def main():
data_loader = DataLoader(args)
if (args.process_data):
data_loader.process_data()
return
torch.cuda.set_device(args.gpu_device)
data_loader.load()
if (args.model == "rnn"):
myModel = model.RNNModel(args, data_loader.vocab_size, 8,
data_loader.id_2_vec).cuda()
elif (args.model == "cnn"):
myModel = model.CNNModel(args, data_loader.vocab_size, 8,
data_loader.id_2_vec).cuda()
elif (args.model == "baseline"):
myModel = model.Baseline(args, data_loader.vocab_size, 8,
data_loader.id_2_vec).cuda()
else:
print("invalid model type")
exit(1)
if (args.test_only):
test(myModel, data_loader, args)
else:
train(myModel, data_loader, args)
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():
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():
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)})
