def main():
path = os.path.join(
ROOT_DIR, 'spikes', 'mnist', 'crop_locally_connected',
'train_2_12_4_100_4_0.01_0.99_60000_250.0_250_1.0_0.05_1e-07_0.5_0.2_10_250'
)
ngram_scores = {}
for i in tqdm(range(200, 240)):
f = os.path.join(path, f'{i}.pt')
spikes, labels = torch.load(f, map_location=location)
ngram_scores = update_ngram_scores(spikes=spikes,
labels=labels,
n_labels=10,
n=1,
ngram_scores=ngram_scores)
all_labels = torch.LongTensor()
all_predictions = torch.LongTensor()
for i in tqdm(range(200, 240)):
f = os.path.join(path, f'{i}.pt')
spikes, labels = torch.load(f, map_location=location)
predictions = ngram(spikes=spikes,
ngram_scores=ngram_scores,
n_labels=10,
n=1)
all_labels = torch.cat([all_labels, labels.long()])
all_predictions = torch.cat([all_predictions, predictions.long()])
accuracy = (all_labels == all_predictions).float().mean() * 100
print(f'Training accuracy: {accuracy:.2f}')
path = os.path.join(
ROOT_DIR, 'spikes', 'mnist', 'crop_locally_connected',
'test_2_12_4_100_4_0.01_0.99_60000_10000_250.0_250_1.0_0.05_1e-07_0.5_0.2_10_250'
)
all_labels = torch.LongTensor()
all_predictions = torch.LongTensor()
for i in tqdm(range(1, 40)):
f = os.path.join(path, f'{i}.pt')
spikes, labels = torch.load(f, map_location=location)
predictions = ngram(spikes=spikes,
ngram_scores=ngram_scores,
n_labels=10,
n=1)
all_labels = torch.cat([all_labels, labels.long()])
all_predictions = torch.cat([all_predictions, predictions.long()])
accuracy = (all_labels == all_predictions).float().mean() * 100
print(f'Test accuracy: {accuracy:.2f}')
def update_curves(curves: Dict[str,
list], labels: torch.Tensor, n_classes: int,
**kwargs) -> Tuple[Dict[str, list], Dict[str, torch.Tensor]]:
# language=rst
"""
Updates accuracy curves for each classification scheme.
:param curves: Mapping from name of classification scheme to list of accuracy evaluations.
:param labels: One-dimensional ``torch.Tensor`` of integer data labels.
:param n_classes: Number of data categories.
:param kwargs: Additional keyword arguments for classification scheme evaluation functions.
:return: Updated accuracy curves and predictions.
"""
predictions = {}
for scheme in curves:
# Branch based on name of classification scheme
if scheme == 'all':
spike_record = kwargs['spike_record']
assignments = kwargs['assignments']
prediction = all_activity(spike_record, assignments, n_classes)
elif scheme == 'proportion':
spike_record = kwargs['spike_record']
assignments = kwargs['assignments']
proportions = kwargs['proportions']
prediction = proportion_weighting(spike_record, assignments,
proportions, n_classes)
elif scheme == 'ngram':
spike_record = kwargs['spike_record']
ngram_scores = kwargs['ngram_scores']
n = kwargs['n']
prediction = ngram(spike_record, ngram_scores, n_classes, n)
elif scheme == 'logreg':
full_spike_record = kwargs['full_spike_record']
logreg = kwargs['logreg']
prediction = logreg_predict(spikes=full_spike_record,
logreg=logreg)
else:
raise NotImplementedError
# Compute accuracy with current classification scheme.
predictions[scheme] = prediction
accuracy = torch.sum(labels.long() == prediction).float() / len(labels)
curves[scheme].append(100 * accuracy)
return curves, predictions
def batch_predictions(self, images):
spike_record = torch.zeros(len(images), 250, self._model.layers['Y'].n)
for i, image in enumerate(images):
self._model.run(inpts={
'X':
poisson(datum=torch.tensor(image), time=250, dt=1)
},
time=250,
dt=1)
self._model.reset_()
spike_record[i] = self._model.monitors['Y_spikes'].get('s').t()
labels = ngram(spike_record, self._ngram_scores, self.num_classes(),
2).numpy()
return labels
def predictions(self, image):
global axes, ims
self._model.run(
inpts={'X': poisson(datum=torch.tensor(image), time=250, dt=1)},
time=250,
dt=1)
spike_record = self._model.monitors['Y_spikes'].get('s').t().unsqueeze(
0)
label = ngram(spike_record, self._ngram_scores, self.num_classes(),
2).numpy()[0]
self._model.reset_()
# axes, ims = plot_input(image.reshape(20, 20), image.reshape(20, 20), axes=axes, ims=ims)
# plt.pause(0.05)
print(label)
return label
def main(seed=0, n_examples=100, gpu=False, plot=False):
np.random.seed(seed)
if gpu:
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.manual_seed_all(seed)
else:
torch.manual_seed(seed)
model_name = '0_12_4_150_4_0.01_0.99_60000_250.0_250_1.0_0.05_1e-07_0.5_0.2_10_250'
network = load_network(os.path.join(params_path, f'{model_name}.pt'))
for l in network.layers:
network.layers[l].dt = network.dt
for c in network.connections:
network.connections[c].dt = network.dt
network.layers['Y'].one_spike = True
network.layers['Y'].lbound = None
kernel_size = 12
side_length = 20
n_filters = 150
time = 250
intensity = 0.5
crop = 4
conv_size = network.connections['X', 'Y'].conv_size
locations = network.connections['X', 'Y'].locations
conv_prod = int(np.prod(conv_size))
n_neurons = n_filters * conv_prod
n_classes = 10
# Voltage recording for excitatory and inhibitory layers.
voltage_monitor = Monitor(network.layers['Y'], ['v'], time=time)
network.add_monitor(voltage_monitor, name='output_voltage')
# Load MNIST data.
dataset = MNIST(path=data_path, download=True)
images, labels = dataset.get_test()
images *= intensity
images = images[:, crop:-crop, crop:-crop]
# Neuron assignments and spike proportions.
path = os.path.join(params_path,
'_'.join(['auxiliary', model_name]) + '.pt')
assignments, proportions, rates, ngram_scores = torch.load(open(
path, 'rb'))
spikes = {}
for layer in set(network.layers):
spikes[layer] = Monitor(network.layers[layer],
state_vars=['s'],
time=time)
network.add_monitor(spikes[layer], name=f'{layer}_spikes')
# Train the network.
print('\nBegin black box adversarial attack.\n')
spike_ims = None
spike_axes = None
weights_im = None
inpt_ims = None
inpt_axes = None
max_iters = 25
delta = 0.1
epsilon = 0.1
for i in range(n_examples):
# Get next input sample.
original = images[i % len(images)].contiguous().view(-1)
label = labels[i % len(images)]
# Check if the image is correctly classified.
sample = poisson(datum=original, time=time)
inpts = {'X': sample}
# Run the network on the input.
network.run(inpts=inpts, time=time)
# Check for incorrect classification.
s = spikes['Y'].get('s').view(1, n_neurons, time)
prediction = ngram(spikes=s,
ngram_scores=ngram_scores,
n_labels=10,
n=2).item()
if prediction != label:
continue
# Create adversarial example.
adversarial = False
while not adversarial:
adv_example = 255 * torch.rand(original.size())
sample = poisson(datum=adv_example, time=time)
inpts = {'X': sample}
# Run the network on the input.
network.run(inpts=inpts, time=time)
# Check for incorrect classification.
s = spikes['Y'].get('s').view(1, n_neurons, time)
prediction = ngram(spikes=s,
ngram_scores=ngram_scores,
n_labels=n_classes,
n=2).item()
if prediction == label:
adversarial = True
j = 0
current = original.clone()
while j < max_iters:
# Orthogonal perturbation.
# perturb = orthogonal_perturbation(delta=delta, image=adv_example, target=original)
# temp = adv_example + perturb
# # Forward perturbation.
# temp = temp.clone() + forward_perturbation(epsilon * get_diff(temp, original), temp, adv_example)
# print(temp)
perturbation = torch.randn(original.size())
unnormed_source_direction = original - perturbation
source_norm = torch.norm(unnormed_source_direction)
source_direction = unnormed_source_direction / source_norm
dot = torch.dot(perturbation, source_direction)
perturbation -= dot * source_direction
perturbation *= epsilon * source_norm / torch.norm(perturbation)
D = 1 / np.sqrt(epsilon**2 + 1)
direction = perturbation - unnormed_source_direction
spherical_candidate = current + D * direction
spherical_candidate = torch.clamp(spherical_candidate, 0, 255)
new_source_direction = original - spherical_candidate
new_source_direction_norm = torch.norm(new_source_direction)
# length if spherical_candidate would be exactly on the sphere
length = delta * source_norm
# length including correction for deviation from sphere
deviation = new_source_direction_norm - source_norm
length += deviation
# make sure the step size is positive
length = max(0, length)
# normalize the length
length = length / new_source_direction_norm
candidate = spherical_candidate + length * new_source_direction
candidate = torch.clamp(candidate, 0, 255)
sample = poisson(datum=candidate, time=time)
inpts = {'X': sample}
# Run the network on the input.
network.run(inpts=inpts, time=time)
# Check for incorrect classification.
s = spikes['Y'].get('s').view(1, n_neurons, time)
prediction = ngram(spikes=s,
ngram_scores=ngram_scores,
n_labels=10,
n=2).item()
# Optionally plot various simulation information.
if plot:
_input = original.view(side_length, side_length)
reconstruction = candidate.view(side_length, side_length)
_spikes = {
'X': spikes['X'].get('s').view(side_length**2, time),
'Y': spikes['Y'].get('s').view(n_neurons, time)
}
w = network.connections['X', 'Y'].w
spike_ims, spike_axes = plot_spikes(spikes=_spikes,
ims=spike_ims,
axes=spike_axes)
weights_im = plot_locally_connected_weights(w,
n_filters,
kernel_size,
conv_size,
locations,
side_length,
im=weights_im)
inpt_axes, inpt_ims = plot_input(_input,
reconstruction,
label=labels[i],
ims=inpt_ims,
axes=inpt_axes)
plt.pause(1e-8)
if prediction == label:
print('Attack failed.')
else:
print('Attack succeeded.')
adv_example = candidate
j += 1
network.reset_() # Reset state variables.
print('\nAdversarial attack complete.\n')
def main():
#TEST
# hyperparameters
n_neurons = 100
n_test = 10000
inhib = 100
time = 350
dt = 1
intensity = 0.25
# extra args
progress_interval = 10
update_interval = 250
plot = True
seed = 0
train = True
gpu = False
n_classes = 10
n_sqrt = int(np.ceil(np.sqrt(n_neurons)))
# TESTING
assert n_test % update_interval == 0
np.random.seed(seed)
save_weights_fn = "plots_snn/weights/weights_test.png"
save_performance_fn = "plots_snn/performance/performance_test.png"
save_assaiments_fn = "plots_snn/assaiments/assaiments_test.png"
# load network
network = load('net_output.pt') # here goes file with network to load
network.train(False)
# pull dataset
data, targets = torch.load(
'data/MNIST/TorchvisionDatasetWrapper/processed/test.pt')
data = data * intensity
data_stretched = data.view(len(data), -1, 784)
testset = torch.utils.data.TensorDataset(data_stretched, targets)
testloader = torch.utils.data.DataLoader(testset,
batch_size=1,
shuffle=True)
# spike init
spike_record = torch.zeros(update_interval, time, n_neurons)
full_spike_record = torch.zeros(n_test, n_neurons).long()
# load parameters
assignments, proportions, rates, ngram_scores = torch.load(
'parameters_output.pt') # here goes file with parameters to load
# accuracy initialization
curves = {'all': [], 'proportion': [], 'ngram': []}
predictions = {scheme: torch.Tensor().long() for scheme in curves.keys()}
spikes = {}
for layer in set(network.layers):
spikes[layer] = Monitor(network.layers[layer],
state_vars=['s'],
time=time)
network.add_monitor(spikes[layer], name='%s_spikes' % layer)
print("Begin test.")
inpt_axes = None
inpt_ims = None
spike_ims = None
spike_axes = None
weights_im = None
assigns_im = None
perf_ax = None
i = 0
current_labels = torch.zeros(update_interval)
# test
test_time = t.time()
time1 = t.time()
for sample, label in testloader:
sample = sample.view(1, 1, 28, 28)
if i % progress_interval == 0:
print(f'Progress: {i} / {n_test} took {(t.time()-time1)*10000} s')
if i % update_interval == 0 and i > 0:
# update accuracy evaluation
curves, preds = update_curves(curves,
current_labels,
n_classes,
spike_record=spike_record,
assignments=assignments,
proportions=proportions,
ngram_scores=ngram_scores,
n=2)
print_results(curves)
for scheme in preds:
predictions[scheme] = torch.cat(
[predictions[scheme], preds[scheme]], -1)
sample_enc = poisson(datum=sample, time=time, dt=dt)
inpts = {'X': sample_enc}
# Run the network on the input.
network.run(inputs=inpts, time=time)
retries = 0
while spikes['Ae'].get('s').sum() < 1 and retries < 3:
retries += 1
sample = sample * 2
inpts = {'X': poisson(datum=sample, time=time, dt=dt)}
network.run(inputs=inpts, time=time)
# Spikes reocrding
spike_record[i % update_interval] = spikes['Ae'].get('s').view(
time, n_neurons)
full_spike_record[i] = spikes['Ae'].get('s').view(
time, n_neurons).sum(0).long()
if plot:
_input = sample.view(28, 28)
reconstruction = inpts['X'].view(time, 784).sum(0).view(28, 28)
_spikes = {layer: spikes[layer].get('s') for layer in spikes}
input_exc_weights = network.connections[('X', 'Ae')].w
square_assignments = get_square_assignments(assignments, n_sqrt)
assigns_im = plot_assignments(square_assignments, im=assigns_im)
if i % update_interval == 0: # plot weights on every update interval
square_weights = get_square_weights(
input_exc_weights.view(784, n_neurons), n_sqrt, 28)
weights_im = plot_weights(square_weights, im=weights_im)
[weights_im,
save_weights_fn] = plot_weights(square_weights,
im=weights_im,
save=save_weights_fn)
inpt_axes, inpt_ims = plot_input(_input,
reconstruction,
label=label,
axes=inpt_axes,
ims=inpt_ims)
spike_ims, spike_axes = plot_spikes(_spikes,
ims=spike_ims,
axes=spike_axes)
assigns_im = plot_assignments(square_assignments,
im=assigns_im,
save=save_assaiments_fn)
perf_ax = plot_performance(curves,
ax=perf_ax,
save=save_performance_fn)
plt.pause(1e-8)
current_labels[i % update_interval] = label[0]
network.reset_state_variables()
if i % 10 == 0 and i > 0:
preds = ngram(
spike_record[i % update_interval - 10:i % update_interval],
ngram_scores, n_classes, 2)
print(f'Predictions: {(preds*1.0).numpy()}')
print(
f'True value: {current_labels[i%update_interval-10:i%update_interval].numpy()}'
)
time1 = t.time()
i += 1
# Compute confusion matrices and save them to disk.
confusions = {}
for scheme in predictions:
confusions[scheme] = confusion_matrix(targets, predictions[scheme])
to_write = 'confusion_test'
f = '_'.join([str(x) for x in to_write]) + '.pt'
torch.save(confusions, os.path.join('.', f))
print("Test completed. Testing took " + str((t.time() - test_time) / 6) +
" min.")
def main(seed=0, n_train=60000, n_test=10000, inhib=250, kernel_size=(16,), stride=(2,), n_filters=25, crop=4, lr=0.01,
lr_decay=1, time=100, dt=1, theta_plus=0.05, theta_decay=1e-7, intensity=1, norm=0.2, progress_interval=10,
update_interval=250, plot=False, train=True, gpu=False):
assert n_train % update_interval == 0 and n_test % update_interval == 0, \
'No. examples must be divisible by update_interval'
params = [
seed, kernel_size, stride, n_filters, crop, lr, lr_decay, n_train, inhib, time, dt,
theta_plus, theta_decay, intensity, norm, progress_interval, update_interval
]
model_name = '_'.join([str(x) for x in params])
if not train:
test_params = [
seed, kernel_size, stride, n_filters, crop, lr, lr_decay, n_train, n_test, inhib, time, dt,
theta_plus, theta_decay, intensity, norm, progress_interval, update_interval
]
np.random.seed(seed)
if gpu:
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.manual_seed_all(seed)
else:
torch.manual_seed(seed)
side_length = 28 - crop * 2
n_examples = n_train if train else n_test
network = load_network(os.path.join(params_path, model_name + '.pt'))
network.layers['X'] = Input(n=400)
network.layers['Y'] = DiehlAndCookNodes(
n=network.layers['Y'].n, thresh=network.layers['Y'].thresh, rest=network.layers['Y'].rest,
reset=network.layers['Y'].reset, theta_plus=network.layers['Y'].theta_plus,
theta_decay=network.layers['Y'].theta_decay
)
network.add_layer(network.layers['X'], name='X')
network.add_layer(network.layers['Y'], name='Y')
network.connections['X', 'Y'].source = network.layers['X']
network.connections['X', 'Y'].target = network.layers['Y']
network.connections['X', 'Y'].update_rule = NoOp(
connection=network.connections['X', 'Y'], nu=network.connections['X', 'Y'].nu
)
network.layers['Y'].theta_decay = 0
network.layers['Y'].theta_plus = 0
conv_size = network.connections['X', 'Y'].conv_size
locations = network.connections['X', 'Y'].locations
conv_prod = int(np.prod(conv_size))
n_neurons = n_filters * conv_prod
n_classes = 10
# Voltage recording for excitatory and inhibitory layers.
voltage_monitor = Monitor(network.layers['Y'], ['v'], time=time)
network.add_monitor(voltage_monitor, name='output_voltage')
# Load MNIST data.
dataset = MNIST(path=data_path, download=True)
images, labels = dataset.get_test()
images *= intensity
images = images[:, crop:-crop, crop:-crop]
# Neuron assignments and spike proportions.
path = os.path.join(params_path, '_'.join(['auxiliary', model_name]) + '.pt')
assignments, proportions, rates, ngram_scores = torch.load(open(path, 'rb'))
spikes = {}
for layer in set(network.layers):
spikes[layer] = Monitor(network.layers[layer], state_vars=['s'], time=time)
network.add_monitor(spikes[layer], name=f'{layer}_spikes')
# Train the network.
print('\nBegin black box adversarial attack.\n')
spike_ims = None
spike_axes = None
weights_im = None
inpt_ims = None
inpt_axes = None
max_iters = 25
delta = 0.1
epsilon = 0.1
for i in range(n_examples):
# Get next input sample.
original = images[i % len(images)].contiguous().view(-1)
label = labels[i % len(images)]
# Check if the image is correctly classified.
sample = poisson(datum=original, time=time)
inpts = {'X': sample}
# Run the network on the input.
network.run(inpts=inpts, time=time)
# Check for incorrect classification.
s = spikes['Y'].get('s').view(1, n_neurons, time)
prediction = ngram(spikes=s, ngram_scores=ngram_scores, n_labels=10, n=2).item()
if prediction != label:
continue
# Create adversarial example.
adversarial = False
while not adversarial:
adv_example = 255 * torch.rand(original.size())
sample = poisson(datum=adv_example, time=time)
inpts = {'X': sample}
# Run the network on the input.
network.run(inpts=inpts, time=time)
# Check for incorrect classification.
s = spikes['Y'].get('s').view(1, n_neurons, time)
prediction = ngram(spikes=s, ngram_scores=ngram_scores, n_labels=n_classes, n=2).item()
if prediction == label:
adversarial = True
j = 0
current = original.clone()
while j < max_iters:
# Orthogonal perturbation.
# perturb = orthogonal_perturbation(delta=delta, image=adv_example, target=original)
# temp = adv_example + perturb
# # Forward perturbation.
# temp = temp.clone() + forward_perturbation(epsilon * get_diff(temp, original), temp, adv_example)
# print(temp)
perturbation = torch.randn(original.size())
unnormed_source_direction = original - perturbation
source_norm = torch.norm(unnormed_source_direction)
source_direction = unnormed_source_direction / source_norm
dot = torch.dot(perturbation, source_direction)
perturbation -= dot * source_direction
perturbation *= epsilon * source_norm / torch.norm(perturbation)
D = 1 / np.sqrt(epsilon ** 2 + 1)
direction = perturbation - unnormed_source_direction
spherical_candidate = current + D * direction
spherical_candidate = torch.clamp(spherical_candidate, 0, 255)
new_source_direction = original - spherical_candidate
new_source_direction_norm = torch.norm(new_source_direction)
# length if spherical_candidate would be exactly on the sphere
length = delta * source_norm
# length including correction for deviation from sphere
deviation = new_source_direction_norm - source_norm
length += deviation
# make sure the step size is positive
length = max(0, length)
# normalize the length
length = length / new_source_direction_norm
candidate = spherical_candidate + length * new_source_direction
candidate = torch.clamp(candidate, 0, 255)
sample = poisson(datum=candidate, time=time)
inpts = {'X': sample}
# Run the network on the input.
network.run(inpts=inpts, time=time)
# Check for incorrect classification.
s = spikes['Y'].get('s').view(1, n_neurons, time)
prediction = ngram(spikes=s, ngram_scores=ngram_scores, n_labels=10, n=2).item()
# Optionally plot various simulation information.
if plot:
_input = original.view(side_length, side_length)
reconstruction = candidate.view(side_length, side_length)
_spikes = {
'X': spikes['X'].get('s').view(side_length ** 2, time),
'Y': spikes['Y'].get('s').view(n_neurons, time)
}
w = network.connections['X', 'Y'].w
spike_ims, spike_axes = plot_spikes(spikes=_spikes, ims=spike_ims, axes=spike_axes)
weights_im = plot_locally_connected_weights(
w, n_filters, kernel_size, conv_size, locations, side_length, im=weights_im
)
inpt_axes, inpt_ims = plot_input(
_input, reconstruction, label=labels[i], ims=inpt_ims, axes=inpt_axes
)
plt.pause(1e-8)
if prediction == label:
print('Attack failed.')
else:
print('Attack succeeded.')
adv_example = candidate
j += 1
network.reset_() # Reset state variables.
print('\nAdversarial attack complete.\n')
print('\nBegin training.\n')
else:
print('\nBegin test.\n')
start = t()
for i in range(n_examples):
if i % progress_interval == 0:
print(f'Progress: {i} / {n_examples} ({t() - start:.4f} seconds)')
start = t()
if i % update_interval == 0 and i > 0:
# Get network predictions.
all_activity_pred = all_activity(spike_record, assignments, 10)
proportion_pred = proportion_weighting(spike_record, assignments,
proportions, 10)
ngram_pred = ngram(spike_record, ngram_scores, 10, 2)
# Compute network accuracy according to available classification strategies.
curves['all'].append(100 * torch.sum(labels[i - update_interval:i].long() \
== all_activity_pred) / update_interval)
curves['proportion'].append(100 * torch.sum(labels[i - update_interval:i].long() \
== proportion_pred) / update_interval)
curves['ngram'].append(100 * torch.sum(labels[i - update_interval:i].long() \
== ngram_pred) / update_interval)
print_results(curves)
if train:
if any([x[-1] > best_accuracy for x in curves.values()]):
print('New best accuracy! Saving network parameters to disk.')
