def forward(self, x):
x1 = self.encoder(x[0])
x2 = self.encoder(x[1])
x3 = self.encoder(x[2])
dap = torch.pairwise_distance(x1, x2)
dan = torch.pairwise_distance(x1, x3)
return dap, dan, self.decoder(x1), self.decoder(x3)
def mmd(x, y, sigmas=[1, 5, 10], normalize=False):
"""Maximum Mean Discrepancy with several Gaussian kernels."""
# Flatten:
x = x.view(x.shape[0], -1)
y = y.view(y.shape[0], -1)
if len(sigmas) == 0:
mean_dist = torch.mean(torch.pow(torch.pairwise_distance(x, y, p=2), 2))
factors = (-1 / (2 * mean_dist)).view(1, 1, 1)
else:
factors = _get_mmd_factor(sigmas, x.device)
if normalize:
x = F.normalize(x, p=2, dim=1)
y = F.normalize(y, p=2, dim=1)
xx = torch.pairwise_distance(x, x, p=2)**2
yy = torch.pairwise_distance(y, y, p=2)**2
xy = torch.pairwise_distance(x, y, p=2)**2
k_xx, k_yy, k_xy = 0, 0, 0
div = 1 / (x.shape[1]**2)
k_xx = div * torch.exp(factors * xx).sum(0).squeeze()
k_yy = div * torch.exp(factors * yy).sum(0).squeeze()
k_xy = div * torch.exp(factors * xy).sum(0).squeeze()
mmd_sq = torch.sum(k_xx) - 2 * torch.sum(k_xy) + torch.sum(k_yy)
return torch.sqrt(mmd_sq)
def forward(self, x1, x2, y):
d = torch.pairwise_distance(x1, x2, keepdim=False)
margin = 1.
d = torch.pairwise_distance(x1, x2, keepdim=False)
zero_tensor = torch.Tensor([0] * d.shape[0])
loss = 0.5 * y * d * d + 0.5 * (1 - y) * torch.max(
zero_tensor, margin - d) * torch.max(zero_tensor, margin - d)
return torch.mean(loss)
def tripletmarginloss_reference(anchor, positive, negative, margin=1.0, p=2, eps=1e-6, swap=False,
size_average=True, reduce=True):
d_p = torch.pairwise_distance(anchor, positive, p, eps)
d_n = torch.pairwise_distance(anchor, negative, p, eps)
if swap:
d_s = torch.pairwise_distance(positive, negative, p, eps)
d_n = torch.min(d_n, d_s)
output = torch.clamp(margin + d_p - d_n, min=0.0)
if reduce and size_average:
return output.mean()
elif reduce:
return output.sum()
return output
def tripletmarginloss_reference(anchor, positive, negative, margin=1.0, p=2, eps=1e-6, swap=False,
reduction='elementwise_mean'):
d_p = torch.pairwise_distance(anchor, positive, p, eps)
d_n = torch.pairwise_distance(anchor, negative, p, eps)
if swap:
d_s = torch.pairwise_distance(positive, negative, p, eps)
d_n = torch.min(d_n, d_s)
output = torch.clamp(margin + d_p - d_n, min=0.0)
if reduction == 'elementwise_mean':
return output.mean()
elif reduction == 'sum':
return output.sum()
return output
def outer_pairwise_distance(A, B=None):
"""
Compute pairwise_distance of Tensors
A (size(A) = n x d, where n - rows count, d - vector size) and
B (size(A) = m x d, where m - rows count, d - vector size)
return matrix C (size n x m), such as C_ij = distance(i-th row matrix A, j-th row matrix B)
if only one Tensor was given, computer pairwise distance to itself (B = A)
"""
if B is None: B = A
max_size = 2**26
n = A.size(0)
m = B.size(0)
d = A.size(1)
if n * m * d <= max_size or m == 1:
return torch.pairwise_distance(
A[:, None].expand(n, m, d).reshape((-1, d)),
B.expand(n, m, d).reshape((-1, d))).reshape((n, m))
else:
batch_size = max(1, max_size // (n * d))
batch_results = []
for i in range((m - 1) // batch_size + 1):
id_left = i * batch_size
id_rigth = min((i + 1) * batch_size, m)
batch_results.append(
outer_pairwise_distance(A, B[id_left:id_rigth]))
return torch.cat(batch_results, dim=1)
def stest_one_versus_many(model, data_dir, img_size):
""" """
data_iterator = iter(
DataLoader(
PairDataset(
data_dir,
transform=transforms.Compose([
transforms.Grayscale(),
transforms.Resize(img_size),
transforms.ToTensor(),
]),
split=SplitEnum.testing,
),
num_workers=0,
batch_size=1,
shuffle=True,
))
x0, *_ = next(data_iterator)
for i in range(10):
_, x1, _ = next(data_iterator)
dis = (torch.pairwise_distance(*model(
to_tensor(x0, device=global_torch_device()),
to_tensor(x1, device=global_torch_device()),
)).cpu().item())
boxed_text_overlay_plot(
torchvision.utils.make_grid(torch.cat((x0, x1), 0)),
f"Dissimilarity: {dis:.2f}",
)
def minimal_mean_distance(self, other, p=2):
mean_distances = []
for i in range(self.vertices):
rolled = self.roll(i).float()
distances = torch.pairwise_distance(rolled, other.to_float(), p=p)
mean_distances.append(distances.mean().item())
return min(mean_distances)
def _test(metric_device):
engine = Engine(update)
m = MeanPairwiseDistance(device=metric_device)
m.attach(engine, "mpwd")
data = list(range(n_iters))
engine.run(data=data, max_epochs=1)
assert "mpwd" in engine.state.metrics
res = engine.state.metrics["mpwd"]
true_res = []
for i in range(n_iters * idist.get_world_size()):
true_res.append(
torch.pairwise_distance(
y_true[i * s : (i + 1) * s, ...], y_preds[i * s : (i + 1) * s, ...], p=m._p, eps=m._eps
)
.cpu()
.numpy()
)
true_res = np.array(true_res).ravel()
true_res = true_res.mean()
assert pytest.approx(res) == true_res
def get_working_points(self, spatial1, spatial2, truth, margin):
"""
Args:
spatial (``torch.tensor``, required): The spatial embedding of the data
truth (``torch.tensor``, required): The truth graph of the data
Returns:
``torch.tensor``: The R95, R98, R99 values
"""
margin_multiple = 1
max_dist = torch.tensor(margin_multiple * margin).float().to(
self.device)
distances = torch.pairwise_distance(spatial1[truth[0]],
spatial2[truth[1]])
# Sort the distances
distances, indices = torch.sort(distances, descending=False)
working_radii = [
min(
distances[int(
len(distances) * (working_point_efficiency / 100))],
max_dist).item()
for working_point_efficiency in self.hparams["working_points"]
]
return working_radii
def eval(model, criterion, dataloader, log_interval=50):
model.eval()
running_loss = 0
number_samples = 0
distances = []
for batch_idx, (x1, x2, y) in enumerate(dataloader):
x1, x2, y = x1.to(device), x2.to(device), y.to(device)
x1, x2 = model(x1, x2)
loss = criterion(x1, x2, y)
distances.extend(
zip(
torch.pairwise_distance(x1, x2, 2).cpu().tolist(),
y.cpu().tolist()))
number_samples += len(x1)
running_loss += loss.item() * len(x1)
if (batch_idx +
1) % log_interval == 0 or batch_idx == len(dataloader) - 1:
print('{}/{}: Loss: {:.4f}'.format(batch_idx + 1, len(dataloader),
running_loss / number_samples))
distances, y = zip(*distances)
distances, y = torch.tensor(distances), torch.tensor(y)
max_accuracy = accuracy(distances, y)
print(f'Max accuracy: {max_accuracy}')
return running_loss / number_samples, max_accuracy
def __call__(self, x):
"""Clustering.
Args:
x (Tensor) - Data points of number n by feature dim m.
"""
idx = np.random.choice(len(x), self.n_clusters)
state = x[idx]
while True:
pre_state = state.clone()
dis = setwise_distance(x, state).squeeze()
result = torch.argmin(dis, dim=1)
for i in range(self.n_clusters):
idx = torch.nonzero(result == i).squeeze()
items = torch.index_select(x, 0, idx)
if items.size(0):
state[i] = items.mean(dim=0)
else:
state[i] = pre_state[i].clone()
shift = torch.pairwise_distance(pre_state, state)
total = torch.pow(torch.sum(shift), 2.0)
if total < self.tol:
return result, state
def stest_many_versus_many(model, data_dir, img_size, threshold=0.5):
""" """
data_iterator = iter(
DataLoader(
PairDataset(
data_dir,
transform=transforms.Compose([
transforms.Grayscale(),
transforms.Resize(img_size),
transforms.ToTensor(),
]),
),
num_workers=0,
batch_size=1,
shuffle=True,
))
for i in range(10):
x0, x1, is_diff = next(data_iterator)
distance = (torch.pairwise_distance(*model(
to_tensor(x0, device=global_torch_device()),
to_tensor(x1, device=global_torch_device()),
)).cpu().item())
boxed_text_overlay_plot(
torchvision.utils.make_grid(torch.cat((x0, x1), 0)),
f"Truth: {'Different' if is_diff.cpu().item() else 'Alike'},"
f" Dissimilarity: {distance:.2f},"
f" Verdict: {'Different' if distance > threshold else 'Alike'}",
)
def get_working_points(self, spatial1, spatial2, truth):
"""
Args:
spatial (``torch.tensor``, required): The spatial embedding of the data
truth (``torch.tensor``, required): The truth graph of the data
Returns:
``torch.tensor``: The R95, R98, R99 values
"""
margin_multiple = 1
max_dist = torch.tensor(margin_multiple *
self.hparams["margin"]).float().to(self.device)
# Get the R95, R98, R99 values
# distances = torch.sum((spatial[truth[0]] - spatial[truth[1]])**2, dim=-1)
distances = torch.pairwise_distance(spatial1[truth[0]],
spatial2[truth[1]])
# Sort the distances
distances, indices = torch.sort(distances, descending=False)
# Get the indices of the 95th, 98th, 99th percentile
R95 = min(distances[int(len(distances) * 0.95)], max_dist)
R98 = min(distances[int(len(distances) * 0.98)], max_dist)
R99 = min(distances[int(len(distances) * 0.99)], max_dist)
return R95.item(), R98.item(), R99.item()
def get_distance(trajectories):
euclid_distance = []
for a, b in zip(trajectories[:, :], trajectories[1:, :]):
dist = torch.pairwise_distance(a, b)
dist = dist.cpu().detach().numpy()
euclid_distance.append(dist.reshape(1, -1))
euclid_distance = torch.from_numpy(np.concatenate(euclid_distance, axis=0)).type(torch.float)
return euclid_distance
def distance_measure(ref, target):
ref = ref.unsqueeze(0)
target = target.unsqueeze(0)
euc_dist = torch.pairwise_distance(ref, target)
cos_dist = 1 - torch.nn.functional.cosine_similarity(ref, target)
query_dist = euc_dist.item() / 200.0 + cos_dist.item(
) # the coefficients are taken from the original implementation.
return query_dist
def fake_speed(fake_traj):
output_speed = []
for a, b in zip(fake_traj[:, :], fake_traj[1:, :]):
dist = torch.pairwise_distance(a, b)
speed = dist / 0.4
output_speed.append(speed.view(1, -1))
output_fake_speed = torch.cat(output_speed,
dim=0).unsqueeze(dim=2).permute(1, 0, 2)
return output_fake_speed
def evaluate_model_pairwise(model,
loss,
g1,
g2,
prec_k=(1, 3, 5, 10, 30),
compute_accuracy=False):
model.eval()
metrics = {}
metrics['pair'] = g1.graph_name + '-->' + g2.graph_name
pos_anchor_idx = g1.anchor_data[g2.gidx]['anchor_edge_index']
neg_anchor_idx = g1.anchor_data[g2.gidx]['negative_anchor_edge_index']
print(">>> Evaluating model...")
x1, x2 = model(g1,
g2,
edge_index_labels=('anchor_edge_index',
'negative_anchor_edge_index'))
if loss == 'contrastive':
euc_dist = torch.pairwise_distance(x1, x2, keepdim=True)
model_pred = 1. / (1. + euc_dist)
elif loss == 'cosine':
cos = torch.nn.CosineSimilarity(dim=1, eps=1e-6)
model_pred = 1. - 0.5 * (1. - cos(x1, x2))
elif loss == 'BCE':
model_pred = torch.sigmoid(torch.einsum("ef,ef->e", x1, x2))
else:
raise NotImplementedError("Loss function not implemented")
model_pred = model_pred.detach().numpy()
anchor_idx = torch.cat([pos_anchor_idx, neg_anchor_idx],
dim=-1).detach().numpy()
true_anchor_idx = list(map(tuple, pos_anchor_idx.numpy().T))
####### Computing Accuracy #######
if compute_accuracy:
print(">>> Computing accuracy...")
pred_anchor_idx = greedy_matching(anchor_idx, model_pred)
metrics['accuracy'] = accuracy(true_anchor_idx, pred_anchor_idx)
####### Computing Precision #######
print(">>> Computing precision metrics...")
ordered_sim = soft_ordering(anchor_idx, model_pred)
for k in prec_k:
metrics['precision@' + str(k)] = precision(true_anchor_idx,
ordered_sim, k)
metrics['MAP'] = MAP(true_anchor_idx, ordered_sim)
return metrics
def batched_eucl(coord):
"""Compute adjacency matrix for a batched coordinates Tensor `c`.
https://github.com/pytorch/pytorch/issues/9406#issuecomment-472269698
"""
B, M, N = coord.shape
return t.pairwise_distance(
coord[:, :, None].expand(B, M, M, N).reshape((-1, N)),
coord[:, None].expand(B, M, M, N).reshape((-1, N)),
).reshape((B, M, M))
def print_closest(word_map, word, n=10):
"""
Sanity check for the pre-trained embedding:
>> WORKING
"""
word_idx = {i: v for v, i in word_map.items()}
idx = word_map[word]
vector = embedding[idx]
distances = torch.pairwise_distance(embedding, vector)
_, idx = distances.sort(descending=False)
for k in range(n):
print(word_idx[idx[k].item()])
def update(self, homography, src_pts, dst_pts, src_confs, dst_confs,
src_img_size, dst_img_size):
"""
Updates the internal evaluation result.
Parameters
----------
homography : torch.Tensor
Homography (from source image to destination one).
src_pts : torch.Tensor
Detected points for the first (source) image.
dst_pts : torch.Tensor
Detected points for the second (destination) image.
src_confs : torch.Tensor
Confidences for detected points on the source image.
dst_confs : torch.Tensor
Confidences for detected points on the destination image.
src_img_size : tuple of 2 int
Size (H, W) of the source image.
dst_img_size : tuple of 2 int
Size (H, W) of the destination image.
"""
# from scipy.spatial.distance import pdist
from scipy.optimize import linear_sum_assignment
with torch.no_grad():
src_hmg_pts = self.calc_homogeneous_coords(src_pts.float())
dst_hmg_pts = self.calc_homogeneous_coords(dst_pts.float())
self.filter_inside_points(src_hmg_pts, src_confs, homography,
dst_img_size)
self.filter_inside_points(dst_hmg_pts, dst_confs,
homography.inverse(), src_img_size)
src_pts_count = src_hmg_pts.shape[0]
dst_pts_count = dst_hmg_pts.shape[0]
pts_count = min(src_pts_count, dst_pts_count, 100)
assert (pts_count > 0)
src_hmg_pts = self.filter_best_points(src_hmg_pts, src_confs,
pts_count)
dst_hmg_pts = self.filter_best_points(dst_hmg_pts, dst_confs,
pts_count)
preds_dst_hmg_pts = self.transform_points(src_hmg_pts, homography)
cost = torch.pairwise_distance(
x1=preds_dst_hmg_pts, x2=dst_hmg_pts).cpu().detach().numpy()
row_ind, col_ind = linear_sum_assignment(cost)
mean_resudual = cost[row_ind, col_ind].sum()
mean_resudual *= (100.0 / dst_img_size[0])
self.sum_metric += mean_resudual
self.global_sum_metric += mean_resudual
self.num_inst += 1
self.global_num_inst += 1
def _test(y_pred, y, batch_size):
mpd.reset()
if batch_size > 1:
n_iters = y.shape[0] // batch_size + 1
for i in range(n_iters):
idx = i * batch_size
mpd.update((y_pred[idx : idx + batch_size], y[idx : idx + batch_size]))
else:
mpd.update((y_pred, y))
np_res = np.mean(torch.pairwise_distance(y_pred, y, p=mpd._p, eps=mpd._eps).numpy())
assert isinstance(mpd.compute(), float)
assert pytest.approx(mpd.compute()) == np_res
def _test_distrib_itegration(device):
import numpy as np
import torch.distributed as dist
from ignite.engine import Engine
rank = dist.get_rank()
torch.manual_seed(12)
n_iters = 100
s = 50
offset = n_iters * s
y_true = torch.rand(offset * dist.get_world_size(), 10).to(device)
y_preds = torch.rand(offset * dist.get_world_size(), 10).to(device)
def update(engine, i):
return (
y_preds[i * s + offset * rank : (i + 1) * s + offset * rank, ...],
y_true[i * s + offset * rank : (i + 1) * s + offset * rank, ...],
)
engine = Engine(update)
m = MeanPairwiseDistance(device=device)
m.attach(engine, "mpwd")
data = list(range(n_iters))
engine.run(data=data, max_epochs=1)
assert "mpwd" in engine.state.metrics
res = engine.state.metrics["mpwd"]
true_res = []
for i in range(n_iters * dist.get_world_size()):
true_res.append(
torch.pairwise_distance(
y_true[i * s : (i + 1) * s, ...],
y_preds[i * s : (i + 1) * s, ...],
p=m._p,
eps=m._eps,
)
.cpu()
.numpy()
)
true_res = np.array(true_res).ravel()
true_res = true_res.mean()
assert pytest.approx(res) == true_res
def forward(self, x1, x2, y):
'''
Shapes:
-------
x1: [B,C]
x2: [B,C]
y: [B,1]
Returns:
--------
loss: [B,1]]
'''
distance = torch.pairwise_distance(x1, x2, p=2)
loss = self.alpha * (1-y) * distance**2 + \
self.beta * y * (torch.max(torch.zeros_like(distance), self.margin - distance)**2)
return torch.mean(loss, dtype=torch.float)
def contrastive_loss(self, x1, x2, y):
margin = 1.
d = torch.pairwise_distance(x1, x2, keepdim=False)
d = d.tolist()
y = y.tolist()
loss = []
for i, item in enumerate(d):
loss.append(0.5 * y[i] * item * item + 0.5 * (1 - y[i]) *
max(0, (margin - item)) * max(0, (margin - item)))
loss = np.array(loss)
loss_mean = torch.Tensor([loss.mean()])
return Variable(loss_mean, requires_grad=True)
def pdist2(x, y):
"""
Compute distance between each pair of row vectors in x and y
Args:
x: tensor of shape n*p
y: tensor of shape m*p
Returns:
dist: tensor of shape n*m
"""
p = x.shape[1]
n = x.shape[0]
m = y.shape[0]
xtile = torch.cat([x] * m, dim=1).view(-1, p)
ytile = torch.cat([y] * n, dim=0)
dist = torch.pairwise_distance(xtile, ytile)
return dist.view(n, m)
def get_query_x(self, Query_x, centroid_per_class, Query_y_labels):
"""
Returns distance matrix from each Query image to each centroid.
"""
centroid_matrix = self.get_centroid_matrix(centroid_per_class,
Query_y_labels)
Query_x = self.f(Query_x)
m = Query_x.size(0)
n = centroid_matrix.size(0)
centroid_matrix = centroid_matrix.expand(m, centroid_matrix.size(0),
centroid_matrix.size(1))
Query_matrix = Query_x.expand(
n, Query_x.size(0),
Query_x.size(1)).transpose(0,
1) # Expanding Query matrix "n" times
Qx = torch.pairwise_distance(centroid_matrix.transpose(1, 2),
Query_matrix.transpose(1, 2))
return Qx
def forward(
self, sequence_a: TextFieldTensors, sequence_b: TextFieldTensors, distance: Optional[torch.Tensor] = None,
) -> Dict[str, torch.Tensor]:
embedded_sequence_a = self.encode_sequence(sequence_a)
embedded_sequence_b = self.encode_sequence(sequence_b)
euclidian_distance = torch.pairwise_distance(embedded_sequence_a, embedded_sequence_b)
output_dict = {
"edit_distance": euclidian_distance,
"emb_sequence_a": embedded_sequence_a,
"emb_sequence_b": embedded_sequence_b,
}
if distance is not None:
loss = self._loss(euclidian_distance, distance.view(-1))
output_dict["loss"] = loss
return output_dict
def get_query_x(self, Query_x, centroid_per_class, Query_y_labels):
"""
Returns distance matrix from each Query image to each centroid.
"""
centroid_matrix = self.get_centroid_matrix(
centroid_per_class, Query_y_labels)
Query_x = self.f(Query_x)
m = Query_x.size(0)
n = centroid_matrix.size(0)
# The below expressions expand both the matrices such that they become compatible to each other in order to caclulate L2 distance.
# Expanding centroid matrix to "m".
centroid_matrix = centroid_matrix.expand(
m, centroid_matrix.size(0), centroid_matrix.size(1))
Query_matrix = Query_x.expand(n, Query_x.size(0), Query_x.size(
1)).transpose(0, 1) # Expanding Query matrix "n" times
Qx = torch.pairwise_distance(centroid_matrix.transpose(
1, 2), Query_matrix.transpose(1, 2))
return Qx
def get_adjacency_eu(self, x, y): # x,y:[1, 2704, 256]
# A = torch.matmul(x, y.transpose(1, 2))
A = torch.rand(x.shape[1], y.shape[1]) # [2704, 2704]
for i in range(x.shape[1]):
# A[i][j] = torch.sqrt(torch.sum((x[0][i] - y[0][j])**2))
A[i] = torch.pairwise_distance(x[0][i].unsqueeze(0).cpu(),
y[0].cpu(),
p=2)
temp = torch.ones_like(A) * 100
A = temp - A
_, idx = torch.topk(A, k=10, dim=-1, largest=True)
for i in range(idx.shape[0]): # [2704, 10]
for index in idx[i]:
A[i][index] = 1
zero = torch.zeros_like(A)
A = torch.where(A == 1, A, zero).unsqueeze(0)
return A
