def forward(self, input, finished=None):
gamma_h = torch.tanh(self.linear_2(input))
weights = torch.sum(self.phi_hidden_state * gamma_h,
dim=2,
keepdim=True)
weights = torch.exp(weights -
torch.max(weights, dim=0, keepdim=True)[0])
weights = torch.divide(
weights, (1e-6 + torch.sum(weights, dim=0, keepdim=True)))
if self.field_size is not None:
alpha_h = torch.tanh(self.linear_5(input))
field_weights = torch.sum(self.phi_field_state * alpha_h,
dim=2,
keepdim=True)
field_weights = torch.exp(
field_weights -
torch.max(field_weights, dim=0, keepdim=True)[0])
field_weights = torch.divide(
field_weights,
(1e-6 + torch.sum(field_weights, dim=0, keepdim=True)))
weights = torch.divide(
weights * field_weights,
(1e-6 +
torch.sum(weights * field_weights, dim=0, keepdim=True)))
context = torch.sum(self.hidden_state * weights, dim=0)
out = self.linear_3(torch.cat([context, input], -1))
if finished is not None:
out = torch.where(finished, torch.zeros_like(out), out)
return out, weights
def _label_attention(self, x):
x_hat = torch.divide(x, torch.norm(x, dim=2, keepdim=True) + 1e-3)
l_hat = torch.divide(self.C,
torch.norm(self.C, dim=1, keepdim=True) + 1e-3)
G_hat = torch.matmul(l_hat.unsqueeze(dim=0), x_hat.permute(0, 2, 1))
return G_hat
def compute_wass_dist(a,
b,
M,
reg,
weights=None,
numItermax=1000,
stopThr=1e-4,
verbose=False,
log=False):
A = torch.vstack((a, b)).T
if weights is None:
weights = torch.ones(A.shape[1], dtype=torch.float64,
device=device) / A.shape[1]
else:
assert (len(weights) == 2)
if log:
log = {'err': []}
# M = M/np.median(M) # suggested by G. Peyre
#print (M)
K = torch.exp(torch.div(-M, reg))
#print (K)
cpt = 0
err = 1
# print (K.shape)
# print (torch.divide(A.T, torch.sum(K, axis=0)).T.shape)
UKv = torch.mm(K, torch.divide(A.T, torch.sum(K, axis=0)).T)
u = (geometricMean(UKv) / UKv.T).T
while (err > stopThr and cpt < numItermax):
cpt = cpt + 1
UKv = u * torch.mm(K, torch.divide(A, torch.mm(K, u)))
u = (u.T * geometricBar2(weights, UKv)).T / UKv
# print (u.shape)
if cpt % 10 == 1:
err = torch.sum(torch.std(UKv, axis=1))
# log and verbose print
if log:
log['err'].append(err)
if verbose:
if cpt % 200 == 0:
print('{:5s}|{:12s}'.format('It.', 'Err') + '\n' +
'-' * 19)
print('{:5d}|{:8e}|'.format(cpt, err))
if log:
log['niter'] = cpt
return geometricBar2(weights, UKv), log
else:
result = geometricBar2(weights, UKv)
#print(torch.sum(result*K))
print(reg * torch.sum(result))
#print((torch.add(torch.log(result), -1)))
return torch.sum(result * M)
def equi_basic(name, input_depth, output_depth, inputs):
'''
:param name: name of layer
:param input_depth: D
:param output_depth: S
:param inputs: N x D x m x m tensor
:return: output: N x S x m x m tensor
'''
basis_dimension = 4
# with tf.variable_scope(name, reuse=tf.AUTO_REUSE) as scope:
# initialization values for variables
coeffs_values = torch.matmul(torch.randn(size=(input_depth, output_depth, basis_dimension), dtype=torch.float32), torch.sqrt(2. / (input_depth + output_depth).type(torch.FloatTensor)))
# coeffs_values = tf.multiply(tf.random_normal([input_depth, output_depth, basis_dimension], dtype=tf.float32), tf.sqrt(2. / tf.to_float(input_depth + output_depth)))
#coeffs_values = tf.random_normal([input_depth, output_depth, basis_dimension], dtype=tf.float32)
# define variables
coeffs = torch.autograd.Variable(coeffs_values, requires_grad=True)
# coeffs = tf.get_variable('coeffs', initializer=coeffs_values)
m = inputs.shape[3].type(torch.IntTensor) # extract dimension
# m = tf.to_int32(tf.shape(inputs)[3]) # extract dimension
float_dim = m.type(torch.FloatTensor)
# float_dim = tf.to_float(m)
# apply ops
ops_out = []
# w1 - identity
ops_out.append(inputs)
# w2 - sum cols
sum_of_cols = torch.divide(torch.sum(inputs, dim=2), float_dim) # N x D x m
# sum_of_cols = tf.divide(tf.reduce_sum(inputs, axis=2), float_dim) # N x D x m
ops_out.append(torch.unsqueeze(sum_of_cols, dim=2).repeat(1, 1, m, 1)) # N x D x m x m
# ops_out.append(tf.tile(tf.expand_dims(sum_of_cols, axis=2), [1, 1, m, 1])) # N x D x m x m
# w3 - sum rows
sum_of_rows = torch.divide(torch.sum(inputs, dim=3), float_dim) # N x D x m
# sum_of_rows = tf.divide(tf.reduce_sum(inputs, axis=3), float_dim) # N x D x m
ops_out.append(torch.unsqueeze(sum_of_rows, dim=3).repeat(1, 1, 1, m)) # N x D x m x m
# ops_out.append(tf.tile(tf.expand_dims(sum_of_rows, axis=3), [1, 1, 1, m])) # N x D x m x m
# w4 - sum all
sum_all = torch.divide(torch.sum(sum_of_rows, dim=2), torch.square(float_dim)) # N x D
# sum_all = tf.divide(tf.reduce_sum(sum_of_rows, axis=2), tf.square(float_dim)) # N x D
ops_out.append(torch.unsqueeze(torch.unsqueeze(sum_all, dim=2), dim=3).repeat(1, 1, m, m)) # N x D x m x m
# ops_out.append(tf.tile(tf.expand_dims(tf.expand_dims(sum_all, axis=2), axis=3), [1, 1, m, m])) # N x D x m x m
ops_out = torch.stack(ops_out, dim=2)
# ops_out = tf.stack(ops_out, axis=2)
output = torch.einsum('dsb,ndbij->nsij', coeffs, ops_out) # N x S x m x m
# output = tf.einsum('dsb,ndbij->nsij', coeffs, ops_out) # N x S x m x m
# bias
bias = torch.autograd.Variable(torch.zeros((1, output_depth, 1, 1), dtype=torch.float32), requires_grad=True)
# bias = tf.get_variable('bias', initializer=tf.zeros([1, output_depth, 1, 1], dtype=tf.float32))
output = output + bias
return output
def cal_loss(fs_list, ft_list, criterion):
tot_loss = 0
for i in range(len(ft_list)):
fs = fs_list[i]
ft = ft_list[i]
_, _, h, w = fs.shape
fs_norm = torch.divide(fs, torch.norm(fs, p=2, dim=1, keepdim=True))
ft_norm = torch.divide(ft, torch.norm(ft, p=2, dim=1, keepdim=True))
f_loss = (0.5 / (w * h)) * criterion(fs_norm, ft_norm)
tot_loss += f_loss
return tot_loss
def get_regularized_slice_dist_matrix(x,y,z,reg):
points1 = torch.vstack((torch.arange(x, device=device, dtype=torch.float32),torch.zeros(x, device=device))).T
dist_matrix1 = (torch.cdist(points1,points1)/(x-1))**2
K1 = torch.divide(dist_matrix1, -reg)
torch.exp(K1, out=K1)
points2 = torch.vstack((torch.arange(y, device=device, dtype=torch.float32),torch.zeros(y, device=device))).T
dist_matrix2 = (torch.cdist(points2,points2)/(y-1))**2
K2 = torch.divide(dist_matrix2, -reg)
torch.exp(K2, out=K2)
return K1,K2
def var(input: Tensor,
dim: DimOrDims = None,
unbiased: Optional[bool] = False,
*,
keepdim: Optional[bool] = False,
dtype: Optional[DType] = None,
mask: Optional[Tensor] = None) -> Tensor:
"""\
{reduction_signature}
{reduction_descr}
The identity value of sample variance operation is undefined. The
elements of output tensor with strided layout, that correspond to
fully masked-out elements, have ``nan`` values.
{reduction_args}
{reduction_example}"""
if dtype is None:
dtype = input.dtype
if not (dtype.is_floating_point or dtype.is_complex):
dtype = torch.float32
compute_dtype = dtype
if not (compute_dtype.is_floating_point or compute_dtype.is_complex):
compute_dtype = torch.float32
if input.layout == torch.strided:
inmask = _input_mask(input, mask=mask)
count = sum(inmask.new_ones(input.shape, dtype=torch.int64),
dim,
keepdim=True,
mask=inmask)
sample_total = sum(input, dim, keepdim=True, dtype=dtype, mask=inmask)
# TODO: replace torch.subtract/divide/square/maximum with
# masked subtract/divide/square/maximum when these will be
# available.
sample_mean = torch.divide(sample_total, count)
x = torch.subtract(input, sample_mean)
total = sum(x * x.conj(),
dim,
keepdim=keepdim,
dtype=compute_dtype,
mask=inmask)
if not keepdim:
count = count.reshape(total.shape)
if unbiased:
count = torch.subtract(count, 1)
count = torch.maximum(count, count.new_zeros([]))
return torch.divide(total, count).to(dtype=dtype)
else:
raise ValueError(
f'masked var expects strided tensor (got {input.layout} tensor)')
def cal_anomaly_map(fs_list, ft_list, out_size=256):
pdist = torch.nn.PairwiseDistance(p=2, keepdim=True)
anomaly_map = torch.ones([ft_list[0].shape[0], 1, out_size,
out_size]).to(device)
for i in range(len(ft_list)):
fs = fs_list[i]
ft = ft_list[i]
fs_norm = torch.divide(fs, torch.norm(fs, p=2, dim=1, keepdim=True))
ft_norm = torch.divide(ft, torch.norm(ft, p=2, dim=1, keepdim=True))
a_map = 0.5 * pdist(fs_norm, ft_norm)**2
a_map = F.interpolate(a_map, size=out_size, mode='bilinear')
anomaly_map *= a_map
return anomaly_map
def update(self, X, y, learning_rate):
z = torch.mul(self.forward(X), y)
grad_W = torch.divide(torch.mul(X, y), (1 + torch.exp(z)))
grad_b = torch.divide(y, 1 + torch.exp(z))
grad_W = -1 * torch.mean(grad_W, dim=0).unsqueeze(dim=1)
grad_b = -1 * torch.mean(grad_b)
self.W -= learning_rate * grad_W
self.b -= learning_rate * grad_b
return grad_W, grad_b
def elbo_grad(Gl, logWl):
nodes = list(
set([
list(G.keys())[i] for G in Gl for i in range(len(list(G.keys())))
]))
g_hat = {}
L = len(logWl)
for node in nodes:
F = []
for l in range(L):
if node in list(Gl[l].keys()):
F.append(torch.multiply(Gl[l][node], logWl[l]))
else:
F.append(torch.tensor([0., 0.]))
Gl[l][node] = torch.tensor([0., 0.])
b_hat = sum([
np.cov(F[l].detach(), Gl[l][node].detach())[0, 1] for l in range(L)
]) / sum([
np.var([Gl[l][node][i].detach().numpy() for l in range(L)])
for i in range(len(Gl[l][node]))
])
# b_hat = 0
temp = [np.multiply(-1 * b_hat, Gl[l][node]) for l in range(L)]
Ftemp = [torch.add(F[i], temp[i]) for i in range(len(F))]
sumFtemp = torch.stack(Ftemp, dim=0).sum(dim=0)
g_hat[node] = torch.divide(sumFtemp, L)
return g_hat
def forward(self, x, adj):
feature_dim = int(adj.shape[-1])
eye = torch.eye(feature_dim).cuda()
if x is None:
AXW = torch.tensordot(
adj, self.kernel,
[[-1], [0]]) # batch_size * num_node * feature_dim
else:
XW = torch.tensordot(
x, self.kernel, [[-1], [0]]) # batch * num_node * feature_dim
AXW = torch.matmul(adj, XW) # batch * num_node * feature_dim
I_cAXW = eye + self.c * AXW
y_relu = torch.nn.functional.relu(I_cAXW)
temp = torch.mean(input=y_relu, dim=-2, keepdim=True) + 1e-6
col_mean = temp.repeat([1, feature_dim, 1])
y_norm = torch.divide(y_relu, col_mean)
output = torch.nn.functional.softplus(y_norm)
if self.neg_penalty != 0:
neg_loss = torch.multiply(
torch.tensor(self.neg_penalty),
torch.sum(torch.nn.functional.relu(1e-6 - self.kernel)))
self.losses.append(neg_loss)
return output
def step(self, reward, act_scalar=1.0, alpha=0.2, epsilon=0.5):
# Update rewards
self.rewards[self.offsets + self.indices] += torch.divide(
reward - self.rewards[self.offsets + self.indices],
self.divs[self.offsets + self.indices])
self.divs[self.offsets + self.indices] += alpha
# Find new indices
self.indices = (
torch.argmax(self.rewards.reshape(
(self.num_params, self.resolution)),
dim=1).float() +
torch.randn(self.num_params, device=self.device) * epsilon +
0.5).long().clamp(min=0, max=self.resolution - 1)
# Generate new parameters
start_index = 0
for p in self.module.parameters():
p.data = (
act_scalar *
(self.indices[start_index:start_index + p.numel()].float() /
self.resolution * 2.0 - 1.0)).reshape(p.shape)
start_index += p.numel()
def masked_mse_loss(y_pred, y_true, null_val):
mask = torch.ne(y_true, null_val).float()
mask = torch.divide(mask, torch.mean(mask))
loss = torch.square(y_pred - y_true)
loss = torch.mul(loss, mask)
loss[torch.isnan(loss)] = 0.
return torch.mean(loss)
def create_torch_stft(samples, FFT=False):
'''
ARGUMENTS
> samples - a pytorch tensor of dimensions (# samples, sample length)
> FFT - if True, FFT is calculated instead of STFT
RETURNS
> spec - a tensor of dimension (# samples, frequency bins)
'''
if not FFT:
# Create Spectrogram
spec = torch.stft(samples,
n_fft=2048,
hop_length=512,
normalized=True,
return_complex=True).abs()
# Collapse spectrogram by summing along the time axis
spec = torch.sum(spec, dim=2)
else:
spec = torch.fft.fft(samples, dim=1).abs()
spec -= spec.min(1, keepdim=True)[0]
spec /= spec.max(1, keepdim=True)[0]
# Normalize
spec_max = torch.max(spec, axis=1)[0].unsqueeze(1).repeat(1, spec.shape[1])
spec_min = torch.min(spec, axis=1)[0].unsqueeze(1).repeat(1, spec.shape[1])
spec = torch.subtract(spec, spec_min)
spec = torch.divide(spec, torch.subtract(spec_max, spec_min))
spec = spec[:, :11025]
return spec.type(torch.float32)
def error_fn_l2_normalized(predictions, dataset):
actual = dataset.X
errors = predictions - actual
error_norms = torch.linalg.norm(torch.tensor(errors), dim=-1, ord=2)
actual_norms = torch.linalg.norm(torch.tensor(actual), dim=-1, ord=2)
normalized_errors = torch.divide(error_norms, actual_norms)
return normalized_errors
def get_group_delay(
raw_data: torch.Tensor,
sampling_rate_in_hz: int,
window_length_in_s: float,
window_shift_in_s: float,
num_fft_points: int,
window_type: str,
):
X_stft_transform = _get_stft(raw_data,
sampling_rate_in_hz,
window_length_in_s,
window_shift_in_s,
num_fft_points,
window_type=window_type)
Y_stft_transform = _get_stft(
raw_data,
sampling_rate_in_hz,
window_length_in_s,
window_shift_in_s,
num_fft_points,
window_type=window_type,
data_transformation="group_delay",
)
X_stft_transform_real = torch.real(X_stft_transform)
X_stft_transform_imag = torch.imag(X_stft_transform)
Y_stft_transform_real = torch.real(Y_stft_transform)
Y_stft_transform_imag = torch.imag(Y_stft_transform)
nominator = torch.multiply(
X_stft_transform_real, Y_stft_transform_real) + torch.multiply(
X_stft_transform_imag, Y_stft_transform_imag)
denominator = torch.square(torch.abs(X_stft_transform))
group_delay = torch.divide(nominator, denominator + 1e-10)
assert not torch.isnan(
group_delay).any(), "There are NaN values in group delay"
return torch.transpose(group_delay, 0, 1)
def topk3d(tensor, k=1):
n = tensor.size(0)
c = tensor.size(1)
d = tensor.size(2)
idx = tensor.contiguous().view(n, -1).topk(k, dim=1)
return torch.cat((torch.divide(idx, d**2, rounding_mode='trunc').view(-1, 1),
(idx % d**2 / d).view(-1, 1),
(idx % d**2 / d).view(-1, 1)), dim=1)
def sample2d(tensor, k=1):
if tensor.dim() == 2:
n = 1
d = tensor.size(-1)
else:
n = tensor.size(0)
d = tensor.size(-1)
idx = torch.multinomial(tensor.reshape(n, -1), k)
return torch.cat((torch.divide(idx, d, rounding_mode='trunc').view(-1, 1), (idx % d).view(-1, 1)), dim=1)
def batch_top_k(batch, k):
bsz, dim = batch.shape
flatten = batch.flatten()
top_values, top_ids = torch.topk(flatten, k=k)
batch_ids = torch.divide(top_ids, dim, rounding_mode="trunc")
dim_ids = torch.remainder(top_ids, dim)
return top_values, batch_ids, dim_ids
def tanimoto_loss(label, pred):
square = torch.square(pred)
sum_square = torch.sum(square)
product = torch.multiply(pred, label)
sum_product = torch.sum(product)
denomintor = torch.subtract(torch.add(sum_square, 1), sum_product)
loss = torch.divide(sum_product, denomintor)
loss = torch.reduce_mean(loss)
return 1.0 - loss
def forward(self, inputs):
inputs = inputs.permute(0, 2, 3, 4, 1).contiguous()
input_shape = inputs.shape
# Flatten input
flat_inputs = inputs.view(-1, self._embedding_dim)
dist = (
torch.sum(flat_inputs.detach()**2, dim=1, keepdims=True) +
torch.sum(self._embedding.weight**2, dim=1)) - 2 * torch.matmul(
flat_inputs.detach(), self._embedding.weight.t())
sij_numer = (1 + dist / self.df).pow(-0.5 * (self.df + 1))
sij_denom = sij_numer.sum(dim=1, keepdims=True)
sij = torch.divide(sij_numer, sij_denom)
sij_Kappa = sij.pow(self.Kappa)
sumi_sij = sij.sum(dim=0, keepdims=True)
tij_numer = torch.divide(sij_Kappa, sumi_sij)
tij_denom = tij_numer.sum(dim=1, keepdims=True)
tij = torch.divide(tij_numer, tij_denom)
log_sij = torch.log(sij)
log_tij = torch.log(tij)
CAH_loss = torch.sum(torch.multiply(tij, log_tij - log_sij))
log_sie = torch.matmul(log_sij, self.adjacency)
sumj_sij_log_sie = torch.sum(torch.multiply(sij, log_sie), dim=1)
SOM_loss = -sumj_sij_log_sie.mean()
total_loss = self.gamma * CAH_loss + self.beta * SOM_loss
return ({
"CAH_loss": CAH_loss,
"SOM_loss": SOM_loss,
"vq_loss": total_loss
})
def std_var(input: Tensor,
dim: DimOrDims = None,
unbiased: Optional[bool] = False,
*,
keepdim: Optional[bool] = False,
dtype: Optional[DType] = None,
mask: Optional[Tensor] = None,
take_sqrt: Optional[bool] = False) -> Tensor:
if dtype is None:
dtype = input.dtype
if not (dtype.is_floating_point or dtype.is_complex):
dtype = torch.float32
compute_dtype = dtype
if not (compute_dtype.is_floating_point or compute_dtype.is_complex):
compute_dtype = torch.float32
if input.layout == torch.strided:
if mask is None:
# TODO: compute count analytically
count = sum(torch.ones(input.shape, dtype=torch.int64, device=input.device), dim, keepdim=True)
sample_total = sum(input, dim, keepdim=True, dtype=dtype)
else:
inmask = _input_mask(input, mask=mask)
count = sum(inmask.new_ones(input.shape, dtype=torch.int64), dim, keepdim=True, mask=inmask)
sample_total = sum(input, dim, keepdim=True, dtype=dtype, mask=inmask)
# TODO: replace torch.subtract/divide/square/maximum with
# masked subtract/divide/square/maximum when these will be
# available.
sample_mean = torch.divide(sample_total, count)
x = torch.subtract(input, sample_mean)
if mask is None:
total = sum(x * x.conj(), dim, keepdim=keepdim, dtype=compute_dtype)
else:
total = sum(x * x.conj(), dim, keepdim=keepdim, dtype=compute_dtype, mask=inmask)
if not keepdim:
count = count.reshape(total.shape)
if unbiased:
count = torch.subtract(count, 1)
count = torch.maximum(count, count.new_zeros([]))
output = torch.divide(total, count).to(dtype=dtype)
if take_sqrt:
output = torch.sqrt(output)
return output
else:
raise ValueError(f'masked std/var expects strided tensor (got {input.layout} tensor)')
def forward(self, x, a):
"""Performs spatial dilation of a periodic function by Fourier
transform. Uses the observation that the Fourier transform has the
following time scaling relation:
f(x) <-----> F(k)
f(ax) <----> 1/|a| F(k/a)
Args:
x (torch.tensor): Observations to be scaled. A tensor of shape
(n_batch, n_channels, n_grid_points). The scaling will be performed
across the last dimension.
a (float): Scaling parameter. x -> lambda * x
"""
a_ = int(a)
# Take Fourier transform
x_dft = torch.fft.fft(x)
# Scale the frequencies k -> k/a
# Idea: we want a new DFT vector which has shape x_dft.shape[-1] * a
# In this vector we want zeros for non-integer values of k/a and the
# elements of x_dft for the integer values of k/a. We will do this via
# 1d convolution with a kernel of zeros except for the first entry
convo_kernel = torch.zeros(x_dft.shape[0], 1, a_, dtype=torch.cfloat)
convo_kernel[:, 0, 0] = torch.ones_like(convo_kernel[:, 0, 0])
convo_kernel.requires_grad = False
print(convo_kernel)
out_x_dft = torch.zeros(
(x_dft.shape[0], x_dft.shape[1], a_ * x_dft.shape[-1]),
dtype=torch.cfloat)
# print(out_x_dft.shape)
for i in range(x_dft.shape[-1]):
lb = a_ * i
ub = a_ * (i + 1)
arg_1 = x_dft[:, :, i].view(x_dft.shape[0], x_dft.shape[1], 1)
# ignoring the batch dimension, convo_kernel is of size (1, a)
# and arg_1 is of size (n_channels, 1)
out_x_dft[:, :, lb:ub] = torch.matmul(arg_1, convo_kernel)
x_dft = out_x_dft
# print(x_dft.shape)
# Scale the whole thing by 1/|a|
x = torch.divide(x_dft, np.abs(a))
x_ifft = torch.fft.ifft(x)
return x_ifft
def argmax4d(tensor):
n = tensor.size(0)
c1 = tensor.size(1)
c2 = tensor.size(2)
d = tensor.size(3)
idx = tensor.contiguous().view(n, -1).topk(k, dim=1)
return torch.cat((torch.divide(idx, d**2 * c2, rounding_mode='trunc').view(-1, 1),
(idx % (d**2 *c2) / d**2).view(-1, 1),
(((idx % (d**2 * c2)) % d**2) / d).view(-1, 1),
(((idx % (d**2 * c2)) % d**2) % d).view(-1, 1)), dim=1)
def average_precision(predictions, ground_truth, iou_threshold):
# return the average precision of a class
# predictions [train_idx, pred_class, confidence, x, y, w, h]
predictions.sort(key=lambda x: x[2], reverse=True)
TP_ = torch.zeros((len(predictions)))
FP_ = torch.zeros((len(predictions)))
gt_used = torch.zeros((len(ground_truth)))
for pred_id, pred in enumerate(predictions):
# get the label in same image
# print("prediction: " + str(pred))
gt_ = []
for bid, bbox in enumerate(ground_truth):
if int(bbox[0]) == int(pred[0]) and int(gt_used[bid]) == 0:
nbb = [bid] + bbox
# print(nbb)
gt_.append(nbb)
if len(gt_) == 0:
FP_[pred_id] = 1
continue
best_iou = 0
best_id = 0
for gt in gt_:
iou_ = intersection_over_union(torch.tensor(pred[3:]),
torch.tensor(gt[4:]),
box_format="midpoint")
# print("iou_: " + str(iou_))
if iou_ > best_iou:
best_iou = iou_
best_id = gt[0]
if best_iou >= iou_threshold and int(
gt_used[best_id]) == 0: # ground truth haven't checked
TP_[pred_id] = 1
gt_used[best_id] = 1
else:
FP_[pred_id] = 1
TP_cumsum = torch.cumsum(TP_, dim=0)
FP_cumsum = torch.cumsum(FP_, dim=0)
rec = TP_cumsum / (len(ground_truth) + 1e-6)
pre = torch.divide(TP_cumsum, (TP_cumsum + FP_cumsum + 1e-6))
pre = torch.cat((torch.tensor([1]), pre))
rec = torch.cat((torch.tensor([0]), rec))
# print("gt_used: " + str(gt_used))
# print("pre: " + str(pre))
# print("rec: " + str(rec))
auc = torch.trapz(pre, rec, dim=-1).item()
# print("AP: " + str(auc))
return auc
def forward(self, x, frequencies, x_grid):
# Input has size (n_batch, n_channels, n_x_points)
n_batch, n_channels, n_x_points = x.shape
x = x.type(torch.cfloat)
exp_argument = torch.mul(frequencies, 1j * 2 * np.pi)
exp_multiplicand = torch.exp(torch.outer(x_grid, exp_argument))
# x has shape b,c,s and exp_multiplicand has shape f,s and we want
# output of shape b,c,f
out = torch.einsum('bcs,fs->bcf', x, exp_multiplicand)
out = torch.divide(out, n_x_points)
return out
def mean_average_precision(self, y_que, X_que, y_pool, X_pool):
if self.rank == None:
raise ("rank function is not provided.")
n_que = len(y_que)
n_pool = len(y_pool)
ap = th.zeros(n_que, device=self.device)
for i in range(n_que):
y = y_que[i]
ranks = self.rank(X_que[i], X_pool)
rel = y_pool[ranks] == y
pre_k = th.cumsum(rel, dim=0) / th.arange(
1, n_pool + 1, device=self.device)
ap[i] = th.divide(th.sum(pre_k * rel), th.sum(rel))
return th.nansum(ap) / (n_que - th.sum(th.isnan(ap)))
def forward(ctx,
input,
scale,
scale_grad,
thd_neg,
thd_pos,
act_mode=True):
# scale grad, thd_neg, thd_pos is scaler else tensor
x_d_s = t.divide(input / scale)
x_ = t.round(t.clamp(x_d_s, thd_neg, thd_pos))
output = x_ * scale
ctx.save_for_backward(x_d_s, x_, scale_grad, thd_neg, thd_pos,
act_mode)
return output
def compute_affinity_matrix(batch_patches: torch.tensor) -> torch.tensor:
"""
Function that computes the affinity matrix for every patch in a batch
:param batch_patches: tensor with shape (batch_size, patch_width, patch_height, no_of_channels)
:return affinity_matrix: torch.tensor containing the affinity Matrix for every patch in the batch
"""
_, h, w, c = batch_patches.shape
x1: torch.tensor = batch_patches.reshape(-1, h * w, c).unsqueeze(1)
x2: torch.tensor = batch_patches.reshape(-1, h * w, c).unsqueeze(2)
diff_image = torch.linalg.norm(x2 - x1, dim=-1)
kernel = torch.topk(diff_image, h * w).values
kernel = torch.mean(kernel[:, :, (h * w) // 4], dim=1)
kernel = torch.reshape(kernel, (-1, 1, 1))
affinity_matrix = torch.exp(-(torch.divide(diff_image, kernel)**2))
return affinity_matrix
def infer_full_image(self, input, C_out, kernel_size=256, stride=128):
self.generator.eval()
B, C, W, H = input.shape
pad_W = kernel_size - W % kernel_size
pad_H = kernel_size - H % kernel_size
x, _, _ = compute_pyramid_patch_weight_loss(kernel_size, kernel_size)
input = F.pad(input, (0, pad_H, 0, pad_W), mode="reflect").squeeze(0)
_, W_pad, H_pad = input.shape
patches = input.unfold(1, kernel_size,
stride).unfold(2, kernel_size, stride)
c, n_w, n_h, w, h = patches.shape
patches = patches.contiguous().view(c, -1, kernel_size, kernel_size)
dataset = torch.utils.data.TensorDataset(patches.permute(1, 0, 2, 3))
batch_size = 4
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size)
op = []
for batch_idx, sample1 in enumerate(dataloader):
patch_op = self.generator(sample1[0])
op.append(patch_op)
op = torch.cat(op).permute(1, 0, 2, 3)
op = op.permute(0, 2, 3, 1).reshape(1, -1, n_w * n_h)
weights_op = (torch.from_numpy(x).unsqueeze(0).unsqueeze(-1).repeat(
1, C_out, 1, n_w * n_h).reshape(1, -1, n_w * n_h)).cuda()
op = torch.mul(weights_op, op)
op = F.fold(
op,
output_size=(W_pad, H_pad),
kernel_size=(kernel_size, kernel_size),
stride=(stride, stride),
)
weights_op = F.fold(
weights_op,
output_size=(W_pad, H_pad),
kernel_size=(kernel_size, kernel_size),
stride=(stride, stride),
)
op = torch.divide(op, weights_op)
output = op #torch.clamp(op, 0.0, 1.0)
output = output[:, :, :W, :H]
return output
