def _graph_fn_unsquash(self, values):
if get_backend() == "tf":
return tf.atanh((values - self.low) /
(self.high - self.low) * 2.0 - 1.0)
elif get_backend() == "tf":
return torch.atanh((values - self.low) /
(self.high - self.low) * 2.0 - 1.0)
def get_log_prob(self, state: rlt.FeatureData,
squashed_action: torch.Tensor):
"""
Action is expected to be squashed with tanh
"""
if self.use_l2_normalization:
# TODO: calculate log_prob for l2 normalization
# https://math.stackexchange.com/questions/3120506/on-the-distribution-of-a-normalized-gaussian-vector
# http://proceedings.mlr.press/v100/mazoure20a/mazoure20a.pdf
pass
loc, scale_log = self._get_loc_and_scale_log(state)
raw_action = torch.atanh(squashed_action)
r = (raw_action - loc) / scale_log.exp()
log_prob = self._normal_log_prob(r, scale_log)
squash_correction = self._squash_correction(squashed_action)
if SummaryWriterContext._global_step % 1000 == 0:
SummaryWriterContext.add_histogram("actor/get_log_prob/loc",
loc.detach().cpu())
SummaryWriterContext.add_histogram("actor/get_log_prob/scale_log",
scale_log.detach().cpu())
SummaryWriterContext.add_histogram("actor/get_log_prob/log_prob",
log_prob.detach().cpu())
SummaryWriterContext.add_histogram(
"actor/get_log_prob/squash_correction",
squash_correction.detach().cpu())
return torch.sum(log_prob - squash_correction, dim=1).reshape(-1, 1)
def sumlogC(x, eps=1e-5):
"""
Numerically stable implementation of sum of logarithm of Continuous Bernoulli constant C
Returns log normalising constant for x in (0, x-eps) and (x+eps, 1)
Uses Taylor 3rd degree approximation in [x-eps, x+eps].
Parameter
----------
x : `torch.Tensor`
[batch_size x C x H x W]. x takes values in (0,1).
"""
# clip x such that x in (0, 1)
x = torch.clamp(x, eps, 1. - eps)
# get mask if x is not in [0.5-eps, 0.5+eps]
mask = torch.abs(x - .5).ge(eps)
# points that are (0, 0.5-eps) and (0.5+eps, 1)
far = x[mask]
# points that are [0.5-eps, 0.5+eps]
close = x[~mask]
# Given by log(|2tanh^-1(1-2x)|) - log(|1-2x|)
far_values = torch.log(torch.abs(
2. * torch.atanh(1 - 2. * far))) - torch.log(torch.abs(1 - 2. * far))
# using Taylor expansion to 3rd degree
close_values = torch.log(2. + (1 - 2 * close).pow(2) / 3 +
(1 - 2 * close).pow(4) / 5)
return far_values.sum() + close_values.sum()
def log_prob(self, value: torch.Tensor) -> torch.Tensor:
value = torch.clamp(value, -1 + 1e-6, 1 - 1e-6)
# Log likelihood for Gaussian with change of variable
log_prob = super().log_prob(torch.atanh(value))
# Squash correction (from original SAC implementation)
# this comes from the fact that tanh is bijective and differentiable
log_prob -= torch.sum(torch.log(1 - value**2), dim=1)
return log_prob
def logp(self, x, ac):
if self.use_squashing:
ac = torch.atanh(ac)
m = self.__act_distribution(x)
logp_pi = m.log_prob(ac).sum(axis=1)
logp_pi -= (2 * (np.log(2) - ac - F.softplus(-2 * ac))).sum(axis=1)
return logp_pi
else:
m = self.__act_distribution(x)
return m.log_prob(ac).sum(axis=1)
def evaluation(self, state, action):
mean, log_std = self.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
y_t = (action - self.action_bias) / self.action_scale
x_t = torch.atanh(y_t)
log_prob = normal.log_prob(x_t)
log_prob -= torch.log(self.action_scale * (1 - y_t.pow(2)) + epsilon)
log_prob = log_prob.sum(1, keepdim=True)
return log_prob
def cosine_loss(a, v, y):
# d = nn.functional.cosine_similarity(a, v)
# print(y)
a = a / a.norm(p=2, dim=1, keepdim=True)
v = v / v.norm(p=2, dim=1, keepdim=True)
cosim = a.matmul(v.transpose(0, 1))
cosim = cosim[torch.eye(len(cosim)).bool()].unsqueeze(1)
cosim = torch.atanh(cosim)
loss = logloss(F.sigmoid(cosim), y)
return loss
def _compute_actor_loss(self, observation, action, weight):
dist = self.policy.dist(observation)
# unnormalize action via inverse tanh function
unnormalized_action = torch.atanh(action.clamp(-0.999999, 0.999999))
# compute log probability
_, log_probs = squash_action(dist, unnormalized_action)
return -(weight * log_probs).mean()
def forward(self, z, inverse_mode=False, context=None):
if inverse_mode:
squashed = 2 * (z - self.low) / self.range - 1
return torch.atanh(
torch.clip(z, -1 + SMALL_NUMBER,
1 - SMALL_NUMBER)), -self._log_abs_det_jac(squashed)
else:
squashed = torch.tanh(z)
squashed01 = (squashed + 1) / 2
return squashed01 * self.range + self.low, self._log_abs_det_jac(
squashed)
def get_logprob(self, state, action):
mean, log_std = self.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
x_t = torch.atanh(action)
y_t = torch.tanh(x_t)
log_prob = normal.log_prob(x_t)
log_prob -= torch.log(1 - y_t.pow(2) + 1e-6)
return log_prob
def reduce_tanh(self, nodes):
w = torch.exp(self.w)
msg = w * nodes.mailbox['m'] + self.b
""" print("MSG max: ", torch.max(msg))
print("MSG min: ", torch.min(msg)) """
fsum = torch.clamp(torch.sum(torch.tanh(msg), dim=1), -0.99, 0.99)
""" print("max: ", torch.max(fsum))
print("min: ", torch.min(fsum)) """
out_h = (torch.atanh(fsum) - self.b) / w
return {'neigh': out_h}
def _compute_actor_loss(self, obs_t, act_t):
dist = self.policy.dist(obs_t)
# unnormalize action via inverse tanh function
unnormalized_act_t = torch.atanh(act_t.clamp(-0.999999, 0.999999))
# compute log probability
_, log_probs = squash_action(dist, unnormalized_act_t)
# compute exponential weight
weights = self._compute_weights(obs_t, act_t)
return -(log_probs * weights).sum()
def reduce_tanh(self, nodes):
alpha = F.softmax(nodes.mailbox['e'], dim=1)
alpha = F.dropout(alpha, self.dropout, training=self.training)
msg = (alpha * nodes.mailbox['z'])
w = torch.exp(self.w)
msg = w * msg + self.b
fsum = torch.clamp(torch.sum(torch.tanh(msg), dim=1), -0.9999999,
0.9999999)
""" print("max: ", torch.max(fsum))
print("min: ", torch.min(fsum)) """
out_h = (torch.atanh(fsum) - self.b) / w
return {'h': out_h}
def to_tanh_space(x, box):
# type: (Union[Variable, torch.FloatTensor], Tuple[float, float]) -> Union[Variable, torch.FloatTensor]
"""
Convert a batch of tensors to tanh-space. This method complements the
implementation of the change-of-variable trick in terms of tanh.
:param x: the batch of tensors, of dimension [B x n_features]
:param box: a tuple of lower bound and upper bound of the box constraint
:return: the batch of tensors in tanh-space, of the same dimension;
the returned tensor is on the same device as ``x``
"""
_box_mul = (box[1] - box[0]) * 0.5
_box_plus = (box[1] + box[0]) * 0.5
return torch.atanh((x - _box_plus) / _box_mul) * 1e4
def compute_actor_loss(self, batch: TorchMiniBatch) -> torch.Tensor:
assert self._policy is not None
dist = self._policy.dist(batch.observations)
# unnormalize action via inverse tanh function
clipped_actions = batch.actions.clamp(-0.999999, 0.999999)
unnormalized_act_t = torch.atanh(clipped_actions)
# compute log probability
_, log_probs = squash_action(dist, unnormalized_act_t)
weight = self._compute_weight(batch.observations, batch.actions)
return -(log_probs * weight).mean()
def _graph_fn_unsquash(self, values):
"""
Reverse operation as _graph_fn_squash (using argus tanh).
Args:
values (DataOp): The values to unsquash.
Returns:
The unsquashed values.
"""
if get_backend() == "tf":
return tf.atanh((values - self.low) /
(self.high - self.low) * 2.0 - 1.0)
elif get_backend() == "tf":
return torch.atanh((values - self.low) /
(self.high - self.low) * 2.0 - 1.0)
def get_repara_action(self, obs, action):
# Generate the latent feature
phi = self.feature_net(obs)
# Reparameterize action with epsilon
action_mean, action_std, _ = self.actor_net.distribution(phi)
if isinstance(self.actor_net, MLPSquashedGaussianActor):
action = torch.clamp(action / self.actor_net.action_lim, -0.999,
0.999)
u = torch.atanh(action)
eps = (u - action_mean) / action_std
repara_action = self.actor_net.action_lim * torch.tanh(
action_mean + action_std * eps.detach())
else:
eps = (action - action_mean) / action_std
repara_action = action_mean + action_std * eps.detach()
return repara_action
def forward(self, phi, action=None, deterministic=False):
# Compute action distribution and the log_pi of given actions
action_mean, action_std, action_distribution = self.distribution(phi)
if action is None:
if deterministic:
u = action_mean
else:
u = action_distribution.rsample(
) if self.rsample else action_distribution.sample()
action = self.action_lim * torch.tanh(u)
else:
u = torch.clamp(action / self.action_lim, -0.999, 0.999)
u = torch.atanh(u)
# Compute logprob from Gaussian, and then apply correction for Tanh squashing.
log_pi = self.log_pi_from_distribution(action_distribution, u)
return action, action_mean, action_std, log_pi
def _compute_actor_loss( # type: ignore
self, obs_t: torch.Tensor, act_t: torch.Tensor
) -> torch.Tensor:
assert self._policy is not None
dist = self._policy.dist(obs_t)
# unnormalize action via inverse tanh function
unnormalized_act_t = torch.atanh(act_t.clamp(-0.999999, 0.999999))
# compute log probability
_, log_probs = squash_action(dist, unnormalized_act_t)
weight = self._compute_weight(obs_t, act_t)
return -(log_probs * weight).mean()
def test_mod_mapping(self):
x, y, z = sympy.symbols("x y z")
expression = x ** 2 + sympy.atanh(sympy.Mod(y + 1, 2) - 1) * 3.2 * z
module = sympy2torch(expression, [x, y, z])
X = torch.rand(100, 3).float() * 10
true_out = (
X[:, 0] ** 2 + torch.atanh(torch.fmod(X[:, 1] + 1, 2) - 1) * 3.2 * X[:, 2]
)
torch_out = module(X)
np.testing.assert_array_almost_equal(
true_out.detach(), torch_out.detach(), decimal=4
)
def rnd2rgb(self, rnd, clip=False):
rnd = torch.atanh(rnd)
rnd_arr = torch.zeros(rnd.shape, device=rnd.get_device())
if not self.linear:
vals = self.distortion + 1e-4
eta = vals[4]
L = rnd_arr[:, 0:1, ]
a = rnd_arr[:, 1:2, ]
b = rnd_arr[:, 2:3, ]
y = (L + vals[1]) / vals[0]
x = (a / vals[2]) + y
z = y - (b / vals[3])
rnd_arr[:, 0:1, ] = x
rnd_arr[:, 1:2, ] = y
rnd_arr[:, 2:3, ] = z
mask = rnd_arr > eta
rnd_arr[mask] = rnd_arr[mask].pow(3.)
rnd_arr[~mask] = (rnd_arr[~mask] - (vals[1] / vals[0])) * 3 * (
eta ** 2)
else:
for i in range(3):
rnd_arr[:, i:i + 1, ] = rnd[:, i:i + 1, ]
# rescale to the reference white (illuminant)
ref_white = self.ref_white + 1e-4
white_arr = torch.zeros(rnd.shape, device=rnd.get_device())
for i in range(3):
white_arr[:, i:i + 1, ] = rnd_arr[:, i:i + 1, ] * ref_white[i]
rgb = torch.zeros(rnd.shape, device=rnd.get_device())
rgb_transform = torch.inverse(self.trans_mat)
for i in range(3):
x_r = white_arr[:, 0:1, ] * rgb_transform[i, 0]
y_g = white_arr[:, 1:2, ] * rgb_transform[i, 1]
z_b = white_arr[:, 2:3, ] * rgb_transform[i, 2]
rgb[:, i:i + 1, ] = x_r + y_g + z_b
if clip:
rgb[torch.isnan(rgb)] = 0
rgb = (rgb * 0.5) + 0.5
rgb[rgb < 0] = 0
rgb[rgb > 1] = 1
return rgb
def _compute_actor_loss(
self,
observation: torch.Tensor,
action: torch.Tensor,
weight: torch.Tensor,
) -> torch.Tensor:
assert self._policy is not None
dist = self._policy.dist(observation)
# unnormalize action via inverse tanh function
unnormalized_action = torch.atanh(action.clamp(-0.999999, 0.999999))
# compute log probability
_, log_probs = squash_action(dist, unnormalized_action)
return -(weight * log_probs).mean()
def rnd2rgb(self, y):
for i in range(self.layers - 1, -1, -1):
trans_mat = getattr(self, 't%.3d' % i)[0].weight.detach().squeeze()
trans_mat = torch.inverse(trans_mat)
y = torch.atanh(y)
if self.bias:
bias_vec = getattr(self, 't%.3d' % i)[0].bias.detach().squeeze()
for i in range(3):
y[:, i, ] -= bias_vec[i]
device = y.get_device()
device = 'cpu' if device == -1 else device
x = torch.zeros(y.shape, device=device)
for i in range(3):
x_r = y[:, 0:1, ] * trans_mat[i, 0]
y_g = y[:, 1:2, ] * trans_mat[i, 1]
z_b = y[:, 2:3, ] * trans_mat[i, 2]
x[:, i:i + 1, ] = x_r + y_g + z_b
y = x.clone()
return x
def forward(self, sample):
return torch.atanh(sample / self.a) / self.b
def atanh(x):
return _torch.atanh(x)
def pointwise_ops(self):
a = torch.randn(4)
b = torch.randn(4)
t = torch.tensor([-1, -2, 3], dtype=torch.int8)
r = torch.tensor([0, 1, 10, 0], dtype=torch.int8)
t = torch.tensor([-1, -2, 3], dtype=torch.int8)
s = torch.tensor([4, 0, 1, 0], dtype=torch.int8)
f = torch.zeros(3)
g = torch.tensor([-1, 0, 1])
w = torch.tensor([0.3810, 1.2774, -0.2972, -0.3719, 0.4637])
return (
torch.abs(torch.tensor([-1, -2, 3])),
torch.absolute(torch.tensor([-1, -2, 3])),
torch.acos(a),
torch.arccos(a),
torch.acosh(a.uniform_(1.0, 2.0)),
torch.add(a, 20),
torch.add(a, torch.randn(4, 1), alpha=10),
torch.addcdiv(torch.randn(1, 3),
torch.randn(3, 1),
torch.randn(1, 3),
value=0.1),
torch.addcmul(torch.randn(1, 3),
torch.randn(3, 1),
torch.randn(1, 3),
value=0.1),
torch.angle(a),
torch.asin(a),
torch.arcsin(a),
torch.asinh(a),
torch.arcsinh(a),
torch.atan(a),
torch.arctan(a),
torch.atanh(a.uniform_(-1.0, 1.0)),
torch.arctanh(a.uniform_(-1.0, 1.0)),
torch.atan2(a, a),
torch.bitwise_not(t),
torch.bitwise_and(t, torch.tensor([1, 0, 3], dtype=torch.int8)),
torch.bitwise_or(t, torch.tensor([1, 0, 3], dtype=torch.int8)),
torch.bitwise_xor(t, torch.tensor([1, 0, 3], dtype=torch.int8)),
torch.ceil(a),
torch.clamp(a, min=-0.5, max=0.5),
torch.clamp(a, min=0.5),
torch.clamp(a, max=0.5),
torch.clip(a, min=-0.5, max=0.5),
torch.conj(a),
torch.copysign(a, 1),
torch.copysign(a, b),
torch.cos(a),
torch.cosh(a),
torch.deg2rad(
torch.tensor([[180.0, -180.0], [360.0, -360.0], [90.0,
-90.0]])),
torch.div(a, b),
torch.divide(a, b, rounding_mode="trunc"),
torch.divide(a, b, rounding_mode="floor"),
torch.digamma(torch.tensor([1.0, 0.5])),
torch.erf(torch.tensor([0.0, -1.0, 10.0])),
torch.erfc(torch.tensor([0.0, -1.0, 10.0])),
torch.erfinv(torch.tensor([0.0, 0.5, -1.0])),
torch.exp(torch.tensor([0.0, math.log(2.0)])),
torch.exp2(torch.tensor([0.0, math.log(2.0), 3.0, 4.0])),
torch.expm1(torch.tensor([0.0, math.log(2.0)])),
torch.fake_quantize_per_channel_affine(
torch.randn(2, 2, 2),
(torch.randn(2) + 1) * 0.05,
torch.zeros(2),
1,
0,
255,
),
torch.fake_quantize_per_tensor_affine(a, 0.1, 0, 0, 255),
torch.float_power(torch.randint(10, (4, )), 2),
torch.float_power(torch.arange(1, 5), torch.tensor([2, -3, 4,
-5])),
torch.floor(a),
# torch.floor_divide(torch.tensor([4.0, 3.0]), torch.tensor([2.0, 2.0])),
# torch.floor_divide(torch.tensor([4.0, 3.0]), 1.4),
torch.fmod(torch.tensor([-3, -2, -1, 1, 2, 3]), 2),
torch.fmod(torch.tensor([1, 2, 3, 4, 5]), 1.5),
torch.frac(torch.tensor([1.0, 2.5, -3.2])),
torch.randn(4, dtype=torch.cfloat).imag,
torch.ldexp(torch.tensor([1.0]), torch.tensor([1])),
torch.ldexp(torch.tensor([1.0]), torch.tensor([1, 2, 3, 4])),
torch.lerp(torch.arange(1.0, 5.0),
torch.empty(4).fill_(10), 0.5),
torch.lerp(
torch.arange(1.0, 5.0),
torch.empty(4).fill_(10),
torch.full_like(torch.arange(1.0, 5.0), 0.5),
),
torch.lgamma(torch.arange(0.5, 2, 0.5)),
torch.log(torch.arange(5) + 10),
torch.log10(torch.rand(5)),
torch.log1p(torch.randn(5)),
torch.log2(torch.rand(5)),
torch.logaddexp(torch.tensor([-1.0]), torch.tensor([-1, -2, -3])),
torch.logaddexp(torch.tensor([-100.0, -200.0, -300.0]),
torch.tensor([-1, -2, -3])),
torch.logaddexp(torch.tensor([1.0, 2000.0, 30000.0]),
torch.tensor([-1, -2, -3])),
torch.logaddexp2(torch.tensor([-1.0]), torch.tensor([-1, -2, -3])),
torch.logaddexp2(torch.tensor([-100.0, -200.0, -300.0]),
torch.tensor([-1, -2, -3])),
torch.logaddexp2(torch.tensor([1.0, 2000.0, 30000.0]),
torch.tensor([-1, -2, -3])),
torch.logical_and(r, s),
torch.logical_and(r.double(), s.double()),
torch.logical_and(r.double(), s),
torch.logical_and(r, s, out=torch.empty(4, dtype=torch.bool)),
torch.logical_not(torch.tensor([0, 1, -10], dtype=torch.int8)),
torch.logical_not(
torch.tensor([0.0, 1.5, -10.0], dtype=torch.double)),
torch.logical_not(
torch.tensor([0.0, 1.0, -10.0], dtype=torch.double),
out=torch.empty(3, dtype=torch.int16),
),
torch.logical_or(r, s),
torch.logical_or(r.double(), s.double()),
torch.logical_or(r.double(), s),
torch.logical_or(r, s, out=torch.empty(4, dtype=torch.bool)),
torch.logical_xor(r, s),
torch.logical_xor(r.double(), s.double()),
torch.logical_xor(r.double(), s),
torch.logical_xor(r, s, out=torch.empty(4, dtype=torch.bool)),
torch.logit(torch.rand(5), eps=1e-6),
torch.hypot(torch.tensor([4.0]), torch.tensor([3.0, 4.0, 5.0])),
torch.i0(torch.arange(5, dtype=torch.float32)),
torch.igamma(a, b),
torch.igammac(a, b),
torch.mul(torch.randn(3), 100),
torch.multiply(torch.randn(4, 1), torch.randn(1, 4)),
torch.mvlgamma(torch.empty(2, 3).uniform_(1.0, 2.0), 2),
torch.tensor([float("nan"),
float("inf"), -float("inf"), 3.14]),
torch.nan_to_num(w),
torch.nan_to_num(w, nan=2.0),
torch.nan_to_num(w, nan=2.0, posinf=1.0),
torch.neg(torch.randn(5)),
# torch.nextafter(torch.tensor([1, 2]), torch.tensor([2, 1])) == torch.tensor([eps + 1, 2 - eps]),
torch.polygamma(1, torch.tensor([1.0, 0.5])),
torch.polygamma(2, torch.tensor([1.0, 0.5])),
torch.polygamma(3, torch.tensor([1.0, 0.5])),
torch.polygamma(4, torch.tensor([1.0, 0.5])),
torch.pow(a, 2),
torch.pow(torch.arange(1.0, 5.0), torch.arange(1.0, 5.0)),
torch.rad2deg(
torch.tensor([[3.142, -3.142], [6.283, -6.283],
[1.570, -1.570]])),
torch.randn(4, dtype=torch.cfloat).real,
torch.reciprocal(a),
torch.remainder(torch.tensor([-3.0, -2.0]), 2),
torch.remainder(torch.tensor([1, 2, 3, 4, 5]), 1.5),
torch.round(a),
torch.rsqrt(a),
torch.sigmoid(a),
torch.sign(torch.tensor([0.7, -1.2, 0.0, 2.3])),
torch.sgn(a),
torch.signbit(torch.tensor([0.7, -1.2, 0.0, 2.3])),
torch.sin(a),
torch.sinc(a),
torch.sinh(a),
torch.sqrt(a),
torch.square(a),
torch.sub(torch.tensor((1, 2)), torch.tensor((0, 1)), alpha=2),
torch.tan(a),
torch.tanh(a),
torch.trunc(a),
torch.xlogy(f, g),
torch.xlogy(f, g),
torch.xlogy(f, 4),
torch.xlogy(2, g),
)
def forward(x):
return TanhInv.alpha*torch.atanh(x)
def l2_attack(input,
target,
model,
targeted,
use_log,
use_tanh,
solver,
reset_adam_after_found=True,
abort_early=True,
batch_size=128,
max_iter=1000,
const=0.01,
confidence=0.0,
early_stop_iters=100,
binary_search_steps=9,
step_size=0.01,
adam_beta1=0.9,
adam_beta2=0.999):
early_stop_iters = early_stop_iters if early_stop_iters != 0 else max_iter // 10
input = torch.from_numpy(input).cuda()
target = torch.from_numpy(target).cuda()
var_len = input.view(-1).size()[0]
modifier_up = np.zeros(var_len, dtype=np.float32)
modifier_down = np.zeros(var_len, dtype=np.float32)
real_modifier = torch.zeros(input.size(), dtype=torch.float32).cuda()
mt = np.zeros(var_len, dtype=np.float32)
vt = np.zeros(var_len, dtype=np.float32)
adam_epoch = np.ones(var_len, dtype=np.int32)
grad = np.zeros(batch_size, dtype=np.float32)
hess = np.zeros(batch_size, dtype=np.float32)
upper_bound = 1e10
lower_bound = 0.0
out_best_attack = input.clone().detach().cpu().numpy()
out_best_const = const
out_bestl2 = 1e10
out_bestscore = -1
if use_tanh:
input = torch.atanh(input * 1.99999)
if not use_tanh:
modifier_up = 0.5 - input.clone().detach().view(-1).cpu().numpy()
modifier_down = -0.5 - input.clone().detach().view(-1).cpu().numpy()
def compare(x, y):
if not isinstance(x, (float, int, np.int64)):
if targeted:
x[y] -= confidence
else:
x[y] += confidence
x = np.argmax(x)
if targeted:
return x == y
else:
return x != y
for step in range(binary_search_steps):
bestl2 = 1e10
prev = 1e6
bestscore = -1
last_loss2 = 1.0
# reset ADAM status
mt.fill(0)
vt.fill(0)
adam_epoch.fill(1)
stage = 0
for iter in range(max_iter):
if (iter + 1) % 100 == 0:
loss, l2, loss2, _, __ = loss_run(input, target, model,
real_modifier, use_tanh,
use_log, targeted,
confidence, const)
print(
"[STATS][L2] iter = {}, loss = {:.5f}, loss1 = {:.5f}, loss2 = {:.5f}"
.format(iter + 1, loss[0], l2[0], loss2[0]))
sys.stdout.flush()
var_list = np.array(range(0, var_len), dtype=np.int32)
indice = var_list[np.random.choice(var_list.size,
batch_size,
replace=False)]
var = np.repeat(real_modifier.detach().cpu().numpy(),
batch_size * 2 + 1,
axis=0)
for i in range(batch_size):
var[i * 2 + 1].reshape(-1)[indice[i]] += 0.0001
var[i * 2 + 2].reshape(-1)[indice[i]] -= 0.0001
var = torch.from_numpy(var)
var = var.view((-1, ) + input.size()[1:]).cuda()
losses, l2s, losses2, scores, pert_images = loss_run(
input, target, model, var, use_tanh, use_log, targeted,
confidence, const)
real_modifier_numpy = real_modifier.clone().detach().cpu().numpy()
if solver == "adam":
coordinate_ADAM(losses,
indice,
grad,
hess,
batch_size,
mt,
vt,
real_modifier_numpy,
adam_epoch,
modifier_up,
modifier_down,
step_size,
adam_beta1,
adam_beta2,
proj=not use_tanh)
if solver == "newton":
coordinate_Newton(losses,
indice,
grad,
hess,
batch_size,
mt,
vt,
real_modifier_numpy,
adam_epoch,
modifier_up,
modifier_down,
step_size,
adam_beta1,
adam_beta2,
proj=not use_tanh)
real_modifier = torch.from_numpy(real_modifier_numpy).cuda()
if losses2[0] == 0.0 and last_loss2 != 0.0 and stage == 0:
if reset_adam_after_found:
mt.fill(0)
vt.fill(0)
adam_epoch.fill(1)
stage = 1
last_loss2 = losses2[0]
if abort_early and (iter + 1) % early_stop_iters == 0:
if losses[0] > prev * .9999:
print("Early stopping because there is no improvement")
break
prev = losses[0]
if l2s[0] < bestl2 and compare(
scores[0], np.argmax(target.cpu().numpy(), -1)):
bestl2 = l2s[0]
bestscore = np.argmax(scores[0])
if l2s[0] < out_bestl2 and compare(
scores[0], np.argmax(target.cpu().numpy(), -1)):
if out_bestl2 == 1e10:
print(
"[STATS][L3](First valid attack found!) iter = {}, loss = {:.5f}, loss1 = {:.5f}, loss2 = {:.5f}"
.format(iter + 1, losses[0], l2s[0], losses2[0]))
sys.stdout.flush()
out_bestl2 = l2s[0]
out_bestscore = np.argmax(scores[0])
out_best_attack = pert_images[0]
out_best_const = const
if compare(bestscore, np.argmax(target.cpu().numpy(),
-1)) and bestscore != -1:
print('old constant: ', const)
upper_bound = min(upper_bound, const)
if upper_bound < 1e9:
const = (lower_bound + upper_bound) / 2
print('new constant: ', const)
else:
print('old constant: ', const)
lower_bound = max(lower_bound, const)
if upper_bound < 1e9:
const = (lower_bound + upper_bound) / 2
else:
const *= 10
print('new constant: ', const)
return out_best_attack, out_bestscore
def _inverse(self, y):
# We do not clamp to the boundary here as it may degrade the performance of certain algorithms.
# one should use `cache_size=1` instead
return torch.atanh(y)
def backward(self, y):
self._low = torch.min(y)
self._high = torch.max(y)
return torch.atanh(y)