def encode(matched, priors, use_yolo_regressors: bool = False):
"""
Encode bboxes matched with each prior into the format
produced by the network. See decode for more details on
this format. Note that encode(decode(x, p), p) = x.
Args:
- matched: A tensor of bboxes in point form with shape [num_priors, 4]
- priors: The tensor of all priors with shape [num_priors, 4]
Return: A tensor with encoded relative coordinates in the format
outputted by the network (see decode). Size: [num_priors, 4]
"""
# print(use_yolo_regressors)
if use_yolo_regressors:
# Exactly the reverse of what we did in decode
# In fact encode(decode(x, p), p) should be x
boxes = center_size(matched)
loc = jt.contrib.concat(
(boxes[:, :2] - priors[:, :2], jt.log(
boxes[:, 2:] / priors[:, 2:])), 1)
else:
variances = [0.1, 0.2]
# dist b/t match center and prior's center
g_cxcy = (matched[:, :2] + matched[:, 2:]) / 2 - priors[:, :2]
# encode variance
g_cxcy /= (variances[0] * priors[:, 2:])
# match wh / prior wh
g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
g_wh = jt.log(g_wh) / variances[1]
# return target for smooth_l1_loss
loc = jt.contrib.concat([g_cxcy, g_wh], 1) # [num_priors,4]
return loc
def execute(self, inputs, targets, mask=None, act=False):
losses = []
for id in range(len(inputs)):
if mask is not None:
input_flatten, target_flatten = self.flatten(
inputs[id], targets[id], mask[id])
else:
input_flatten, target_flatten = self.flatten(
inputs[id], targets[id])
if act:
MIN = 1e-9
input_flatten = jt.clamp(input_flatten,
min_v=MIN,
max_v=1 - MIN)
input_flatten = jt.log(input_flatten) - jt.log(1 -
input_flatten)
losses.append(self.lovasz_hinge_flat(input_flatten,
target_flatten))
losses = jt.stack(losses)
if self.reduction == "mean":
losses = losses.mean()
elif self.reduction == "sum":
losses = losses.sum()
return losses
def encode(self, reference_boxes, proposals):
"""
Encode a set of proposals with respect to some
reference boxes
Arguments:
reference_boxes (Tensor): reference boxes
proposals (Tensor): boxes to be encoded
"""
TO_REMOVE = 1 # TODO remove
ex_widths = proposals[:, 2] - proposals[:, 0] + TO_REMOVE
ex_heights = proposals[:, 3] - proposals[:, 1] + TO_REMOVE
ex_ctr_x = proposals[:, 0] + 0.5 * ex_widths
ex_ctr_y = proposals[:, 1] + 0.5 * ex_heights
gt_widths = reference_boxes[:, 2] - reference_boxes[:, 0] + TO_REMOVE
gt_heights = reference_boxes[:, 3] - reference_boxes[:, 1] + TO_REMOVE
gt_ctr_x = reference_boxes[:, 0] + 0.5 * gt_widths
gt_ctr_y = reference_boxes[:, 1] + 0.5 * gt_heights
wx, wy, ww, wh = self.weights
targets_dx = wx * (gt_ctr_x - ex_ctr_x) / ex_widths
targets_dy = wy * (gt_ctr_y - ex_ctr_y) / ex_heights
targets_dw = ww * jt.log(gt_widths / ex_widths)
targets_dh = wh * jt.log(gt_heights / ex_heights)
targets = jt.stack((targets_dx, targets_dy, targets_dw, targets_dh), dim=1)
return targets
def bce_loss(output, target, size_average=True):
if size_average:
return -(target * jt.log(jt.maximum(output, 1e-20)) +
(1 - target) * jt.log(jt.maximum(1 - output, 1e-20))).mean()
else:
return -(target * jt.log(jt.maximum(output, 1e-20)) +
(1 - target) * jt.log(jt.maximum(1 - output, 1e-20))).sum()
def bce_loss(output, target, weight=None, size_average=True):
loss = - (target * jt.log(jt.maximum(output, 1e-20)) + (1 - target) * jt.log(jt.maximum(1 - output, 1e-20)))
if weight is not None:
loss *= weight
if size_average:
return loss.mean()
else:
return loss.sum()
def sigmoid_focal_loss(logits, targets, gamma, alpha):
num_classes = logits.shape[1]
dtype = targets.dtype
class_range = jt.arange(1, num_classes + 1, dtype=dtype).unsqueeze(0)
t = targets.unsqueeze(1)
p = logits.sigmoid()
term1 = (1 - p)**gamma * jt.log(p)
term2 = p**gamma * jt.log(1 - p)
return -(t == class_range).float() * term1 * alpha - (
(t != class_range) * (t >= 0)).float() * term2 * (1 - alpha)
def execute(self, x):
batch_size = x.shape[0]
x = nn.relu(self.fc1(x))
x = nn.relu(self.fc2(x))
# decoder follows NMR
centroid = self.fc_centroid(x) * self.centroid_scale
bias = self.fc_bias(x) * self.bias_scale
bias = bias.view(-1, self.nv, 3)
base = self.vertices_base * self.obj_scale
sign = nn.sign(base)
base = base.abs()
base = jt.log(base / (1 - base))
centroid = jt.tanh(centroid[:, None, :])
scale_pos = 1 - centroid
scale_neg = centroid + 1
vertices = (base + bias).sigmoid() * sign
vertices = nn.relu(vertices) * scale_pos - nn.relu(
-vertices) * scale_neg
vertices = vertices + centroid
vertices = vertices * 0.5
faces = self.faces[None, :, :].repeat(batch_size, 1, 1)
return vertices, faces
def __init__(self, p=None, logits=None):
assert (p is not None) or (logits is not None)
assert 0 < p and p < 1
if p is None:
self.prob = jt.sigmoid(logits)
self.logits = logits
elif logits is None:
self.prob = p
self.logits = -jt.log(1. / p - 1)
def log_sum_exp(x):
"""Utility function for computing log_sum_exp while determining
This will be used to determine unaveraged confidence loss across
all examples in a batch.
Args:
x (Variable(tensor)): conf_preds from conf layers
"""
x_max = x.data.max()
return jt.log(jt.sum(jt.exp(x - x_max), 1)) + x_max
def execute(self, batch_size):
base = jt.log(self.vertices.abs() / (1 - self.vertices.abs()))
centroid = jt.tanh(self.center)
vertices = (base + self.displace).sigmoid() * nn.sign(self.vertices)
vertices = nn.relu(vertices) * (1 - centroid) - nn.relu(-vertices) * (centroid + 1)
vertices = vertices + centroid
# apply Laplacian and flatten geometry constraints
laplacian_loss = self.laplacian_loss(vertices).mean()
flatten_loss = self.flatten_loss(vertices).mean()
return jr.Mesh(vertices.repeat(batch_size, 1, 1),
self.faces.repeat(batch_size, 1, 1), dr_type='n3mr'), laplacian_loss, flatten_loss
def kl_divergence(cur_dist, old_dist):
assert isinstance(cur_dist, type(old_dist))
if isinstance(cur_dist, Normal):
vr = (cur_dist.sigma / old_dist.sigma)**2
t1 = ((cur_dist.mu - old_dist.mu) / old_dist.sigma)**2
return 0.5 * (vr + t1 - 1 - jt.log(vr))
if isinstance(cur_dist, Categorical) or isinstance(cur_dist,
OneHotCategorical):
t = cur_dist.probs * (cur_dist.logits - old_dist.logits)
t[jt.array((old_dist.probs == 0))] = math.inf
t[jt.array((cur_dist.probs == 0))] = 0
return t.sum(-1)
if isinstance(cur_dist, Uniform):
res = jt.log(
(old_dist.high - old_dist.low) / (cur_dist.high - cur_dist.low))
if old_dist.low > cur_dist.low or old_dist.high < cur_dist.high:
res = math.inf
return res
if isinstance(cur_dist, Geometric):
return -cur_dist.entropy() - jt.log(
-old_dist.prob + 1) / cur_dist.prob - old_dist.logits
def kl_divergence(cur_dist, old_dist):
assert isinstance(cur_dist, type(old_dist))
if isinstance(cur_dist, Normal):
vr = (cur_dist.sigma / old_dist.sigma)**2
t1 = ((cur_dist.mu - old_dist.mu) / old_dist.sigma)**2
return 0.5 * (vr + t1 - 1 - jt.log(vr))
if isinstance(cur_dist, Categorical) or isinstance(cur_dist,
OneHotCategorical): # ?
t = cur_dist.probs * (cur_dist.logits - old_dist.logits)
t[jt.array((old_dist.probs == 0))] = math.inf
t[jt.array((cur_dist.probs == 0))] = 0
return t.sum(-1)
def __init__(self, probs=None, logits=None):
assert not (probs is None and logits is None)
if probs is None:
# cannot align to pytorch
probs = jt.sigmoid(logits)
elif logits is None:
logits = jt.log(probs)
with jt.no_grad():
self.probs = probs / probs.sum(-1, True)
self.logits = logits
self.cum_probs = simple_presum(probs)
self.cum_probs_l = self.cum_probs[..., :-1]
self.cum_probs_r = self.cum_probs[..., 1:]
def bbox2loc(src_bbox,dst_bbox):
width = src_bbox[:, 2:3] - src_bbox[:, 0:1]
height = src_bbox[:, 3:4] - src_bbox[:, 1:2]
center_x = src_bbox[:, 0:1] + 0.5 * width
center_y = src_bbox[:, 1:2] + 0.5 * height
base_width = dst_bbox[:, 2:3] - dst_bbox[:, 0:1]
base_height = dst_bbox[:, 3:4] - dst_bbox[:, 1:2]
base_center_x = dst_bbox[:, 0:1] + 0.5 * base_width
base_center_y = dst_bbox[:, 1:2] + 0.5 * base_height
eps = 1e-5
height = jt.maximum(height, eps)
width = jt.maximum(width, eps)
dy = (base_center_y - center_y) / height
dx = (base_center_x - center_x) / width
dw = jt.log(base_width / width)
dh = jt.log(base_height / height)
loc = jt.contrib.concat([dx,dy,dw,dh],dim=1)
return loc
def execute(self, pred, target, weight=None):
pred_left = pred[:, 0]
pred_top = pred[:, 1]
pred_right = pred[:, 2]
pred_bottom = pred[:, 3]
target_left = target[:, 0]
target_top = target[:, 1]
target_right = target[:, 2]
target_bottom = target[:, 3]
target_area = (target_left + target_right) * \
(target_top + target_bottom)
pred_area = (pred_left + pred_right) * \
(pred_top + pred_bottom)
w_intersect = jt.minimum(pred_left, target_left) + jt.minimum(
pred_right, target_right)
g_w_intersect = jt.maximum(pred_left, target_left) + jt.maximum(
pred_right, target_right)
h_intersect = jt.minimum(pred_bottom, target_bottom) + jt.minimum(
pred_top, target_top)
g_h_intersect = jt.maximum(pred_bottom, target_bottom) + jt.maximum(
pred_top, target_top)
ac_uion = g_w_intersect * g_h_intersect + 1e-7
area_intersect = w_intersect * h_intersect
area_union = target_area + pred_area - area_intersect
ious = (area_intersect + 1.0) / (area_union + 1.0)
gious = ious - (ac_uion - area_union) / ac_uion
if self.loc_loss_type == 'iou':
losses = -jt.log(ious)
elif self.loc_loss_type == 'linear_iou':
losses = 1 - ious
elif self.loc_loss_type == 'giou':
losses = 1 - gious
else:
raise NotImplementedError
if weight is not None and weight.sum() > 0:
return (losses * weight).sum() / weight.sum()
else:
assert losses.numel() != 0
return losses.mean()
def soft_cross_entropy_loss(output, target, smoothing=True):
''' Calculate cross entropy loss, apply label smoothing if needed. '''
target = target.view(-1)
softmax = nn.Softmax(dim=1)
if smoothing:
eps = 0.2
b, n_class = output.shape
one_hot = jt.zeros(output.shape)
for i in range(b):
one_hot[i, target[i].data] = 1
one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
# print (one_hot[0].data)
log_prb = jt.log(softmax(output))
loss = -(one_hot * log_prb).sum(dim=1).mean()
else:
loss = nn.cross_entropy_loss(output, target)
return loss
def log2(x):
return jt.log(x)/math.log(2.0)
def log_sigmoid(x):
return jt.log(jt.sigmoid(x))
def log_softmax(x, dim=None):
x = softmax(x, dim=dim)
return jt.log(x)
def execute(self, x):
return 1 / self.beta * jt.log(1 + (self.beta * x).exp())
def softplus(x, beta=1, threshold=20):
return 1 / beta * jt.log(1 + (beta * x).exp())
def __init__(self, loc, scale):
self.loc = loc
self.scale = scale
self.log_scale = jt.log(self.scale)
def entropy(self):
return -jt.sum(jt.mean(self.probs) * jt.log(self.probs))
def log_prob(self, x):
return jt.log(self.probs)[0,x]
def logits(self):
if self._logits is None:
return jt.log(jt.clamp(self.probs, min_v=eps, max_v=1-eps))
else:
return self._logits
import jittor as jt
from jittor import nn
import numpy as np
# Misc
img2mse = lambda x, y: jt.mean((x - y)**2)
mse2psnr = lambda x: -10. * jt.log(x) / jt.log(jt.array(np.array([10.])))
to8b = lambda x: (255 * np.clip(x, 0, 1)).astype(np.uint8)
# Positional encoding (section 5.1)
class Embedder:
def __init__(self, **kwargs):
self.kwargs = kwargs
self.create_embedding_fn()
def create_embedding_fn(self):
embed_fns = []
d = self.kwargs['input_dims']
out_dim = 0
if self.kwargs['include_input']:
embed_fns.append(lambda x: x)
out_dim += d
max_freq = self.kwargs['max_freq_log2']
N_freqs = self.kwargs['num_freqs']
if self.kwargs['log_sampling']:
freq_bands = 2.**jt.linspace(0., max_freq, steps=N_freqs)
else:
freq_bands = jt.linspace(2.**0., 2.**max_freq, steps=N_freqs)
def bce_loss(output, target):
return -(target * jt.log(jt.maximum(output, 1e-20)) +
(1 - target) * jt.log(jt.maximum(1 - output, 1e-20))).mean()
def prod(x,dim=0):
x = jt.log(x)
x = x.sum(dim=dim)
return jt.exp(x)
def cumprod(x,dim=0):
x = jt.log(x)
x = cumsum(x,dim=dim)
return jt.exp(x)
def execute(self, input, target):
bs_idx = jt.array(range(input.shape[0]))
ret = (-jt.log(nn.softmax(input, dim=1)))[bs_idx, target]
if self.reduction != None:
ret = jt.mean(ret) if self.reduction == 'mean' else jt.sum(ret)
return ret