalar('Gradient norm', total_norm, epoch)
writer.add_scalar('Validation error', 1 - results['accuracy'])
#for param_group in optimizer.param_groups:
#print(param_group['lr'])
if not options.validate:
torch.save(mdl.state_dict(), os.path.join(modelfile, 'model.pth'))
with io.open(os.path.join(modelfile, 'model.vars'), 'w') as f:
f.writelines('%s=%s\n' % kv for kv in cfg.items())
def main():
# parse command line
parser = opts_parser()
options = parser.parse_args()
outdir = options.outdir
if options.load_spectra != 'memory' and not options.cache_spectra:
parser.error('option --load-spectra=%s requires --cache-spectra' %
options.load_spectra)
# read configuration files and immediate settings
cfg = {}
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
# read some settings into local variables
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
# prepare dataset
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
# - load filelist
filelist = []
ranges = {}
for part in 'train', 'valid', 'test':
a = len(filelist)
with io.open(
os.path.join(datadir, 'filelists',
cfg.get('filelist.%s' % part, part))) as f:
filelist.extend(l.rstrip() for l in f if l.rstrip())
ranges[part] = slice(a, len(filelist))
# - compute spectra
print("Computing%s spectra..." %
(" or loading" if options.cache_spectra else ""))
spects = []
for fn in progress(filelist, 'File '):
cache_fn = (options.cache_spectra
and os.path.join(options.cache_spectra, fn + '.npy'))
spects.append(
cached(cache_fn,
audio.extract_spect,
os.path.join(datadir, 'audio', fn),
sample_rate,
frame_len,
fps,
loading_mode=options.load_spectra))
# - load and convert corresponding labels
print("Loading labels...")
labels = []
for fn, spect in zip(filelist, spects):
fn = os.path.join(datadir, 'labels', fn.rsplit('.', 1)[0] + '.lab')
with io.open(fn) as f:
segments = [l.rstrip().split() for l in f if l.rstrip()]
segments = [(float(start), float(end), label == 'sing')
for start, end, label in segments]
timestamps = np.arange(len(spect)) / float(fps)
labels.append(create_aligned_targets(segments, timestamps, np.bool))
# compute and save different variants of summarized magnitudes
print("Saving files...")
# - ground truth
outfile = os.path.join(outdir, '%s_gt.pkl' % options.dataset)
print(outfile)
with io.open(outfile, 'wb') as f:
pickle.dump({'labels': labels, 'splits': ranges}, f, protocol=-1)
# - summarized spectra
save_spectral_sums(
os.path.join(outdir, '%s_spect_sum.pkl' % options.dataset), spects)
# - summarized mel spectra
bank = audio.create_mel_filterbank(sample_rate, frame_len, mel_bands,
mel_min, mel_max).astype(np.float32)
spects = [np.dot(spect[:, ], bank) for spect in spects]
save_spectral_sums(
os.path.join(outdir, '%s_spect_mel_sum.pkl' % options.dataset), spects)
# - summarized log-mel spectra
spects = [np.log(np.maximum(1e-7, spect)) for spect in spects]
save_spectral_sums(
os.path.join(outdir, '%s_spect_mel_log_sum.pkl' % options.dataset),
spects)
# - summarized standardized log-mel spectra
m, s = znorm.compute_mean_std(spects[ranges['train']], axis=0)
spects = [((spect - m) / s).astype(np.float32) for spect in spects]
save_spectral_sums(
os.path.join(outdir, '%s_spect_mel_log_std_sum.pkl' % options.dataset),
spects)
def main():
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
if options.load_spectra != 'memory' and not options.cache_spectra:
parser.error('option --load-spectra=%s requires --cache-spectra' %
options.load_spectra)
# read configuration files and immediate settings
cfg = {}
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
# read some settings into local variables
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
bin_nyquist = frame_len // 2 + 1
if cfg['filterbank'] == 'mel_learn':
bin_mel_max = bin_nyquist
else:
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
# - load filelist
with io.open(
os.path.join(datadir, 'filelists',
cfg.get('filelist.train', 'train'))) as f:
filelist = [l.rstrip() for l in f if l.rstrip()]
if options.validate:
with io.open(
os.path.join(datadir, 'filelists',
cfg.get('filelist.valid', 'valid'))) as f:
filelist_val = [l.rstrip() for l in f if l.rstrip()]
filelist.extend(filelist_val)
else:
filelist_val = []
# - compute spectra
print("Computing%s spectra..." %
(" or loading" if options.cache_spectra else ""))
spects = []
for fn in progress(filelist, 'File '):
cache_fn = (options.cache_spectra
and os.path.join(options.cache_spectra, fn + '.npy'))
spects.append(
cached(cache_fn,
audio.extract_spect,
os.path.join(datadir, 'audio', fn),
sample_rate,
frame_len,
fps,
loading_mode=options.load_spectra))
# - load and convert corresponding labels
print("Loading labels...")
labels = []
for fn, spect in zip(filelist, spects):
fn = os.path.join(datadir, 'labels', fn.rsplit('.', 1)[0] + '.lab')
with io.open(fn) as f:
segments = [l.rstrip().split() for l in f if l.rstrip()]
segments = [(float(start), float(end), label == 'sing')
for start, end, label in segments]
timestamps = np.arange(len(spect)) / float(fps)
labels.append(create_aligned_targets(segments, timestamps, np.bool))
# - split off validation data, if needed
if options.validate:
spects_val = spects[-len(filelist_val):]
spects = spects[:-len(filelist_val)]
labels_val = labels[-len(filelist_val):]
labels = labels[:-len(filelist_val)]
# - prepare training data generator
print("Preparing training data feed...")
if not options.augment:
# Without augmentation, we just create a generator that returns
# mini-batches of random excerpts
batches = augment.grab_random_excerpts(spects, labels, batchsize,
blocklen, bin_mel_max)
batches = augment.generate_in_background([batches], num_cached=15)
else:
# For time stretching and pitch shifting, it pays off to preapply the
# spline filter to each input spectrogram, so it does not need to be
# applied to each mini-batch later.
spline_order = cfg['spline_order']
if spline_order > 1 and options.load_spectra == 'memory':
from scipy.ndimage import spline_filter
spects = [
spline_filter(spect, spline_order).astype(floatX)
for spect in spects
]
prefiltered = True
else:
prefiltered = False
# We define a function to create the mini-batch generator. This allows
# us to easily create multiple generators for multithreading if needed.
def create_datafeed(spects, labels):
# With augmentation, as we want to apply random time-stretching,
# we request longer excerpts than we finally need to return.
max_stretch = cfg['max_stretch']
batches = augment.grab_random_excerpts(
spects,
labels,
batchsize=batchsize,
frames=int(blocklen / (1 - max_stretch)))
# We wrap the generator in another one that applies random time
# stretching and pitch shifting, keeping a given number of frames
# and bins only.
max_shift = cfg['max_shift']
batches = augment.apply_random_stretch_shift(
batches,
max_stretch,
max_shift,
keep_frames=blocklen,
keep_bins=bin_mel_max,
order=spline_order,
prefiltered=prefiltered)
# We apply random frequency filters
max_db = cfg['max_db']
batches = augment.apply_random_filters(batches, mel_max, max_db)
# We apply random loudness changes
max_loudness = cfg['max_loudness']
if max_loudness:
batches = augment.apply_random_loudness(batches, max_loudness)
return batches
# We start the mini-batch generator and augmenter in one or more
# background threads or processes (unless disabled).
bg_threads = cfg['bg_threads']
bg_processes = cfg['bg_processes']
if not bg_threads and not bg_processes:
# no background processing: just create a single generator
batches = create_datafeed(spects, labels)
elif bg_threads:
# multithreading: create a separate generator per thread
batches = augment.generate_in_background(
[create_datafeed(spects, labels) for _ in range(bg_threads)],
num_cached=bg_threads * 5)
elif bg_processes:
# multiprocessing: single generator is forked along with processes
batches = augment.generate_in_background(
[create_datafeed(spects, labels)] * bg_processes,
num_cached=bg_processes * 25,
in_processes=True)
print("Preparing training function...")
# instantiate neural network
input_var = T.tensor3('input')
inputs = input_var.dimshuffle(0, 'x', 1, 2) # insert "channels" dimension
network = model.architecture(inputs, (None, 1, blocklen, bin_mel_max), cfg)
print(
"- %d layers (%d with weights), %f mio params" %
(len(lasagne.layers.get_all_layers(network)),
sum(hasattr(l, 'W') for l in lasagne.layers.get_all_layers(network)),
lasagne.layers.count_params(network, trainable=True) / 1e6))
print("- weight shapes: %r" % [
l.W.get_value().shape for l in lasagne.layers.get_all_layers(network)
if hasattr(l, 'W') and hasattr(l.W, 'get_value')
])
# create cost expression
target_var = T.vector('targets')
targets = (0.02 + 0.96 * target_var) # map 0 -> 0.02, 1 -> 0.98
targets = targets.dimshuffle(0, 'x') # turn into column vector
outputs = lasagne.layers.get_output(network, deterministic=False)
cost = T.mean(lasagne.objectives.binary_crossentropy(outputs, targets))
if cfg.get('l2_decay', 0):
cost_l2 = lasagne.regularization.regularize_network_params(
network, lasagne.regularization.l2) * cfg['l2_decay']
else:
cost_l2 = 0
# prepare and compile training function
params = lasagne.layers.get_all_params(network, trainable=True)
initial_eta = cfg['initial_eta']
eta_decay = cfg['eta_decay']
eta_decay_every = cfg.get('eta_decay_every', 1)
patience = cfg.get('patience', 0)
trials_of_patience = cfg.get('trials_of_patience', 1)
patience_criterion = cfg.get(
'patience_criterion',
'valid_loss' if options.validate else 'train_loss')
momentum = cfg['momentum']
first_params = params[:cfg['first_params']]
first_params_eta_scale = cfg['first_params_eta_scale']
if cfg['learn_scheme'] == 'nesterov':
learn_scheme = lasagne.updates.nesterov_momentum
elif cfg['learn_scheme'] == 'momentum':
learn_scheme = lasagne.update.momentum
elif cfg['learn_scheme'] == 'adam':
learn_scheme = lasagne.updates.adam
else:
raise ValueError('Unknown learn_scheme=%s' % cfg['learn_scheme'])
eta = theano.shared(lasagne.utils.floatX(initial_eta))
if not first_params or first_params_eta_scale == 1:
updates = learn_scheme(cost + cost_l2, params, eta, momentum)
else:
grads = theano.grad(cost + cost_l2, params)
updates = learn_scheme(grads[len(first_params):],
params[len(first_params):], eta, momentum)
if first_params_eta_scale > 0:
updates.update(
learn_scheme(grads[:len(first_params)], first_params,
eta * first_params_eta_scale, momentum))
print("Compiling training function...")
train_fn = theano.function([input_var, target_var], cost, updates=updates)
# prepare and compile validation function, if requested
if options.validate:
print("Compiling validation function...")
import model_to_fcn
network_test = model_to_fcn.model_to_fcn(network, allow_unlink=False)
outputs_test = lasagne.layers.get_output(network_test,
deterministic=True)
cost_test = T.mean(
lasagne.objectives.binary_crossentropy(outputs_test, targets))
val_fn = theano.function([input_var, target_var],
[cost_test, outputs_test])
# run training loop
print("Training:")
epochs = cfg['epochs']
epochsize = cfg['epochsize']
batches = iter(batches)
if options.save_errors:
errors = []
if first_params and cfg['first_params_log']:
first_params_hist = []
if patience > 0:
best_error = np.inf
best_state = get_state(network, updates)
for epoch in range(epochs):
# actual training
err = 0
for batch in progress(range(epochsize),
min_delay=.5,
desc='Epoch %d/%d: Batch ' %
(epoch + 1, epochs)):
err += train_fn(*next(batches))
if not np.isfinite(err):
print("\nEncountered NaN loss in training. Aborting.")
sys.exit(1)
if first_params and cfg['first_params_log'] and (
batch % cfg['first_params_log'] == 0):
first_params_hist.append(
tuple(param.get_value() for param in first_params))
np.savez(
modelfile[:-4] + '.hist.npz', **{
'param%d' % i: param
for i, param in enumerate(zip(*first_params_hist))
})
# report training loss
print("Train loss: %.3f" % (err / epochsize))
if options.save_errors:
errors.append(err / epochsize)
# compute and report validation loss, if requested
if options.validate:
val_err = 0
preds = []
max_len = int(fps * cfg.get('val.max_len', 30))
for spect, label in zip(spects_val, labels_val):
# pick excerpt of val.max_len seconds in center of file
excerpt = slice(max(0, (len(spect) - max_len) // 2),
(len(spect) + max_len) // 2)
# crop to maximum length and required spectral bins
spect = spect[None, excerpt, :bin_mel_max]
# crop to maximum length and remove edges lost in the network
label = label[excerpt][blocklen // 2:-(blocklen // 2)]
e, pred = val_fn(spect, label)
val_err += e
preds.append((pred[:, 0], label))
print("Validation loss: %.3f" % (val_err / len(filelist_val)))
from eval import evaluate
_, results = evaluate(*zip(*preds))
print("Validation error: %.3f" % (1 - results['accuracy']))
if options.save_errors:
errors.append(val_err / len(filelist_val))
errors.append(1 - results['accuracy'])
# update learning rate and/or apply early stopping, if needed
if patience > 0:
if patience_criterion == 'train_loss':
cur_error = err / epochsize
elif patience_criterion == 'valid_loss':
cur_error = val_err / len(filelist_val)
elif patience_criterion == 'valid_error':
cur_error = 1 - results['accuracy']
if cur_error <= best_error:
best_error = cur_error
best_state = get_state(network, updates)
patience = cfg['patience']
else:
patience -= 1
if patience == 0:
if eta_decay_every == 'trial_of_patience' and eta_decay != 1:
eta.set_value(eta.get_value() *
lasagne.utils.floatX(eta_decay))
restore_state(network, updates, best_state)
patience = cfg['patience']
trials_of_patience -= 1
print("Lost patience (%d remaining trials)." %
trials_of_patience)
if trials_of_patience == 0:
break
if eta_decay_every != 'trial_of_patience' and eta_decay != 1 and \
(epoch + 1) % eta_decay_every == 0:
eta.set_value(eta.get_value() * lasagne.utils.floatX(eta_decay))
# save final network
print("Saving final model")
np.savez(
modelfile, **{
'param%d' % i: p
for i, p in enumerate(lasagne.layers.get_all_param_values(network))
})
with io.open(modelfile + '.vars', 'wb') as f:
f.writelines('%s=%s\n' % kv for kv in cfg.items())
if options.save_errors:
np.savez(modelfile[:-len('.npz')] + '.err.npz',
np.asarray(errors).reshape(epoch + 1, -1))
def main():
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
cfg = {}
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
bin_nyquist = frame_len // 2 + 1
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
# - load filelist
with io.open(os.path.join(datadir, 'filelists', 'train')) as f:
filelist = [l.rstrip() for l in f if l.rstrip()]
if options.validate:
with io.open(os.path.join(datadir, 'filelists', 'valid')) as f:
filelist_val = [l.strip() for l in f if l.strip()]
filelist.extend(filelist_val)
else:
filelist_val = []
# - compute spectra
print("Computing%s spectra..." %
(" or loading" if options.cache_spectra else ""))
spects = []
for fn in progress(filelist, 'File '):
cache_fn = (options.cache_spectra
and os.path.join(options.cache_spectra, fn + '.npy'))
spects.append(
cached(cache_fn, audio.extract_spect,
os.path.join(datadir, 'audio', fn), sample_rate, frame_len,
fps))
# - load and convert corresponding labels
print("Loading labels...")
labels = []
for fn, spect in zip(filelist, spects):
fn = os.path.join(datadir, 'labels', fn.rsplit('.', 1)[0] + '.lab')
with io.open(fn) as f:
segments = [l.rstrip().split() for l in f if l.rstrip()]
segments = [(float(start), float(end), label == 'sing')
for start, end, label in segments]
timestamps = np.arange(len(spect)) / float(fps)
labels.append(create_aligned_targets(segments, timestamps, np.bool))
# - prepare mel filterbank
filterbank = audio.create_mel_filterbank(sample_rate, frame_len, mel_bands,
mel_min, mel_max)
filterbank = filterbank[:bin_mel_max].astype(floatX)
if options.validate:
spects_val = spects[-len(filelist_val):]
spects = spects[:-len(filelist_val)]
labels_val = labels[-len(filelist_val):]
labels = labels[:-len(filelist_val)]
# - precompute mel spectra, if needed, otherwise just define a generator
mel_spects = (np.log(
np.maximum(np.dot(spect[:, :bin_mel_max], filterbank), 1e-7))
for spect in spects)
if not options.augment:
mel_spects = list(mel_spects)
del spects
# - load mean/std or compute it, if not computed yet
meanstd_file = os.path.join(os.path.dirname(__file__),
'%s_meanstd.npz' % options.dataset)
try:
with np.load(meanstd_file) as f:
mean = f['mean']
std = f['std']
except (IOError, KeyError):
print("Computing mean and standard deviation...")
mean, std = znorm.compute_mean_std(mel_spects)
np.savez(meanstd_file, mean=mean, std=std)
mean = mean.astype(floatX)
istd = np.reciprocal(std).astype(floatX)
# - prepare training data generator
print("Preparing training data feed...")
if not options.augment:
# Without augmentation, we just precompute the normalized mel spectra
# and create a generator that returns mini-batches of random excerpts
mel_spects = [(spect - mean) * istd for spect in mel_spects]
batches = augment.grab_random_excerpts(mel_spects, labels, batchsize,
blocklen)
else:
# For time stretching and pitch shifting, it pays off to preapply the
# spline filter to each input spectrogram, so it does not need to be
# applied to each mini-batch later.
spline_order = cfg['spline_order']
if spline_order > 1:
from scipy.ndimage import spline_filter
spects = [
spline_filter(spect, spline_order).astype(floatX)
for spect in spects
]
# We define a function to create the mini-batch generator. This allows
# us to easily create multiple generators for multithreading if needed.
def create_datafeed(spects, labels):
# With augmentation, as we want to apply random time-stretching,
# we request longer excerpts than we finally need to return.
max_stretch = cfg['max_stretch']
batches = augment.grab_random_excerpts(
spects,
labels,
batchsize=batchsize,
frames=int(blocklen / (1 - max_stretch)))
# We wrap the generator in another one that applies random time
# stretching and pitch shifting, keeping a given number of frames
# and bins only.
max_shift = cfg['max_shift']
batches = augment.apply_random_stretch_shift(batches,
max_stretch,
max_shift,
keep_frames=blocklen,
keep_bins=bin_mel_max,
order=spline_order,
prefiltered=True)
# We transform the excerpts to mel frequency and log magnitude.
batches = augment.apply_filterbank(batches, filterbank)
batches = augment.apply_logarithm(batches)
# We apply random frequency filters
max_db = cfg['max_db']
batches = augment.apply_random_filters(batches,
filterbank,
mel_max,
max_db=max_db)
# We apply normalization
batches = augment.apply_znorm(batches, mean, istd)
return batches
# We start the mini-batch generator and augmenter in one or more
# background threads or processes (unless disabled).
bg_threads = cfg['bg_threads']
bg_processes = cfg['bg_processes']
if not bg_threads and not bg_processes:
# no background processing: just create a single generator
batches = create_datafeed(spects, labels)
elif bg_threads:
# multithreading: create a separate generator per thread
batches = augment.generate_in_background(
[create_datafeed(spects, labels) for _ in range(bg_threads)],
num_cached=bg_threads * 5)
elif bg_processes:
# multiprocessing: single generator is forked along with processes
batches = augment.generate_in_background(
[create_datafeed(spects, labels)] * bg_processes,
num_cached=bg_processes * 25,
in_processes=True)
###########################################################################
#-----------Main changes to code to make it work with pytorch-------------#
###########################################################################
print("preparing training function...")
mdl = model.CNNModel()
mdl = mdl.to(device)
#Setting up learning rate and learning rate parameters
initial_eta = cfg['initial_eta']
eta_decay = cfg['eta_decay']
momentum = cfg['momentum']
eta_decay_every = cfg.get('eta_decay_every', 1)
eta = initial_eta
#set up loss
criterion = torch.nn.BCELoss()
#set up optimizer
optimizer = torch.optim.SGD(mdl.parameters(),
lr=eta,
momentum=momentum,
nesterov=True)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=eta_decay_every,
gamma=eta_decay)
#set up optimizer
writer = SummaryWriter(os.path.join(modelfile, 'runs'))
epochs = cfg['epochs']
epochsize = cfg['epochsize']
batches = iter(batches)
#conditions to save model
best_val_loss = 100000.
best_val_error = 1.
for epoch in range(epochs):
# - Initialize certain parameters that are used to monitor training
err = 0
total_norm = 0
loss_accum = 0
mdl.train(True)
# - Compute the L-2 norm of the gradients
for p in mdl.parameters():
if p.grad is not None:
param_norm = p.grad.data.norm(2)
total_norm += param_norm.item()**2
total_norm = total_norm**(1. / 2)
# - Start the training for this epoch
for batch in progress(range(epochsize),
min_delay=0.5,
desc='Epoch %d/%d: Batch ' %
(epoch + 1, epochs)):
data = next(batches)
input_data = np.transpose(data[0][:, :, :, np.newaxis],
(0, 3, 1, 2))
labels = data[1][:, np.newaxis].astype(np.float32)
#map labels to make them softer
labels = (0.02 + 0.96 * labels)
optimizer.zero_grad()
outputs = mdl(torch.from_numpy(input_data).to(device))
loss = criterion(outputs, torch.from_numpy(labels).to(device))
loss.backward()
optimizer.step()
loss_accum += loss.item()
# - Compute validation loss and error if desired
if options.validate:
from eval import evaluate
mdl.train(False)
val_loss = 0
preds = []
labs = []
max_len = fps
mel_spects_val = (np.log(
np.maximum(np.dot(spect[:, :bin_mel_max], filterbank), 1e-7))
for spect in spects_val)
mel_spects_val = [(spect - mean) * istd
for spect in mel_spects_val]
num_iter = 0
for spect, label in zip(mel_spects_val, labels_val):
num_excerpts = len(spect) - blocklen + 1
excerpts = np.lib.stride_tricks.as_strided(
spect,
shape=(num_excerpts, blocklen, spect.shape[1]),
strides=(spect.strides[0], spect.strides[0],
spect.strides[1]))
# - Pass mini-batches through the network and concatenate results
for pos in range(0, num_excerpts, batchsize):
input_data = np.transpose(
excerpts[pos:pos + batchsize, :, :, np.newaxis],
(0, 3, 1, 2))
if (pos + batchsize > num_excerpts):
label_batch = label[blocklen // 2 + pos:blocklen // 2 +
num_excerpts,
np.newaxis].astype(np.float32)
else:
label_batch = label[blocklen // 2 + pos:blocklen // 2 +
pos + batchsize,
np.newaxis].astype(np.float32)
pred = mdl(torch.from_numpy(input_data).to(device))
e = criterion(pred,
torch.from_numpy(label_batch).to(device))
preds = np.append(preds, pred[:, 0].cpu().detach().numpy())
labs = np.append(labs, label_batch)
val_loss += e.item()
num_iter += 1
print("Validation loss: %.3f" % (val_loss / num_iter))
_, results = evaluate(preds, labs)
print("Validation error: %.3f" % (1 - results['accuracy']))
if (val_loss / num_iter < best_val_loss
and (1 - results['accuracy']) < best_val_error):
torch.save(mdl.state_dict(),
os.path.join(modelfile, 'model.pth'))
best_val_loss = val_loss / num_iter
best_val_error = 1 - results['accuracy']
print('New saved model', best_val_loss, best_val_error)
#Update the learning rate
scheduler.step()
print('Training Loss per epoch', loss_accum / epochsize)
# - Save parameters for examining
writer.add_scalar('Training Loss', loss_accum / epochsize, epoch)
writer.add_scalar('Validation loss', val_loss / num_iter, epoch)
writer.add_scalar('Gradient norm', total_norm, epoch)
writer.add_scalar('Validation error', 1 - results['accuracy'])
for param_group in optimizer.param_groups:
print(param_group['lr'])
if not options.validate:
torch.save(mdl.state_dict(), os.path.join(modelfile, 'model.pth'))
def main():
print(torch.cuda.is_available())
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# parse command line
parser = opts_parser()
options = parser.parse_args()
modelfile = options.modelfile
lossgradient = options.lossgradient
cfg = {}
for fn in options.vars:
cfg.update(config.parse_config_file(fn))
cfg.update(config.parse_variable_assignments(options.var))
outfile = options.outfile
sample_rate = cfg['sample_rate']
frame_len = cfg['frame_len']
fps = cfg['fps']
mel_bands = cfg['mel_bands']
mel_min = cfg['mel_min']
mel_max = cfg['mel_max']
blocklen = cfg['blocklen']
batchsize = cfg['batchsize']
bin_nyquist = frame_len // 2 + 1
bin_mel_max = bin_nyquist * 2 * mel_max // sample_rate
# prepare dataset
print("Preparing data reading...")
datadir = os.path.join(os.path.dirname(__file__), os.path.pardir,
'datasets', options.dataset)
# - load filelist
with io.open(os.path.join(datadir, 'filelists', 'valid')) as f:
filelist = [l.rstrip() for l in f if l.rstrip()]
with io.open(os.path.join(datadir, 'filelists', 'test')) as f:
filelist += [l.rstrip() for l in f if l.rstrip()]
# - load mean/std
meanstd_file = os.path.join(os.path.dirname(__file__),
'%s_meanstd.npz' % options.dataset)
dataloader = DatasetLoader(options.dataset,
options.cache_spectra,
datadir,
input_type=options.input_type,
filelist=filelist)
mel_spects, labels = dataloader.prepare_batches(sample_rate,
frame_len,
fps,
mel_bands,
mel_min,
mel_max,
blocklen,
batchsize,
batch_data=False)
with np.load(meanstd_file) as f:
mean = f['mean']
std = f['std']
mean = mean.astype(floatX)
istd = np.reciprocal(std).astype(floatX)
mdl = model.CNNModel(input_type='mel_spects_norm',
is_zeromean=False,
meanstd_file=meanstd_file,
device=device)
mdl.load_state_dict(torch.load(modelfile))
mdl.to(device)
mdl.eval()
if (lossgradient != 'None'):
mdl_lossgrad = model.CNNModel(input_type=options.input_type,
is_zeromean=False,
sample_rate=sample_rate,
frame_len=frame_len,
fps=fps,
mel_bands=mel_bands,
mel_min=mel_min,
mel_max=mel_max,
bin_mel_max=bin_mel_max,
meanstd_file=meanstd_file,
device=device)
mdl_lossgrad.load_state_dict(torch.load(lossgradient))
mdl_lossgrad.to(device)
mdl_lossgrad.eval()
criterion = torch.nn.BCELoss()
loss_grad_val = dataloader.prepare_loss_grad_batches(
options.loss_grad_save, mel_spects, labels, mdl_lossgrad,
criterion, blocklen, batchsize, device)
# run prediction loop
print("Predicting:")
predictions = []
#for spect, g in zip(mel_spects, loss_grad_val):
c = 0
for spect in progress(mel_spects, total=len(filelist), desc='File '):
if (lossgradient != 'None'):
g = loss_grad_val[c]
c += 1
# naive way: pass excerpts of the size used during training
# - view spectrogram memory as a 3-tensor of overlapping excerpts
num_excerpts = len(spect) - blocklen + 1
excerpts = np.lib.stride_tricks.as_strided(
spect.astype(floatX),
shape=(num_excerpts, blocklen, spect.shape[1]),
strides=(spect.strides[0], spect.strides[0], spect.strides[1]))
preds = np.zeros((num_excerpts, 1))
count = 0
for pos in range(0, num_excerpts, batchsize):
input_data = np.transpose(
excerpts[pos:pos + batchsize, :, :, np.newaxis], (0, 3, 1, 2))
input_data = (input_data - mean) * istd
if lossgradient != 'None':
for i in range(input_data.shape[0]):
if (options.lossgrad_algorithm == 'grad'):
rank_matrix = np.abs(g[i + pos])
elif (options.lossgrad_algorithm == 'gradxinp'):
rank_matrix = np.squeeze(g[i + pos] *
input_data[i, :, :, :])
elif (options.lossgrad_algorithm == 'gradorig'):
rank_matrix = g[i + pos]
if (options.ROAR == 1):
v = np.argsort(rank_matrix,
axis=None)[-cfg['occlude']:]
else:
v = np.argsort(rank_matrix, axis=None)[:cfg['occlude']]
input_data[i, :, v // 80, v % 80] = 0
else:
for i in range(input_data.shape[0]):
#print('random')
v = np.random.choice(115 * 80,
cfg['occlude'],
replace=False)
input_data[i, :, v // 80, v % 80] = 0
count += 1
#print('Here')
#preds = np.vstack(mdl.forward(torch.from_numpy(
# np.transpose(excerpts[pos:pos + batchsize,:,:,
# np.newaxis],(0,3,1,2))).to(device)).cpu().detach().numpy()
# for pos in range(0, num_excerpts, batchsize))
preds[pos:pos + batchsize, :] = mdl(
torch.from_numpy(input_data).to(
device)).cpu().detach().numpy()
print('Here')
predictions.append(preds)
# save predictions
print("Saving predictions")
np.savez(outfile, **{fn: pred for fn, pred in zip(filelist, predictions)})