def _get_single_double_config(self, nocc, nvirt):
"""Get the confs of the single + double
Args:
nelec (int): number of electrons in the active space
norb (int) : number of orbitals in the active space
"""
_gs_up = list(range(self.nup))
_gs_down = list(range(self.ndown))
cup, cdown = self._get_single_config(nocc, nvirt)
cup = cup.tolist()
cdown = cdown.tolist()
idx_occ_up = list(
range(self.nup - 1, self.nup - 1 - nocc, -1))
idx_vrt_up = list(range(self.nup, self.nup + nvirt, 1))
idx_occ_down = list(range(
self.ndown - 1, self.ndown - 1 - nocc, -1))
idx_vrt_down = list(range(self.ndown, self.ndown + nvirt, 1))
# ground, single and double with 1 elec excited per spin
for iocc_up in idx_occ_up:
for ivirt_up in idx_vrt_up:
for iocc_down in idx_occ_down:
for ivirt_down in idx_vrt_down:
_xt_up = self._create_excitation(
_gs_up.copy(), iocc_up, ivirt_up)
_xt_down = self._create_excitation(
_gs_down.copy(), iocc_down, ivirt_down)
cup, cdown = self._append_excitations(
cup, cdown, _xt_up, _xt_down)
# double with 2elec excited on spin up
for occ1, occ2 in torch.combinations(torch.as_tensor(idx_occ_up), r=2):
for vrt1, vrt2 in torch.combinations(torch.as_tensor(idx_vrt_up), r=2):
_xt_up = self._create_excitation(
_gs_up.copy(), occ1, vrt2)
_xt_up = self._create_excitation(_xt_up, occ2, vrt1)
cup, cdown = self._append_excitations(
cup, cdown, _xt_up, _gs_down)
# double with 2elec excited per spin
for occ1, occ2 in torch.combinations(torch.as_tensor(idx_occ_down), r=2):
for vrt1, vrt2 in torch.combinations(torch.as_tensor(idx_vrt_down), r=2):
_xt_down = self._create_excitation(
_gs_down.copy(), occ1, vrt2)
_xt_down = self._create_excitation(
_xt_down, occ2, vrt1)
cup, cdown = self._append_excitations(
cup, cdown, _gs_up, _xt_down)
return (torch.LongTensor(cup), torch.LongTensor(cdown))
def similarity_computing_inner(self, doc_output, segment_idx):
similarities = []
cos = nn.CosineSimilarity(dim=1, eps=1e-6)
seg_outputs = []
index = 0
doc_output = F.softmax(doc_output)
for i, idx in enumerate(segment_idx):
if i == 0:
seg_output = doc_output[0:segment_idx[i] + 1, :]
elif i == len(segment_idx) - 1:
seg_output = doc_output[segment_idx[i - 1] + 1:, :]
else:
seg_output = doc_output[segment_idx[i - 1] + 1:segment_idx[i] +
1, :]
seg_outputs.append(seg_output)
for i in range(len(seg_outputs)):
sent_idx = maybe_cuda(
torch.LongTensor([k for k in range(seg_outputs[i].size()[0])]))
if seg_outputs[i].size()[0] > 1:
pairs = torch.combinations(sent_idx)
pair_sims = []
for p in pairs:
pair_sims.append(
cos(seg_outputs[i][p[0], :].unsqueeze(0),
seg_outputs[i][p[1], :].unsqueeze(0)))
similarities.append(sum(pair_sims) / len(pair_sims))
else:
continue
return Variable(maybe_cuda(torch.Tensor(similarities)))
def __init__(self, hparams):
super(HeteroConv, self).__init__()
self.hparams = hparams
concatenation_factor = (
3 if (self.hparams["aggregation"] in ["sum_max", "mean_max", "mean_sum"]) else 2
)
# Make module list
self.node_encoders = nn.ModuleList([
make_mlp(
concatenation_factor * hparams["hidden"],
[hparams["hidden"]] * hparams["nb_node_layer"],
output_activation=hparams["output_activation"],
hidden_activation=hparams["hidden_activation"],
layer_norm=hparams["layernorm"],
batch_norm=hparams["batchnorm"],
) for _ in hparams["model_ids"]
])
# Make edge encoder combos (this is an N-choose-2 with replacement situation)
self.all_combos = torch.combinations(torch.arange(len(self.hparams["model_ids"])), r=2, with_replacement=True)
self.edge_encoders = nn.ModuleList([
make_mlp(
3 * hparams["hidden"],
[hparams["hidden"]] * hparams["nb_edge_layer"],
layer_norm=hparams["layernorm"],
batch_norm=hparams["batchnorm"],
output_activation=hparams["output_activation"],
hidden_activation=hparams["hidden_activation"],
) for _ in self.all_combos
])
def multiview_covariance_matrix(dims,
constructor,
options=dict(),
symmetric=True,
**kwargs):
num_views = dims.size(0)
indices = torch.arange(num_views, dtype=torch.long)
diag_indices = indices.repeat(2, 1).t_()
upper_indices = torch.combinations(indices, r=2)
diag_and_upper_indices = torch.cat([diag_indices, upper_indices])
sum_dimensions = dims.sum()
options.update(kwargs)
out = torch.empty(sum_dimensions, sum_dimensions, **options)
for j, r in diag_and_upper_indices:
j_indices = slice(dims[:j].sum(), dims[:j + 1].sum())
r_indices = slice(dims[:r].sum(), dims[:r + 1].sum())
jr_cross_covariance = constructor(j, r)
if not torch.is_tensor(
jr_cross_covariance) or jr_cross_covariance.nelement() == 1:
jr_cross_covariance = torch.empty(
dims[j], dims[r], **options).fill_(jr_cross_covariance)
elif jr_cross_covariance.nelement() == 0:
jr_cross_covariance = torch.zeros(dims[j], dims[r], **options)
out[j_indices, r_indices] = jr_cross_covariance
if j != r:
out[r_indices, j_indices] = jr_cross_covariance.t(
) if symmetric else constructor(r, j)
return out
def __init__(self, hparams):
super(HeteroEncoder, self).__init__()
self.hparams = hparams
# Make module list
self.node_encoders = nn.ModuleList([
make_mlp(
model["num_features"],
[hparams["hidden"]] * hparams["nb_node_layer"],
output_activation=hparams["output_activation"],
hidden_activation=hparams["hidden_activation"],
layer_norm=hparams["layernorm"],
batch_norm=hparams["batchnorm"],
) for model in hparams["model_ids"]
])
# Make edge encoder combos (this is an N-choose-2 with replacement situation)
self.all_combos = torch.combinations(torch.arange(
len(self.hparams["model_ids"])),
r=2,
with_replacement=True)
self.edge_encoders = nn.ModuleList([
make_mlp(
2 * hparams["hidden"],
[hparams["hidden"]] * hparams["nb_edge_layer"],
layer_norm=hparams["layernorm"],
batch_norm=hparams["batchnorm"],
output_activation=hparams["output_activation"],
hidden_activation=hparams["hidden_activation"],
) for _ in self.all_combos
])
def get_triplets(self, embeddings, labels):
# calculate distances between embedded vectors
distance_matrix = pdist(embeddings)
# logging.debug(distance_matrix)
# triplets..
triplets = []
# get unique labels in this batch
unique_labels = labels.unique()
for anchor_label in unique_labels:
# anchor indices
anchor_mask = labels == anchor_label
anchor_indices = (anchor_mask == 1).nonzero()[:, 0]
# skip the following procedure if we have one or zero positive case
if anchor_indices.size(0) < 2:
continue
# negative indices
negative_indices = (anchor_mask == 0).nonzero()[:, 0]
# generate anchor-positive pairs using combination
anchor_positive_pairs = torch.combinations(anchor_indices)
# for each pair
for ap_pair in anchor_positive_pairs:
anchor_idx = ap_pair[0]
positive_idx = ap_pair[1]
# distance between anchor-positive
ap_distance = distance_matrix[anchor_idx, positive_idx]
# distanceS between anchor-negativeS
an_distances = distance_matrix[anchor_idx][negative_indices]
# loss_valueS = ap - anS + margin
loss_values = ap_distance - an_distances + self.margin
# for each fn in func_list (ordered by high priority)
negative_idx = None
for fn in self.func_list:
negative_idx = fn(loss_values, self.margin)
if negative_idx is not None:
break
# if no negative_idx is found, just pick a random one
if negative_idx is None:
logging.debug("No negative idx found. Picking random..")
choice = torch.randint(0, negative_indices.size(0), (1,))
negative_idx = negative_indices[choice].view(-1).squeeze(0)
triplet = torch.stack([anchor_idx, positive_idx, negative_idx], 0)
triplets.append(triplet)
triplet_tensor = torch.stack(triplets)
logging.debug(triplet_tensor.size())
return triplet_tensor
def _create_pixel_pairs(self, input):
batch_size = input.shape[0]
# Divide image into nxn blocks
n = 16
unfold = torch.nn.Unfold(kernel_size=(n, n), stride=n)
input_blocks = unfold(input) # [B, 256, 196]
num_patches = input_blocks.shape[-1]
flattened_blocks = input_blocks.view(-1)
total_patches = batch_size * input_blocks.shape[-1]
# Randomly sample 1 pixel from each block
random_pixel_idxs = torch.Tensor([
256 * i + torch.randint(0, 256, size=(1, ))
for i in range(total_patches)
])
pixel_samples = flattened_blocks[random_pixel_idxs.type(torch.long)]
pixel_samples_batched = pixel_samples.view(batch_size, num_patches)
# Create combinations of each pixel pair index
pixel_pairs = torch.stack(([
torch.combinations(pixel_samples_batched[i])
for i in range(batch_size)
]))
pixel_pairs = pixel_pairs.view(-1, 2)
return pixel_pairs[..., 0], pixel_pairs[..., 1]
def erdos_renyi_graph(num_nodes, edge_prob, directed=False):
r"""Returns the :obj:`edge_index` of a random Erdos-Renyi graph.
Args:
num_nodes (int): The number of nodes.
edge_prob (float): Probability of an edge.
directed (bool, optional): If set to :obj:`True`, will return a
directed graph. (default: :obj:`False`)
"""
if directed:
idx = torch.arange((num_nodes - 1) * num_nodes)
idx = idx.view(num_nodes - 1, num_nodes)
idx = idx + torch.arange(1, num_nodes).view(-1, 1)
idx = idx.view(-1)
else:
idx = torch.combinations(torch.arange(num_nodes))
# Filter edges.
mask = torch.rand(idx.size(0)) < edge_prob
idx = idx[mask]
if directed:
row = idx // num_nodes
col = idx % num_nodes
edge_index = torch.stack([row, col], dim=0)
else:
edge_index = to_undirected(idx.t(), num_nodes)
return edge_index
def __init__(self, hparams):
super().__init__(hparams)
"""
Initialise the Lightning Module that can scan over different filter training regimes
"""
self.all_combos = torch.combinations(torch.arange(
len(self.hparams["model_ids"])),
r=2,
with_replacement=True)
# Still need this??
self.vol_matrix = get_vol_matrix(
self.all_combos,
[model_id["volume_ids"] for model_id in self.hparams["model_ids"]])
self.edge_encoders = nn.ModuleList([
make_mlp(
hparams["model_ids"][combo[0]]["num_features"] +
hparams["model_ids"][combo[1]]["num_features"] +
2 * hparams["cell_channels"],
[
hparams["hidden"] // (2**i)
for i in range(hparams["nb_layer"])
] + [1],
layer_norm=hparams["layernorm"],
batch_norm=hparams["batchnorm"],
output_activation=None,
hidden_activation=hparams["hidden_activation"],
) for combo in self.all_combos
])
def IP_loss(phase, mask):
osci_num = phase.shape[1]
phase_sin = torch.sin(phase)
phase_cos = torch.cos(phase)
masked_sin = phase_sin.unsqueeze(1) * mask
masked_cos = phase_cos.unsqueeze(1) * mask
# mask.shape=(batch, groups, N)
product1 = torch.matmul(masked_sin.unsqueeze(3), masked_sin.unsqueeze(2)) +\
torch.matmul(masked_cos.unsqueeze(3), masked_cos.unsqueeze(2))
sync_loss = (product1.sum(3).sum(2) - osci_num) / (osci_num**2 - osci_num)
product2 = torch.matmul(masked_sin.unsqueeze(2).unsqueeze(4), masked_sin.unsqueeze(1).unsqueeze(3)) +\
torch.matmul(masked_cos.unsqueeze(2).unsqueeze(4), masked_cos.unsqueeze(1).unsqueeze(3))
product2 = product2.sum(2, keepdim=True) - product1.unsqueeze(2)
desync_loss = product2.squeeze().sum(3).sum(2).sum(
1) / 2 / torch.combinations(torch.arange(osci_num)).shape[0]
sync_loss_mean = torch.exp(-sync_loss.mean())
desync_loss_mean = torch.exp(desync_loss.mean())
tot_loss_mean = 0.1 * sync_loss_mean + desync_loss_mean
return tot_loss_mean, sync_loss_mean, desync_loss_mean
def _combinations(self, tensor, dim=0):
n = tensor.shape[dim]
if n == 0:
return tensor, tensor
r = torch.arange(n, dtype=torch.long, device=tensor.device)
index1, index2 = torch.combinations(r).unbind(-1)
return tensor.index_select(dim, index1), \
tensor.index_select(dim, index2)
def triu_index(num_species):
species = torch.arange(num_species)
species1, species2 = torch.combinations(species, r=2, with_replacement=True).unbind(-1)
pair_index = torch.arange(species1.shape[0])
ret = torch.zeros(num_species, num_species, dtype=torch.long)
ret[species1, species2] = pair_index
ret[species2, species1] = pair_index
return ret
def __init__(self,
points_per_patch=256,
patch_radius=[0.05],
dim_pts=3,
num_gpts=128,
dim_gpts=1,
hyps=64,
inlier_params=[0.01, 0.5],
use_mask=False,
sym_op='max',
ith=0,
use_point_stn=True,
use_feat_stn=True,
decoder='PointPredNet',
device=0,
normal_loss='ms_euclidean',
seed=3627474):
'''
Constructor.
hyps -- number of planes hypotheses sampled for each patch
inlier_thresh -- threshold used in the soft inlier count,
inlier_beta -- scaling factor within the sigmoid of the soft inlier count
inlier_alpha -- scaling factor for the soft inlier scores (controls the peakiness of the hypothesis distribution)
'''
super(WDSAC, self).__init__()
self.hyps = hyps
self.num_gpts = num_gpts
if len(inlier_params) == 2:
self.scorer = Scorer_dist(inlier_params[0])
else:
self.scorer = Scorer_inlier_count(inlier_params[0],
inlier_params[1])
self.inlier_alpha = inlier_params[-1]
self.normal_loss = normal_loss
self.use_point_stn = use_point_stn
self.device = device
torch.manual_seed(seed)
gpts_idx = torch.tensor([i for i in range(num_gpts)])
self.idx_combi = torch.combinations(gpts_idx, 3).to(device)
#self.plane_weight = torch.ones(self.idx_combi.size(0), dtype=torch.float).to(device)
self.wpcp = PCPNet(num_pts=points_per_patch,
dim_pts=dim_pts,
num_gpts=points_per_patch,
dim_gpts=1,
use_point_stn=use_point_stn,
use_feat_stn=use_feat_stn,
device=device,
b_pred=True,
use_mask=False,
sym_op=sym_op,
ith=0)
def main(args):
res = []
combos = torch.combinations(torch.arange(0, args.n_modules), args.n_active).numpy().tolist()
for i, subset in enumerate(combos):
if args.head_id in subset:
res.append(str(i))
print("{}".format(" ".join(res)))
def take_all_pwrs(self, vec, pwr): #todo: vectorize (kinda) combins=torch.combinations(vec, r=pwr, with_replacement=True) out=torch.ones(combins.size()[0]).to(device).to(torch.float) for i in torch.t(combins).to(device).to(torch.float): out *= i if pwr == 1: return out else: return torch.cat((out,self.take_all_pwrs(vec, pwr-1)))
def neighbor_pairs(padding_mask, coordinates, cell, shifts, cutoff):
"""Compute pairs of atoms that are neighbors
Copyright 2018- Xiang Gao and other ANI developers
(https://github.com/aiqm/torchani/blob/master/torchani/aev.py)
Arguments:
padding_mask (:class:`torch.Tensor`): boolean tensor of shape
(molecules, atoms) for padding mask. 1 == is padding.
coordinates (:class:`torch.Tensor`): tensor of shape
(molecules, atoms, 3) for atom coordinates.
cell (:class:`torch.Tensor`): tensor of shape (3, 3) of the three vectors
defining unit cell: tensor([[x1, y1, z1], [x2, y2, z2], [x3, y3, z3]])
cutoff (float): the cutoff inside which atoms are considered pairs
shifts (:class:`torch.Tensor`): tensor of shape (?, 3) storing shifts
"""
# type: (Tensor, Tensor, Tensor, Tensor, float) -> Tuple[Tensor, Tensor, Tensor, Tensor]
coordinates = coordinates.detach()
cell = cell.detach()
num_atoms = padding_mask.shape[0]
all_atoms = torch.arange(num_atoms, device=cell.device)
# Step 2: center cell
p1_center, p2_center = torch.combinations(all_atoms).unbind(-1)
shifts_center = shifts.new_zeros(p1_center.shape[0], 3)
# Step 3: cells with shifts
# shape convention (shift index, molecule index, atom index, 3)
num_shifts = shifts.shape[0]
all_shifts = torch.arange(num_shifts, device=cell.device)
shift_index, p1, p2 = torch.cartesian_prod(all_shifts, all_atoms, all_atoms).unbind(
-1
)
shifts_outside = shifts.index_select(0, shift_index)
# Step 4: combine results for all cells
shifts_all = torch.cat([shifts_center, shifts_outside])
p1_all = torch.cat([p1_center, p1])
p2_all = torch.cat([p2_center, p2])
shift_values = torch.mm(shifts_all.to(cell.dtype), cell)
# step 5, compute distances, and find all pairs within cutoff
distances = (coordinates[p1_all] - coordinates[p2_all] + shift_values).norm(2, -1)
padding_mask = (padding_mask[p1_all]) | (padding_mask[p2_all])
distances.masked_fill_(padding_mask, math.inf)
in_cutoff = torch.nonzero(distances < cutoff, as_tuple=False)
pair_index = in_cutoff.squeeze()
atom_index1 = p1_all[pair_index]
atom_index2 = p2_all[pair_index]
shifts = shifts_all.index_select(0, pair_index)
return atom_index1, atom_index2, shifts
def get_iterator(N, get_all=True):
if get_all:
# every possible pair
total = (N * (N - 1) // 2)
iters = torch.combinations(torch.from_numpy(np.arange(N)),
2).to(DEVICE).long()
else:
total = 10000
iters = torch.from_numpy(np.random.randint(
N, size=(total, 2))).to(DEVICE).long()
return total, iters
def test_sample_hyp(tries_in=9999):
gpu_idx = -3
device = torch.device("cpu" if gpu_idx < 0 else "cuda:%d" % gpu_idx)
batchsize = 2
rng = np.random.RandomState(3627474)
num_gpts = 256
pts = torch.rand(batchsize, num_gpts, 3)
hyps = 32
gpts_idx = torch.tensor([i for i in range(num_gpts)])
combi = torch.combinations(gpts_idx, 3).to(device)
t = time.process_time()
tries = tries_in
while tries:
tmp = torch.randint(0, combi.size(0), (hyps * batchsize, ))
index = combi[tmp, :].view(batchsize, hyps * 3, 1)
if tries:
tries -= 1
else:
break
print(time.process_time() - t)
plane_weight = torch.ones(combi.size(0), dtype=torch.float).to(device)
t = time.process_time()
tries = tries_in
while tries:
tmp = torch.multinomial(plane_weight,
hyps * batchsize,
replacement=False)
index = combi[tmp, :].view(batchsize, hyps * 3, 1)
if tries:
tries -= 1
else:
break
print(time.process_time() - t)
t = time.process_time()
tries = tries_in
while tries:
index = torch.stack([
torch.from_numpy(
np.stack(
rng.choice(pts.size(1), 3, replace=False)
for _ in range(hyps)).reshape(-1))
for _ in range(batchsize)
])
index = index.view(batchsize, hyps * 3, 1)
if tries:
tries -= 1
else:
break
print(time.process_time() - t)
def neighbor_pairs(padding_mask: Tensor, coordinates: Tensor, cell: Tensor,
shifts: Tensor,
cutoff: float) -> Tuple[Tensor, Tensor, Tensor]:
"""Compute pairs of atoms that are neighbors
Arguments:
padding_mask (:class:`torch.Tensor`): boolean tensor of shape
(molecules, atoms) for padding mask. 1 == is padding.
coordinates (:class:`torch.Tensor`): tensor of shape
(molecules, atoms, 3) for atom coordinates.
cell (:class:`torch.Tensor`): tensor of shape (3, 3) of the three vectors
defining unit cell: tensor([[x1, y1, z1], [x2, y2, z2], [x3, y3, z3]])
cutoff (float): the cutoff inside which atoms are considered pairs
shifts (:class:`torch.Tensor`): tensor of shape (?, 3) storing shifts
"""
coordinates = coordinates.detach()
cell = cell.detach()
num_atoms = padding_mask.shape[1]
all_atoms = torch.arange(num_atoms, device=cell.device)
# Step 2: center cell
p1_center, p2_center = torch.combinations(all_atoms).unbind(-1)
shifts_center = shifts.new_zeros((p1_center.shape[0], 3))
# Step 3: cells with shifts
# shape convention (shift index, molecule index, atom index, 3)
num_shifts = shifts.shape[0]
all_shifts = torch.arange(num_shifts, device=cell.device)
shift_index, p1, p2 = torch.cartesian_prod(all_shifts, all_atoms,
all_atoms).unbind(-1)
shifts_outide = shifts.index_select(0, shift_index)
# Step 4: combine results for all cells
shifts_all = torch.cat([shifts_center, shifts_outide])
p1_all = torch.cat([p1_center, p1])
p2_all = torch.cat([p2_center, p2])
shift_values = shifts_all.to(cell.dtype) @ cell
# step 5, compute distances, and find all pairs within cutoff
distances = (coordinates.index_select(1, p1_all) -
coordinates.index_select(1, p2_all) + shift_values).norm(
2, -1)
padding_mask = (padding_mask.index_select(
1, p1_all)) | (padding_mask.index_select(1, p2_all))
distances.masked_fill_(padding_mask, math.inf)
in_cutoff = (distances <= cutoff).nonzero()
molecule_index, pair_index = in_cutoff.unbind(1)
molecule_index *= num_atoms
atom_index1 = p1_all[pair_index]
atom_index2 = p2_all[pair_index]
shifts = shifts_all.index_select(0, pair_index)
return molecule_index + atom_index1, molecule_index + atom_index2, shifts
def forward(self, vertex_pred, vertex_gt, c_pred, points, target, model_points, model_kp, idx, target_r, target_t):
vertex_loss = smooth_l1_loss(vertex_pred.view(1, self.num_pt_mesh, -1), vertex_gt.view(1, self.num_pt_mesh, -1))
kp_set = vertex_pred + points.repeat(1, 1, 9).view(1, points.shape[1], 9, 3)
confidence = c_pred / (0.00001 + torch.sum(c_pred, 1))
points_pred = torch.sum(confidence * kp_set, 1)
all_index = torch.combinations(torch.arange(9), 3)
all_r, all_t = batch_least_square(model_kp.squeeze()[all_index, :], points_pred.squeeze()[all_index, :], torch.ones([all_index.shape[0], 3]).cuda())
all_e = calculate_error(all_r, all_t, model_points, points)
e = all_e.unsqueeze(0).unsqueeze(2)
w = torch.softmax(1 / e, 1).squeeze().unsqueeze(1)
all_qua = tgm.rotation_matrix_to_quaternion(torch.cat((all_r, torch.tensor([0., 0., 1.]).cuda().unsqueeze(1).repeat(all_index.shape[0], 1, 1)), dim=2))
pred_qua = torch.sum(w * all_qua, 0)
pred_r = pred_qua.view(1, 1, -1)
bs, num_p, _ = pred_r.size()
pred_r = pred_r / (torch.norm(pred_r, dim=2).view(1, 1, 1))
pred_r = torch.cat(((1.0 - 2.0 * (pred_r[:, :, 2] ** 2 + pred_r[:, :, 3] ** 2)).view(bs, num_p, 1), \
(2.0 * pred_r[:, :, 1] * pred_r[:, :, 2] - 2.0 * pred_r[:, :, 0] * pred_r[:, :, 3]).view(bs, num_p,1), \
(2.0 * pred_r[:, :, 0] * pred_r[:, :, 2] + 2.0 * pred_r[:, :, 1] * pred_r[:, :, 3]).view(bs, num_p, 1), \
(2.0 * pred_r[:, :, 1] * pred_r[:, :, 2] + 2.0 * pred_r[:, :, 3] * pred_r[:, :, 0]).view(bs, num_p, 1), \
(1.0 - 2.0 * (pred_r[:, :, 1] ** 2 + pred_r[:, :, 3] ** 2)).view(bs, num_p, 1), \
(-2.0 * pred_r[:, :, 0] * pred_r[:, :, 1] + 2.0 * pred_r[:, :, 2] * pred_r[:, :, 3]).view(bs, num_p, 1), \
(-2.0 * pred_r[:, :, 0] * pred_r[:, :, 2] + 2.0 * pred_r[:, :, 1] * pred_r[:, :, 3]).view(bs, num_p, 1), \
(2.0 * pred_r[:, :, 0] * pred_r[:, :, 1] + 2.0 * pred_r[:, :, 2] * pred_r[:, :, 3]).view(bs, num_p, 1), \
(1.0 - 2.0 * (pred_r[:, :, 1] ** 2 + pred_r[:, :, 2] ** 2)).view(bs, num_p, 1)), dim=2).contiguous().view(bs * num_p, 3, 3)
pred_r = pred_r.squeeze()
pred_t = torch.sum(w * all_t, 0)
target_r = target_r.squeeze()
target_t = target_t.squeeze()
pose_loss = torch.norm(pred_t - target_t) + 0.01 * torch.norm(torch.mm(pred_r, torch.transpose(target_r, 1, 0)) - torch.eye(3).cuda())
pred = torch.mm(model_points[0], torch.transpose(pred_r, 1, 0)) + pred_t
knn = KNN(k=1, transpose_mode=True)
if idx in self.sym_list:
dist, inds = knn(pred.unsqueeze(0), target.unsqueeze(0))
dis = torch.mean(dist.squeeze())
else:
dis = torch.mean(torch.norm(pred - target[0], dim=1), dim=0)
ori_r = torch.unsqueeze(pred_r, 0).cuda()
ori_t = torch.unsqueeze(pred_t, 0).cuda()
ori_t = ori_t.repeat(self.num_pt_mesh, 1).contiguous().view(1, self.num_pt_mesh, 3)
new_points = torch.bmm((points - ori_t), ori_r).contiguous()
new_target = torch.bmm((target - ori_t), ori_r).contiguous()
del knn
return vertex_loss, pose_loss, dis, new_points.detach(), new_target.detach()
def fast_pcmvda_loss(Ys, y, y_unique=None, beta=1, q=1):
if len(set(X.size(1) for X in Ys)) > 1:
raise ValueError(
f"pc-MvDA only works on projected data with same dimensions, "
f"got dimensions of {tuple(X.size(1) for X in Ys)}.")
y_unique_present = torch.unique(y)
if y_unique is None:
y_unique = y_unique_present
y_present_mask = torch.zeros_like(y_unique, dtype=torch.bool)
y_present_mask[y_unique_present.tolist()] = 1
else:
y_present_mask = torch.ones_like(y_unique, dtype=torch.bool)
y_present_mask = torch.where(y_present_mask)[0]
options = dict(dtype=Ys[0].dtype, device=Ys[0].device)
num_views = len(Ys)
num_components = y.size(0)
num_classes = y_unique.size(0)
ecs = class_vectors(y, y_unique).to(dtype=options['dtype'])
ucs = torch.stack(class_means(Ys, ecs))
us = ucs.mean(0)
y_unique_counts = ecs.sum(1)
out_dimension = Ys[0].size(1)
pairs = torch.combinations(y_unique_present, r=2)
class_Sw = torch.zeros(num_classes, out_dimension, out_dimension,
**options)
for ci in y_unique_present:
for vj in range(num_views):
for k in torch.where(ecs[ci])[0]:
d_ijk = Ys[vj][k] - us[ci]
class_Sw[ci] += d_ijk.unsqueeze(1) @ d_ijk.unsqueeze(0)
Sw = class_Sw[y_present_mask].sum(0)
out = torch.tensor(0, **options)
for ca, cb in pairs:
Sw_ab = beta * (y_unique_counts[ca] * class_Sw[ca] +
y_unique_counts[cb] * class_Sw[cb])
Sw_ab.div_(y_unique_counts[ca] + y_unique_counts[cb]).add_(
(1 - beta) * Sw)
du_ab = sum(uc[ca] for uc in ucs).sub(sum(
uc[cb] for uc in ucs)).div_(num_views).unsqueeze_(0)
Sb_ab = du_ab.t().mm(du_ab)
out += y_unique_counts[ca] * y_unique_counts[cb] * (
torch.trace(Sb_ab) / torch.trace(Sw_ab)).pow_(-q)
out /= num_components * num_components
return out
def _gen_pairs(input, dim=-2, reducer=lambda a, b: ((a - b)**2).sum(dim=-1)):
"""Generates all pairs of different rows and then applies the reducer
Args:
input: a tensor
dim: a dimension to generate pairs across
reducer: a function of generated pair of rows to apply (beyond just concat)
Returns:
for default args, for A x B x C input, will output A x (B choose 2)
"""
n = input.size()[dim]
range = torch.arange(n)
idx = torch.combinations(range).to(input).long()
left = input.index_select(dim, idx[:, 0])
right = input.index_select(dim, idx[:, 1])
return reducer(left, right)
def forward(self, xs, y_true):
_, x = xs
bs = x.shape[0]
idx = torch.combinations(torch.range(0, bs - 1), 2).long().cuda()
x1 = x[idx[:, 0]]
x2 = x[idx[:, 1]]
y1 = y_true[idx[:, 0]]
y2 = y_true[idx[:, 1]]
y = 2 * ((y1 == y2).float()) - 1
loss = self.cos_dis(x1, x2, y)
return loss
def stochastic_blockmodel_graph(block_sizes, edge_probs, directed=False):
r"""Returns the :obj:`edge_index` of a stochastic blockmodel graph.
Args:
block_sizes ([int] or LongTensor): The sizes of blocks.
edge_probs ([[float]] or FloatTensor): The density of edges going
from each block to each other block. Must be symmetric if the graph is
undirected.
directed (bool, optional): If set to :obj:`True`, will return a
directed graph. (default: :obj:`False`)
"""
size, prob = block_sizes, edge_probs
if not torch.is_tensor(size):
size = torch.tensor(size, dtype=torch.long)
if not torch.is_tensor(prob):
prob = torch.tensor(prob, dtype=torch.float)
assert size.dim() == 1
assert prob.dim() == 2 and prob.size(0) == prob.size(1)
assert size.size(0) == prob.size(0)
if not directed:
assert torch.allclose(prob, prob.t())
node_idx = torch.cat([size.new_full((b, ), i) for i, b in enumerate(size)])
num_nodes = node_idx.size(0)
if directed:
idx = torch.arange((num_nodes - 1) * num_nodes)
idx = idx.view(num_nodes - 1, num_nodes)
idx = idx + torch.arange(1, num_nodes).view(-1, 1)
idx = idx.view(-1)
row = idx // num_nodes
col = idx % num_nodes
else:
row, col = torch.combinations(torch.arange(num_nodes)).t()
mask = torch.bernoulli(prob[node_idx[row], node_idx[col]]).to(torch.bool)
edge_index = torch.stack([row[mask], col[mask]], dim=0)
if not directed:
edge_index = to_undirected(edge_index, num_nodes)
return edge_index
def update(self, y_pred: torch.Tensor, s_c: torch.Tensor):
assert y_pred.shape == s_c.shape
if self.k is None:
order = torch.argsort(input=y_pred, descending=True)
else:
sequence_length = y_pred.shape[0]
if sequence_length < self.k:
k = sequence_length
else:
k = self.k
_, order = torch.topk(input=y_pred, k=k, largest=True)
senti_score = torch.take(s_c, order)
senti_score = torch.combinations(senti_score)
senti_score = torch.abs(torch.diff(senti_score, dim=-1)) / 2
senti_score = torch.sum(senti_score) / senti_score.size(0)
self.ils_senti += senti_score
self.count += 1.0
def __init__(self,
args,
npoints=8,
input=64,
hidden=64,
hidden2=256,
sz=28,
sigma2=1e-2,
allow_points=True,
hard=False):
super().__init__(args)
self.hard = hard
# build the coordinate grid:
r = torch.linspace(-1, 1, sz)
c = torch.linspace(-1, 1, sz)
grid = torch.meshgrid(r, c)
grid = torch.stack(grid, dim=2)
self.register_buffer("grid", grid)
# if we allow points, we compute the upper-triangular part of the symmetric connection
# matrix including the diagonal. If points are not allowed, we don't need the diagonal values
# as they would be implictly zero
if allow_points:
nlines = int((npoints**2 + npoints) / 2)
else:
nlines = int(npoints * (npoints - 1) / 2)
self.coordpairs = torch.combinations(torch.arange(0,
npoints,
dtype=torch.long),
r=2,
with_replacement=allow_points)
# shared part of the encoder
self.enc1 = nn.Sequential(nn.Linear(input, hidden), nn.ReLU(),
nn.Linear(hidden, hidden2), nn.ReLU())
# second part for computing npoints 2d coordinates (using tanh because we use a -1..1 grid)
self.enc_pts = nn.Sequential(nn.Linear(hidden2, npoints * 2),
nn.Tanh())
# second part for computing upper triangular part of the connection matrix
self.enc_con = nn.Sequential(nn.Linear(hidden2, nlines), nn.Sigmoid())
self.sigma2 = sigma2
def get_pair_indices(inputs: Tensor, ordered_pair: bool = False) -> Tensor:
"""
Get pair indices between each element in input tensor
Args:
inputs: input tensor
ordered_pair: if True, will return ordered pairs. (e.g. both inputs[i,j] and inputs[j,i] are included)
Returns: a tensor of shape (K, 2) where K = choose(len(inputs),2) if ordered_pair is False.
Else K = 2 * choose(len(inputs),2). Each row corresponds to two indices in inputs.
"""
indices = torch.combinations(torch.tensor(range(len(inputs))), r=2)
if ordered_pair:
# make pairs ordered (e.g. both (0,1) and (1,0) are included)
indices = torch.cat((indices, indices[:, [1, 0]]), dim=0)
return indices
def compute_action(obs,
user_choice_model: UserChoiceModel,
q_model: QModel,
slate_size: int = 3):
user = pack_state_user(obs)
doc = pack_state_doc(obs)
user = torch.tensor(user).unsqueeze(0)
doc = torch.tensor(doc).unsqueeze(0)
with torch.no_grad():
scores = user_choice_model(user, doc).squeeze(0) # shape=[num_docs+1]
scores = torch.exp(scores - torch.max(scores, dim=-1)[0])
scores_doc = scores[:-1] # shape=[num_docs]
score_no_click = scores[-1] # shape=[]
q_values = q_model(user, doc).squeeze(0) # shape=[num_docs+1]
q_values_doc = q_values[:-1] # shape=[num_docs]
q_values_no_click = q_values[-1] # shape=[]
num_docs = len(obs["doc"])
indices = torch.tensor(np.arange(num_docs)).long()
slates = torch.combinations(indices,
r=slate_size) # shape=[num_slates, num_docs]
num_slates, _ = slates.shape
slate_decomp_q_values = torch.gather(
q_values_doc.unsqueeze(0).expand(num_slates, num_docs), 1,
slates) # shape=[num_slates, slate_size]
slate_scores = torch.gather(
scores_doc.unsqueeze(0).expand(num_slates, num_docs), 1,
slates) # shape=[num_slates, slate_size]
slate_q_values = (slate_decomp_q_values * slate_scores +
q_values_no_click * score_no_click).sum(dim=1) / (
slate_scores.sum(dim=1) + score_no_click
) # shape=[num_slates]
idx = np.argmax(slate_q_values.detach().numpy())
selected = slates[idx].detach().numpy().tolist()
action = tuple(selected)
print("compute_action",
q_values.detach().numpy(),
scores.detach().numpy(), action)
return action
def main(args):
subsets = {}
modules = {}
reindex_fn = lambda x: x
if args.n_modules is not None and args.n_active is not None:
print("n-modules and n-active set... using module reindexing...", file=sys.stderr)
# We use torch to guarantee that the subset indexing will be same as with our model
combos = torch.combinations(torch.arange(0, args.n_modules), args.n_active).numpy().tolist()
reindex_fn = lambda x: combos[x]
n_lines = 0
with open(args.input_file, "r") as fh:
for line in fh:
n_lines += 1
line = line.strip().split("\t")
curr_modules = [int(x) for x in line[-1].split(",")]
curr_modules = [y for x in curr_modules for y in reindex_fn(x)]
curr_modules = [str(x) for x in curr_modules]
s = ",".join(curr_modules)
if s in subsets:
subsets[s] += 1
else:
subsets[s] = 1
for m in curr_modules:
if m in modules:
modules[m] += 1
else:
modules[m] = 1
print("N-subsets\t{}".format(len(subsets.keys())))
print("N-modules\t{}".format(len(modules.keys())))
print("N-ratio\t{}".format(len(subsets.keys()) / float(len(combos))))
for k, v in sorted(subsets.items(), key=lambda item: item[1], reverse=True):
print("S-{}\t{}\t{}".format(k, v, (v / float(n_lines))))
for k, v in sorted(modules.items(), key=lambda item: item[1], reverse=True):
print("M-{}\t{}\t{}".format(k, v, (v / float(n_lines))))
def forward(self, node_features):
"""
FORWARD ROUTINE
- computes Adjacency Matrix
"""
n_nodes = node_features.shape[-1]
index = torch.arange(n_nodes)
index_combinations = torch.combinations(index, r=2)
adjacency_matrix = torch.zeros(node_features.shape[0], n_nodes,
n_nodes)
for i, j in index_combinations:
adjacency_matrix[:, j,
i] = adjacency_matrix[:, i, j] = torch.sigmoid(
self.fc1(
torch.cat((node_features[:, :, i],
node_features[:, :, j]),
dim=1))).squeeze()
return adjacency_matrix
