def _interpolate(im, x, y, out_size):
num_batch, height, width, channels = im.size(
) # to be sure the input dims is NHWC
x = torch._cast_Float(x).cuda()
y = torch._cast_Float(y).cuda()
height_f = torch._cast_Float(torch.Tensor([height]))[0].cuda()
width_f = torch._cast_Float(torch.Tensor([width]))[0].cuda()
out_height = out_size[0]
out_width = out_size[1]
zero = torch.zeros([], dtype=torch.int32).cuda()
max_y = torch._cast_Long(torch.Tensor([height - 1]))[0].cuda()
max_x = torch._cast_Long(torch.Tensor([width - 1]))[0].cuda()
# scale indices from [-1, 1] to [0, width/height]
x = (x + 1.0) * width_f / 2.0
y = (y + 1.0) * height_f / 2.0
# do sampling
x0 = torch._cast_Long(torch.floor(x)).cuda()
x1 = x0 + 1
y0 = torch._cast_Long(torch.floor(y)).cuda()
y1 = y0 + 1
x0 = torch.clamp(x0, zero, max_x)
x1 = torch.clamp(x1, zero, max_x)
y0 = torch.clamp(y0, zero, max_y)
y1 = torch.clamp(y1, zero, max_y)
dim2 = width
dim1 = width * height
base = _repeat(torch.arange(num_batch) * dim1,
out_height * out_width).cuda()
base_y0 = base + y0 * dim2
base_y1 = base + y1 * dim2
idx_a = base_y0 + x0
idx_b = base_y1 + x0
idx_c = base_y0 + x1
idx_d = base_y1 + x1
# use indices to look up pixels in the flate images
# and restore channels dim
im_flat = im.contiguous().view(-1, channels)
im_flat = torch._cast_Float(im_flat)
Ia = im_flat[idx_a] # as in tf, the default dim is row first
Ib = im_flat[idx_b]
Ic = im_flat[idx_c]
Id = im_flat[idx_d]
# calculate interpolated values
x0_f = torch._cast_Float(x0).cuda()
x1_f = torch._cast_Float(x1).cuda()
y0_f = torch._cast_Float(y0).cuda()
y1_f = torch._cast_Float(y1).cuda()
wa = ((x1_f - x) * (y1_f - y)).unsqueeze(1)
wb = ((x1_f - x) * (y - y0_f)).unsqueeze(1)
wc = ((x - x0_f) * (y1_f - y)).unsqueeze(1)
wd = ((x - x0_f) * (y - y0_f)).unsqueeze(1)
return wa * Ia + wb * Ib + wc * Ic + wd * Id
def aggregate_accuracy(r_dict, metric_name_list):
m_list = []
for metric_name in metric_name_list:
m_list.append(r_dict[metric_name][0])
m_list[0] = torch._cast_Long(m_list[0])
m_list[1] = torch._cast_Long(m_list[1])
m_list[2] = torch._cast_Long(m_list[2])
agg = torch.stack(m_list, dim=0)
agg = agg.prod(0, keepdim=False)
return (agg.sum(), agg.numel())
def loss(self, score, truth):
loss = 0
B = len(truth)
agg_num_score, agg_score = score
#loss for the column number
truth_num = [len(t) for t in truth] # double check truth format and for test cases
data = torch.from_numpy(np.array(truth_num))
data = torch._cast_Long(data)
if self.gpu:
truth_num_var = Variable(data.cuda())
else:
truth_num_var = Variable(data)
loss += self.CE(agg_num_score, truth_num_var)
#loss for the key words
T = len(agg_score[0])
truth_prob = np.zeros((B, T), dtype=np.float32)
for b in range(B):
truth_prob[b][truth[b]] = 1
data = torch.from_numpy(truth_prob)
if self.gpu:
truth_var = Variable(data.cuda())
else:
truth_var = Variable(data)
#loss += self.mlsml(agg_score, truth_var)
#loss += self.bce_logit(agg_score, truth_var) # double check no sigmoid
pred_prob = self.sigm(agg_score)
bce_loss = -torch.mean( 3*(truth_var * \
torch.log(pred_prob+1e-10)) + \
(1-truth_var) * torch.log(1-pred_prob+1e-10) )
loss += bce_loss
return loss
def __init__(self, ndim, ind, scale=None):
"""
Constructor
:param ndim: Int, number of dimensions
:param ind: Iterable, indices of uniformly distributed entries
:param scale: Iterable, standard deviation of Gaussian or width of
uniform distribution
"""
super().__init__()
self.ndim = ndim
# Set up indices and permutations
self.ndim = ndim
if torch.is_tensor(ind):
self.register_buffer('ind', torch._cast_Long(ind))
else:
self.register_buffer('ind', torch.tensor(ind, dtype=torch.long))
ind_ = []
for i in range(self.ndim):
if not i in self.ind:
ind_ += [i]
self.register_buffer('ind_', torch.tensor(ind_, dtype=torch.long))
perm_ = torch.cat((self.ind, self.ind_))
inv_perm_ = torch.zeros_like(perm_)
for i in range(self.ndim):
inv_perm_[perm_[i]] = i
self.register_buffer('inv_perm', inv_perm_)
if scale is None:
self.register_buffer('scale', torch.ones(self.ndim))
else:
self.register_buffer('scale', scale)
def loss(self, score, truth):
loss = 0
B = len(truth)
op_num_score, op_score = score
truth = [t if len(t) <= 2 else t[:2] for t in truth]
# loss for the op number
truth_num = [len(t) - 1
for t in truth] #num_score 0 maps to 1 in truth
data = torch.from_numpy(np.array(truth_num))
data = torch._cast_Long(data)
if self.gpu:
truth_num_var = Variable(data.cuda())
else:
truth_num_var = Variable(data)
loss += self.CE(op_num_score, truth_num_var)
# loss for op
T = len(op_score[0])
truth_prob = np.zeros((B, T), dtype=np.float32)
for b in range(B):
truth_prob[b][truth[b]] = 1
data = torch.from_numpy(np.array(truth_prob))
if self.gpu:
truth_var = Variable(data.cuda())
else:
truth_var = Variable(data)
#loss += self.mlsml(op_score, truth_var)
#loss += self.bce_logit(op_score, truth_var)
pred_prob = self.sigm(op_score)
bce_loss = -torch.mean( 3*(truth_var * \
torch.log(pred_prob+1e-10)) + \
(1-truth_var) * torch.log(1-pred_prob+1e-10) )
loss += bce_loss
return loss
def _repeat(x, n_repeats):
rep = torch._cast_Long(
torch.transpose(torch.ones([
n_repeats,
]).unsqueeze(1), 1, 0))
x = torch.matmul(x.view(-1, 1), rep)
return x.view(-1)
def loss(self, score, truth):
loss = 0
data = torch.from_numpy(np.array(truth))
data = torch._cast_Long(data)
if self.gpu:
truth_var = Variable(data.cuda())
else:
truth_var = Variable(data)
loss = self.CE(score, truth_var)
return loss
def forward(self, inputs: torch.Tensor):
outputs = torch.zeros(
*inputs.size(), self.hidden_size,
device=inputs.device) # [batch, label_num, hidden_dim]
for left, right, emb in zip(self.group_offset[:-1],
self.group_offset[1:], self.emb):
index = (left <= inputs) & (inputs < right
) # bool类型 [batch, label_num]
group_inputs = torch._cast_Long(
(inputs[index] - left).to(emb.weight.device))
outputs[index] = emb(group_inputs).to(inputs.device)
return outputs # [batch, label_num, hidden_dim]
def warp_pts(self, pts, flow):
x = pts[:, :, 0]
x = torch.clamp((x + 1) / 2 * width, 0, width - 1)
x = torch._cast_Int(torch.round(x))
y = pts[:, :, 1]
y = torch.clamp((y + 1) / 2 * height, 0, height - 1)
y = torch._cast_Int(torch.round(y))
out = []
for i in range(batch_size):
flow_ = flow[i, :, :, :].view([-1, 2])
xy = x[i, :] + y[i, :] * width
xy = torch._cast_Long(xy)
temp = flow_[xy]
out.append(temp.view([1, max_matches, 2]))
return torch.cat(out, 0)
def __init__(self, ndim, ind, scale=1., bias=False, activation=None):
"""
Constructor
:param ndim: Int, number of dimensions
:param ind: Iterable, indices of input elements to convert to
periodic features
:param scale: Scalar or iterable, used to scale inputs before
converting them to periodic features
:param bias: Flag, whether to add a bias
:param activation: Function or None, activation function to be
applied
"""
super(PeriodicFeatures, self).__init__()
# Set up indices and permutations
self.ndim = ndim
if torch.is_tensor(ind):
self.register_buffer('ind', torch._cast_Long(ind))
else:
self.register_buffer('ind', torch.tensor(ind, dtype=torch.long))
ind_ = []
for i in range(self.ndim):
if not i in self.ind:
ind_ += [i]
self.register_buffer('ind_', torch.tensor(ind_, dtype=torch.long))
perm_ = torch.cat((self.ind, self.ind_))
inv_perm_ = torch.zeros_like(perm_)
for i in range(self.ndim):
inv_perm_[perm_[i]] = i
self.register_buffer('inv_perm', inv_perm_)
self.weights = nn.Parameter(torch.ones(len(self.ind), 2))
if torch.is_tensor(scale):
self.register_buffer('scale', scale)
else:
self.scale = scale
self.apply_bias = bias
if self.apply_bias:
self.bias = nn.Parameter(torch.zeros(len(self.ind)))
if activation is None:
self.activation = lambda input: input
else:
self.activation = activation
def loss(self, score, truth):
#here suppose truth looks like [[[1, 4], 3], [], ...]
loss = 0
B = len(truth)
col_num_score, col_score = score
#loss for the column number
truth_num = [len(t) - 1 for t in truth
] # double check truth format and for test cases
data = torch.from_numpy(np.array(truth_num, dtype=np.long))
data = torch._cast_Long(data)
if self.gpu:
truth_num_var = Variable(data.cuda())
else:
truth_num_var = Variable(data)
loss += self.CE(col_num_score, truth_num_var)
#loss for the key words
T = len(col_score[0])
# print("T {}".format(T))
truth_prob = np.zeros((B, T), dtype=np.float32)
for b in range(B):
gold_l = []
for t in truth[b]:
if isinstance(t, list):
gold_l.extend(t)
else:
gold_l.append(t)
truth_prob[b][gold_l] = 1
data = torch.from_numpy(truth_prob)
# print("data {}".format(data))
# print("data {}".format(data.cuda()))
if self.gpu:
truth_var = Variable(data.cuda())
else:
truth_var = Variable(data)
#loss += self.mlsml(col_score, truth_var)
#loss += self.bce_logit(col_score, truth_var) # double check no sigmoid
pred_prob = self.sigm(col_score)
bce_loss = -torch.mean( 3*(truth_var * \
torch.log(pred_prob+1e-10)) + \
(1-truth_var) * torch.log(1-pred_prob+1e-10) )
loss += bce_loss
return loss
