for t in targets:
# search for adversary
X_adv, P_adv = find_adversary(
model=dnn_model,
X0=X0,
label=t,
P0=Phs,
mu=mu,
epsilon=epsilon,
maxits=maxits,
stop_thresh=stop_thresh,
griffin_lim=True)
# get time-domain representation
x_adv = overlap_add( np.hstack((X_adv, X_adv[:,-2:-nfft/2-1:-1])) * np.exp(1j*P_adv))
out_snr = 20*np.log10(np.linalg.norm(x[nfft:-nfft]) / np.linalg.norm(x[nfft:-nfft]-x_adv[nfft:-nfft]))
# BEFORE FILTERING
# ===========================================
# dnn prediction
pred = np.argmax(np.sum(fprop(X_adv), axis=0))
if pred == t:
dnn_file.write('{}\t'.format(int(out_snr+.5)))
else:
dnn_file.write('{}\t'.format('na'))
# aux prediction
X_adv_agg = aggregate_features(dnn_model, X_adv, which_layers)
pred = np.argmax(np.bincount(np.array(aux_model.predict(X_adv_agg), dtype='int')))
if pred == t:
def file_misclass_error_printf(dnn_model,
root_dir,
dataset,
save_file,
mode='all_same',
label=0,
snr=30,
aux_model=None,
aux_save_file=None,
which_layers=None,
save_adversary_audio=None,
fwd_xform=None,
back_xform=None):
"""
Function to compute the file-level classification error by classifying
individual frames and then voting for the class with highest cumulative probability
"""
if fwd_xform is None:
print 'fwd_xform=None, using identity'
fwd_xform = lambda X: X
if back_xform is None:
print 'back_xform=None, using identity'
back_xform = lambda X: X
n_classes = len(dataset.targets)
X = dnn_model.get_input_space().make_theano_batch()
Y = dnn_model.fprop(X)
fprop_theano = theano.function([X], Y)
input_space = dnn_model.get_input_space()
if isinstance(input_space, Conv2DSpace):
tframes, dim = input_space.shape
view_converter = DefaultViewConverter((tframes, dim, 1))
else:
dim = input_space.dim
tframes = 1
view_converter = None
if view_converter is not None:
def fprop(batch):
nframes = batch.shape[0]
thop = 1.
sup = np.arange(0, nframes - tframes + 1, np.int(tframes / thop))
data = np.vstack([
np.reshape(batch[i:i + tframes, :], (tframes * dim, ))
for i in sup
])
data = fwd_xform(data)
return fprop_theano(
view_converter.get_formatted_batch(data, input_space))
else:
fprop = fprop_theano
n_examples = len(dataset.file_list)
target_space = dnn_model.get_output_space() #VectorSpace(dim=n_classes)
feat_space = dnn_model.get_input_space(
) #VectorSpace(dim=dataset.nfft//2+1, dtype='complex64')
data_specs = (CompositeSpace(
(feat_space, target_space)), ("songlevel-features", "targets"))
iterator = dataset.iterator(mode='sequential',
batch_size=1,
data_specs=data_specs)
if aux_model:
aux_fname = open(aux_save_file, 'w')
aux_writer = csv.writer(aux_fname, delimiter='\t')
with open(save_file, 'w') as fname:
dnn_writer = csv.writer(fname, delimiter='\t')
for i, el in enumerate(iterator):
# display progress indicator
'Progress: %2.0f%%\r' % (100 * i / float(n_examples))
Mag, Phs = np.abs(el[0], dtype=np.float32), np.angle(el[0])
epsilon = np.linalg.norm(Mag) / Mag.shape[0] / 10**(snr / 20.)
if mode == 'all_same':
target = label
elif mode == 'perfect':
target = el[1]
elif mode == 'random':
target = np.random.randint(n_classes)
elif mode == 'all_wrong':
cand = np.setdiff1d(np.arange(n_classes), np.array(
el[1])) # remove ground truth label from set of options
target = cand[np.random.randint(len(cand))]
if 1: # re-read audio (seems to be bug when reading from h5)
f = el[2]
if f.endswith('.wav'):
read_fun = audiolab.wavread
elif f.endswith('.au'):
read_fun = audiolab.auread
elif f.endswith('.mp3'):
read_fun = read_mp3
x, fstmp, _ = read_fun(os.path.join(root_dir, f))
# make mono
if len(x.shape) != 1:
x = np.sum(x, axis=1) / 2.
seglen = 30
x = x[:fstmp * seglen]
fs = 22050
if fstmp != fs:
x = samplerate.resample(x, fs / float(fstmp), 'sinc_best')
Mag, Phs = compute_fft(x)
Mag = Mag[:1200, :513]
Phs = Phs[:1200, :513]
epsilon = np.linalg.norm(Mag) / Mag.shape[0] / 10**(snr / 20.)
else:
raise ValueError(
"Check that song-level iterator is indeed returning 'raw data'"
)
X_adv, P_adv = find_adversary(
model=dnn_model,
X0=Mag,
label=target,
fwd_xform=fwd_xform,
back_xform=back_xform,
P0=np.hstack((Phs, -Phs[:, -2:-dataset.nfft / 2 - 1:-1])),
mu=.01,
epsilon=epsilon,
maxits=50,
stop_thresh=0.5,
griffin_lim=False #True
)
if save_adversary_audio:
nfft = 2 * (X_adv.shape[1] - 1)
nhop = nfft // 2
x_adv = overlap_add(
np.hstack((X_adv, X_adv[:, -2:-nfft // 2 - 1:-1])) *
np.exp(1j * P_adv), nfft, nhop)
audiolab.wavwrite(x_adv,
os.path.join(save_adversary_audio, el[2]),
22050, 'pcm16')
#frame_labels = np.argmax(fprop(X_adv), axis=1)
#hist = np.bincount(frame_labels, minlength=n_classes)
fpass = fprop(X_adv)
conf = np.sum(fpass, axis=0) / float(fpass.shape[0])
dnn_label = np.argmax(conf) #np.argmax(hist) # most used label
true_label = el[1]
# truncate to correct length
ext = min(Mag.shape[0], X_adv.shape[0])
Mag = Mag[:ext, :]
X_adv = X_adv[:ext, :]
X_diff = Mag - X_adv
out_snr = 20 * np.log10(
np.linalg.norm(Mag) / np.linalg.norm(X_diff))
dnn_writer.writerow([
dataset.file_list[i], true_label, dnn_label, out_snr,
conf[dnn_label]
])
print 'Mode:{}, True label:{}, Adv label:{}, Sel label:{}, Conf:{}, Out snr: {}'.format(
mode, true_label, target, dnn_label, conf[dnn_label], out_snr)
if aux_model:
fft_agg = aggregate_features(dnn_model, X_adv, which_layers)
aux_vote = np.argmax(
np.bincount(
np.array(aux_model.predict(fft_agg), dtype='int')))
aux_writer.writerow(
[dataset.file_list[i], true_label, aux_vote])
print 'AUX adversarial label: {}'.format(aux_vote)
if aux_model:
aux_fname.close()
print ''
for t in targets:
# search for adversary
X_adv, P_adv = find_adversary(model=dnn_model,
X0=X0,
label=t,
P0=Phs,
mu=mu,
epsilon=epsilon,
maxits=maxits,
stop_thresh=stop_thresh,
griffin_lim=True)
# get time-domain representation
x_adv = overlap_add(
np.hstack((X_adv, X_adv[:, -2:-nfft / 2 - 1:-1])) *
np.exp(1j * P_adv))
out_snr = 20 * np.log10(
np.linalg.norm(x[nfft:-nfft]) /
np.linalg.norm(x[nfft:-nfft] - x_adv[nfft:-nfft]))
# BEFORE FILTERING
# ===========================================
# dnn prediction
pred = np.argmax(np.sum(fprop(X_adv), axis=0))
if pred == t:
dnn_file.write('{}\t'.format(int(out_snr + .5)))
else:
dnn_file.write('{}\t'.format('na'))
# aux prediction
def file_misclass_error_printf(dnn_model, root_dir, dataset, save_file, mode='all_same', label=0, snr=30, aux_model=None, aux_save_file=None, which_layers=None, save_adversary_audio=None, fwd_xform=None, back_xform=None):
"""
Function to compute the file-level classification error by classifying
individual frames and then voting for the class with highest cumulative probability
"""
if fwd_xform is None:
print 'fwd_xform=None, using identity'
fwd_xform = lambda X: X
if back_xform is None:
print 'back_xform=None, using identity'
back_xform = lambda X: X
n_classes = len(dataset.targets)
X = dnn_model.get_input_space().make_theano_batch()
Y = dnn_model.fprop(X)
fprop_theano = theano.function([X],Y)
input_space = dnn_model.get_input_space()
if isinstance(input_space, Conv2DSpace):
tframes, dim = input_space.shape
view_converter = DefaultViewConverter((tframes, dim, 1))
else:
dim = input_space.dim
tframes = 1
view_converter = None
if view_converter is not None:
def fprop(batch):
nframes = batch.shape[0]
thop = 1.
sup = np.arange(0,nframes-tframes+1, np.int(tframes/thop))
data = np.vstack([np.reshape(batch[i:i+tframes, :],(tframes*dim,)) for i in sup])
data = fwd_xform(data)
return fprop_theano(view_converter.get_formatted_batch(data, input_space))
else:
fprop = fprop_theano
n_examples = len(dataset.file_list)
target_space = dnn_model.get_output_space() #VectorSpace(dim=n_classes)
feat_space = dnn_model.get_input_space() #VectorSpace(dim=dataset.nfft//2+1, dtype='complex64')
data_specs = (CompositeSpace((feat_space, target_space)), ("songlevel-features", "targets"))
iterator = dataset.iterator(mode='sequential', batch_size=1, data_specs=data_specs)
if aux_model:
aux_fname = open(aux_save_file, 'w')
aux_writer = csv.writer(aux_fname, delimiter='\t')
with open(save_file, 'w') as fname:
dnn_writer = csv.writer(fname, delimiter='\t')
for i,el in enumerate(iterator):
# display progress indicator
'Progress: %2.0f%%\r' % (100*i/float(n_examples))
Mag, Phs = np.abs(el[0], dtype=np.float32), np.angle(el[0])
epsilon = np.linalg.norm(Mag)/Mag.shape[0]/10**(snr/20.)
if mode == 'all_same':
target = label
elif mode == 'perfect':
target = el[1]
elif mode == 'random':
target = np.random.randint(n_classes)
elif mode == 'all_wrong':
cand = np.setdiff1d(np.arange(n_classes),np.array(el[1])) # remove ground truth label from set of options
target = cand[np.random.randint(len(cand))]
if 1: # re-read audio (seems to be bug when reading from h5)
f = el[2]
if f.endswith('.wav'):
read_fun = audiolab.wavread
elif f.endswith('.au'):
read_fun = audiolab.auread
elif f.endswith('.mp3'):
read_fun = read_mp3
x, fstmp, _ = read_fun(os.path.join(root_dir, f))
# make mono
if len(x.shape) != 1:
x = np.sum(x, axis=1)/2.
seglen=30
x = x[:fstmp*seglen]
fs = 22050
if fstmp != fs:
x = samplerate.resample(x, fs/float(fstmp), 'sinc_best')
Mag, Phs = compute_fft(x)
Mag = Mag[:1200,:513]
Phs = Phs[:1200,:513]
epsilon = np.linalg.norm(Mag)/Mag.shape[0]/10**(snr/20.)
else:
raise ValueError("Check that song-level iterator is indeed returning 'raw data'")
X_adv, P_adv = find_adversary(
model=dnn_model,
X0=Mag,
label=target,
fwd_xform=fwd_xform,
back_xform=back_xform,
P0=np.hstack((Phs, -Phs[:,-2:-dataset.nfft/2-1:-1])),
mu=.01,
epsilon=epsilon,
maxits=50,
stop_thresh=0.5,
griffin_lim=False#True
)
if save_adversary_audio:
nfft = 2*(X_adv.shape[1]-1)
nhop = nfft//2
x_adv = overlap_add(np.hstack((X_adv, X_adv[:,-2:-nfft//2-1:-1])) * np.exp(1j*P_adv), nfft, nhop)
audiolab.wavwrite(x_adv, os.path.join(save_adversary_audio, el[2]), 22050, 'pcm16')
#frame_labels = np.argmax(fprop(X_adv), axis=1)
#hist = np.bincount(frame_labels, minlength=n_classes)
fpass = fprop(X_adv)
conf = np.sum(fpass, axis=0) / float(fpass.shape[0])
dnn_label = np.argmax(conf) #np.argmax(hist) # most used label
true_label = el[1]
# truncate to correct length
ext = min(Mag.shape[0], X_adv.shape[0])
Mag = Mag[:ext,:]
X_adv = X_adv[:ext,:]
X_diff = Mag-X_adv
out_snr = 20*np.log10(np.linalg.norm(Mag)/np.linalg.norm(X_diff))
dnn_writer.writerow([dataset.file_list[i], true_label, dnn_label, out_snr, conf[dnn_label]])
print 'Mode:{}, True label:{}, Adv label:{}, Sel label:{}, Conf:{}, Out snr: {}'.format(mode, true_label, target, dnn_label, conf[dnn_label], out_snr)
if aux_model:
fft_agg = aggregate_features(dnn_model, X_adv, which_layers)
aux_vote = np.argmax(np.bincount(np.array(aux_model.predict(fft_agg), dtype='int')))
aux_writer.writerow([dataset.file_list[i], true_label, aux_vote])
print 'AUX adversarial label: {}'.format(aux_vote)
if aux_model:
aux_fname.close()
print ''