def save(self, signal, target_path, orig_path, lr, iterations, i, bandwidth):
outprefix = "generated_stage2"
if not os.path.exists(outprefix):
os.makedirs(outprefix)
signal = signal.unsqueeze(0)
id_original = self.audio_path.split("/")[-1].split('_')[0]
id_target = target_path.split("/")[-1].split('.')[0]
save_path = os.path.join(outprefix, id_original + '_to_' + id_target + "_final" + ".wav")
torchaudio.save(save_path, signal, sample_rate=16000, precision=32)
print("Adversarial audio has been saved to " + save_path + "!")
return save_path
self.activations = activations
def __lt__(self, other):
return torch.sum(self.activations, dim=0) < torch.sum(
other.activations, dim=0)
import heapq
from collections import defaultdict
### Maximum Activations ###
d = defaultdict(list)
attr_algo = LayerActivation(model, modules["conv8"])
for signal in tqdm(X):
# Get neuron activation
attributions = attr_algo.attribute(signal.unsqueeze(0).cuda(),
attribute_to_layer_input=False)
# print(attributions.shape)
for i, act in enumerate(attributions[0]):
heapq.heappush(d[i], neuron(signal, act))
if len(d[i]) > 5:
d[i].pop(-1)
# %%
from ipywidgets import interact_manual, widgets
def get_top_activations_spectogram_for_unit(channel):
# for u in d[]:
place = 1
# plt.figure(place)
def analyze_ood(signal, range):
output = signal.unsqueeze(0)
for i in modules:
layer = modules[i]
layer.eval().to(device)
output = layer(output)
features = torch.flatten(output, start_dim=1)
# print(features.shape)
coordinates = pca.transform(features.cpu().detach().numpy())
plotly.offline.init_notebook_mode()
# print(coordinates)
new_df = pd.DataFrame({
'X': [coordinates[0][0]],
'Y': [coordinates[0][1]],
'Z': [coordinates[0][2]],
'label': [200]
})
new_df = df.append(new_df)
# Configure the trace.
trace = px.scatter_3d(
new_df,
x='X',
y='Y',
z='Z',
color='label',
)
trace.show()
fig = plt.figure(figsize=(40, 40))
x1, y1 = purturbate(signal, range)
#print(x1)
#print(y1)
close = closest_node(coordinates, points)
x2, y2 = purturbate(X[close], range)
new_points = list(points.copy())
new_points.pop(close)
close_2 = closest_node(coordinates, new_points)
x3, y3 = purturbate(X[close_2], range)
##
print("Out of Distribution Signal Prediction: ",
int(torch.round(torch.sigmoid(model(signal.unsqueeze(0)))).item()))
ax = fig.add_subplot(411)
# print(signal)
signal = torch.abs(torch.sum(signal, dim=0))
# print((signal).shape)
# print(signal)
ax.plot(signal)
ax.set_xlabel('Time')
ax.set_ylabel('Motion')
rect = Rectangle((x1, 0),
y1 - x1,
20,
color='red',
fc=(1, 0, 0, 0.2),
ec=(0, 0, 0, 1))
plt.ylim((0, 20))
ax.add_patch(rect)
plt.show()
##
print("Closest Signal Prediction: ",
int(torch.round(torch.sigmoid(model(X[close].unsqueeze(0)))).item()))
fig = plt.figure(figsize=(40, 40))
ax = fig.add_subplot(411)
close_signal = torch.abs(torch.sum(X[close_2], dim=0))
ax.plot(close_signal)
ax.set_xlabel('Time')
ax.set_ylabel('Motion')
rect = Rectangle((x3, 0),
y3 - x3,
20,
color='red',
fc=(1, 0, 0, 0.2),
ec=(0, 0, 0, 1))
plt.ylim((0, 20))
ax.add_patch(rect)
plt.show()