def compute_loss(y, gt_map):
# 1. segmentation loss
target_labels = torch.argmax(gt_map[:, MapOrdering.SEG_WORD:MapOrdering.SEG_BACKGROUND + 1], dim=1)
loss_seg = F.cross_entropy(y[:, MapOrdering.SEG_WORD:MapOrdering.SEG_BACKGROUND + 1], target_labels)
# 2. geometry loss
# distances to all sides of aabb
t = torch.minimum(y[:, MapOrdering.GEO_TOP], gt_map[:, MapOrdering.GEO_TOP])
b = torch.minimum(y[:, MapOrdering.GEO_BOTTOM], gt_map[:, MapOrdering.GEO_BOTTOM])
l = torch.minimum(y[:, MapOrdering.GEO_LEFT], gt_map[:, MapOrdering.GEO_LEFT])
r = torch.minimum(y[:, MapOrdering.GEO_RIGHT], gt_map[:, MapOrdering.GEO_RIGHT])
# area of predicted aabb
y_width = y[:, MapOrdering.GEO_LEFT, ...] + y[:, MapOrdering.GEO_RIGHT, ...]
y_height = y[:, MapOrdering.GEO_TOP, ...] + y[:, MapOrdering.GEO_BOTTOM, ...]
area1 = y_width * y_height
# area of gt aabb
gt_width = gt_map[:, MapOrdering.GEO_LEFT, ...] + gt_map[:, MapOrdering.GEO_RIGHT, ...]
gt_height = gt_map[:, MapOrdering.GEO_TOP, ...] + gt_map[:, MapOrdering.GEO_BOTTOM, ...]
area2 = gt_width * gt_height
# compute intersection over union
intersection = (r + l) * (b + t)
union = area1 + area2 - intersection
eps = 0.01 # avoid division by 0
iou = intersection / (union + eps)
iou = iou[gt_map[:, MapOrdering.SEG_WORD] > 0]
loss_aabb = -torch.log(torch.mean(iou))
# total loss is simply the sum of both losses
loss = loss_seg + loss_aabb
return loss
def small_IoU(box_i, box_j):
device = box_i.device
dtype = box_i.dtype
pw_a = box_i[2] - box_i[0]
ph_a = box_i[3] - box_i[1]
area_p = (pw_a * ph_a)
bw_a = box_j[2] - box_j[0]
bh_a = box_j[3] - box_j[1]
area_b = (bw_a * bh_a)
area_u = area_p + area_b
x_val = torch.minimum(box_i[2], box_j[2]) - torch.maximum(
box_i[0], box_j[0])
x_val_zero = torch.zeros(x_val.shape, device=device, dtype=dtype)
y_val = torch.minimum(box_i[3], box_j[3]) - torch.maximum(
box_i[1], box_j[1])
y_val_zero = torch.zeros(y_val.shape, device=device, dtype=dtype)
area_i = torch.maximum(x_val, x_val_zero) * torch.maximum(
y_val, y_val_zero)
area_u -= area_i
return area_i / area_u
def compute_iou(pred, gt):
""" Calculates IoU (Jaccard index) of two sets of bboxes:
IOU = pred ∩ gt / (area(pred) + area(gt) - pred ∩ gt)
Parameters:
Coordinates of bboxes are supposed to be in the following form: [x1, y1, x2, y2]
pred (torch.tensor): predicted bboxes
gt (torch.tensor): ground truth bboxes
Return value:
iou (torch.tensor): intersection over union
"""
def get_box_area(box):
return (box[:, 2] - box[:, 0] + 1.) * (box[:, 3] - box[:, 1] + 1.)
#_gt = torch.tile(gt, (pred.shape[0], 1))
_gt = gt.repeat(pred.shape[0], 1)
_pred = torch.repeat_interleave(pred, gt.shape[0], dim=0)
ixmin = torch.maximum(_gt[:, 0], _pred[:, 0])
iymin = torch.maximum(_gt[:, 1], _pred[:, 1])
ixmax = torch.minimum(_gt[:, 2], _pred[:, 2])
iymax = torch.minimum(_gt[:, 3], _pred[:, 3])
width = torch.maximum(ixmax - ixmin + 1., torch.tensor(0))
height = torch.maximum(iymax - iymin + 1., torch.tensor(0))
intersection_area = width * height
union_area = get_box_area(_gt) + get_box_area(_pred) - intersection_area
iou = (intersection_area / union_area).reshape(pred.shape[0], gt.shape[0])
return iou
def find_active_constraints(x, lb, ub, rtol=1e-10):
"""Determine which constraints are active in a given point.
The threshold is computed using `rtol` and the absolute value of the
closest bound.
Returns
-------
active : ndarray of int with shape of x
Each component shows whether the corresponding constraint is active:
* 0 - a constraint is not active.
* -1 - a lower bound is active.
* 1 - a upper bound is active.
"""
active = torch.zeros_like(x, dtype=torch.long)
if rtol == 0:
active[x <= lb] = -1
active[x >= ub] = 1
return active
lower_dist = x - lb
upper_dist = ub - x
lower_threshold = rtol * lb.abs().clamp(1, None)
upper_threshold = rtol * ub.abs().clamp(1, None)
lower_active = (lb.isfinite() &
(lower_dist <= torch.minimum(upper_dist, lower_threshold)))
active[lower_active] = -1
upper_active = (ub.isfinite() &
(upper_dist <= torch.minimum(lower_dist, upper_threshold)))
active[upper_active] = 1
return active
def rbox_loss(pred_geometry_map: Tensor,
target_geometry_map: Tensor,
train_ignore_mask: Tensor,
train_boundary_mask: Tensor,
angle_lambda: int = 10) -> Tensor:
pred_top, pred_right, pred_bottom, pred_left, pred_angle = torch.split(
pred_geometry_map, split_size_or_sections=[1, 1, 1, 1, 1], dim=2)
target_top, target_right, target_bottom, target_left, target_angle = torch.split(
target_geometry_map, split_size_or_sections=[1, 1, 1, 1, 1], dim=2)
pred_area = (pred_top + pred_bottom) * (pred_left + pred_right)
target_area = (target_top + target_bottom) * (target_left + target_right)
h_inter = torch.minimum(pred_right, target_right) + torch.minimum(
pred_left, target_left)
w_inter = torch.minimum(pred_top, target_top) + torch.minimum(
pred_bottom, target_bottom)
intersection = h_inter * w_inter
union = pred_area + target_area - intersection
box_loss = -torch.log((intersection + 1e-6) / (union + 1e-6))
angle_loss = 1 - torch.cos(pred_angle - target_angle)
rbox_loss = box_loss + angle_lambda * angle_loss
train_mask = torch.unsqueeze(train_ignore_mask * train_boundary_mask,
dim=2)
rbox_loss = torch.mean(
torch.sum(rbox_loss * target_geometry_map * train_mask))
return rbox_loss
def bbox_iou(bboxes1, bboxes2):
eps = 0.001
EPS = 1.0e+20
x1, y1 = bboxes1[..., 0], bboxes1[..., 1]
w1, h1 = bboxes1[..., 2], bboxes1[..., 3]
x2, y2 = bboxes2[..., 0], bboxes2[..., 1]
w2, h2 = bboxes2[..., 2], bboxes2[..., 3]
xmin1 = x1 - (w1 * 0.5)
xmax1 = x1 + (w1 * 0.5)
ymin1 = y1 - (h1 * 0.5)
ymax1 = y1 + (h1 * 0.5)
xmin2 = x2 - (w2 * 0.5)
xmax2 = x2 + (w2 * 0.5)
ymin2 = y2 - (h2 * 0.5)
ymax2 = y2 + (h2 * 0.5)
xmini = torch.maximum(xmin1, xmin2)
ymini = torch.maximum(ymin1, ymin2)
xmaxi = torch.minimum(xmax1, xmax2)
ymaxi = torch.minimum(ymax1, ymax2)
wi = torch.clamp(xmaxi - xmini, min=0.0)
hi = torch.clamp(ymaxi - ymini, min=0.0)
area1 = torch.clamp(w1 * h1, max=EPS)
area2 = torch.clamp(w2 * h2, max=EPS)
areai = torch.clamp(wi * hi, max=EPS)
areau = area1 + area2 - areai
iou = areai / (areau + eps)
return iou
def multiplex(self, subframes: torch.Tensor):
"""
Assuming that the number of subframes is divisible by the number of pixles in a neighborhood (n_pixels), bucket 1 is active in the following cases:
* pixel 1 of each neighborhood in the first S / n_pixels
* pixel 2 of each neighborhood in the second S / n_pixels
* pixel 3 of each neighborhood in the third S / n_pixels
...
* last pixel of each neighborhood in the last S / n_pixels
subframes: intensities of subframes of a frame as a numpy array with shape (S, H, W)
nbhd_size: tuple of width and height of the neighborhood (For now W must be divisible by the width of the nbhd
and H must be divisible by the height of the neighborhood)
"""
# S, height, width = raw_subframes.shape
# scheme_ = torch.tile(self.scheme, (height // self.nbhd_height, width // self.nbhd_width)).to(raw_subframes.device) # shape: (n_pixels, height, width)
# scheme_ = scheme_.unsqueeze(1).repeat(1, S // n_pixels, 1, 1).view(S, height, width) # shape: (S, height, width)
c2b_frame_bucket0 = torch.minimum(torch.sum(subframes * self.scheme_.unsqueeze(1), dim=1), self.max_intensity.to(self.device))
c2b_frame_bucket1 = torch.minimum(torch.sum(subframes * (1 - self.scheme_.unsqueeze(1)), dim=1), self.max_intensity.to(self.device))
if self.noise_std is not None:
c2b_frame_bucket0 += torch.randn_like(c2b_frame_bucket0) * self.noise_std
c2b_frame_bucket1 += torch.randn_like(c2b_frame_bucket1) * self.noise_std
return c2b_frame_bucket0, c2b_frame_bucket1
def non_maximum_suppress(confidences, coordinates, over_thres=.4):
""" Params
TODO: wrapping the non_maximum_suppress and channel nms as an integral module of the neural network
TODO: Considering is it possible to archive different overlap threshold on rbc cells and platelet.
------
confidences: {Tensor: (n_posi_m,)},
coordinates: {Tensor: (n_posi_m, x-y-w-h)},
class_scores: {Tensor: (n_posi_m, n_class)},
over_thresh: float
"""
x1, y1, x2, y2 = coordinates.permute(1, 0)
areas = (x2 - x1) * (y2 - y1)
order = confidences.argsort()
pick = []
while len(order):
pick.append(idx := order[-1])
xx1 = torch.maximum(x1[order[:-1]], x1[idx])
yy1 = torch.maximum(y1[order[:-1]], y1[idx])
xx2 = torch.minimum(x2[order[:-1]], x2[idx])
yy2 = torch.minimum(y2[order[:-1]], y2[idx])
# inter: w * h
inter = torch.clamp(xx2 - xx1, min=0.) * torch.clamp(yy2 - yy1, min=0.)
over = inter / (areas[idx] + areas[order[:-1]] - inter)
order = order[:-1][over < over_thres]
return pick
def compute_crop_pad_image_location(
bbox_tight: "torch.Tensor", image: "torch.Tensor"
) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor",
"torch.Tensor"]:
"""
Get the valid image coordinates for the context region in target or search region in full image
:param bbox_tight: Coordinates of bounding box [x1, y1, x2, y2].
:param image: Frame to be cropped and padded.
:return: x-coordinate of the bounding box center.
"""
# Center of the bounding box
# bbox_center_x = bbox_tight.get_center_x()
# bbox_center_y = bbox_tight.get_center_y()
bbox_center_x = get_center_x_f(bbox_tight)
bbox_center_y = get_center_y_f(bbox_tight)
image_height = image.shape[0]
image_width = image.shape[1]
# Padded output width and height
# output_width = bbox_tight.compute_output_width()
# output_height = bbox_tight.compute_output_height()
output_width = compute_output_width_f(bbox_tight)
output_height = compute_output_height_f(bbox_tight)
roi_left = torch.maximum(
torch.tensor(0.0).to(self.device),
bbox_center_x - (output_width / 2.0))
roi_bottom = torch.maximum(
torch.tensor(0.0).to(self.device),
bbox_center_y - (output_height / 2.0))
# New ROI width
# -------------
# 1. left_half should not go out of bound on the left side of the
# image
# 2. right_half should not go out of bound on the right side of the
# image
left_half = torch.minimum(output_width / 2.0, bbox_center_x)
right_half = torch.minimum(output_width / 2.0,
image_width - bbox_center_x)
roi_width = torch.maximum(
torch.tensor(1.0).to(self.device), left_half + right_half)
# New ROI height
# Similar logic applied that is applied for 'New ROI width'
top_half = torch.minimum(output_height / 2.0, bbox_center_y)
bottom_half = torch.minimum(output_height / 2.0,
image_height - bbox_center_y)
roi_height = torch.maximum(
torch.tensor(1.0).to(self.device), top_half + bottom_half)
# Padded image location in the original image
# objPadImageLocation = BoundingBox(roi_left, roi_bottom, roi_left + roi_width, roi_bottom + roi_height)
#
# return objPadImageLocation
return roi_left, roi_bottom, roi_left + roi_width, roi_bottom + roi_height
def forward(self, x):
start, end = self._logits(x)
_, start_idx = start.max(dim=1)
_, end_idx = end.max(dim=1)
_, ctxs_len = (x[:, 0] == self.info.sep_token).max(dim=1)
start_idx = torch.minimum(start_idx, ctxs_len)
end_idx = torch.minimum(end_idx, ctxs_len)
return start_idx, end_idx
def limiter(cr):
return torch.maximum(
torch.tensor([0.0], device=cr.device),
torch.maximum(
torch.minimum(torch.tensor([1.0], device=cr.device), 2 * cr),
torch.minimum(torch.tensor([2.0], device=cr.device), cr),
),
)
def layout_bbox(self, final_pred, batch_size, num_bboxes, num_classes,
output_height, output_width):
# 5, 188, 20
final_pred = torch.reshape(final_pred,
[batch_size, num_bboxes, 4 + num_classes])
#print('Final pred:',final_pred.shape)
return self.rectangle_render(final_pred)
0 / 0
final_pred = torch.reshape(final_pred,
[batch_size, 4 + num_classes, num_bboxes])
print('Final pred requires grad:', final_pred.requires_grad)
bbox_reg = final_pred[:, :4, :]
cls_prob = final_pred[:, 4:, :]
print('bbox requires grad:', bbox_reg.requires_grad)
bbox_reg = torch.reshape(bbox_reg, [batch_size, num_bboxes, 4])
x_c = bbox_reg[:, :, 0] * output_width
y_c = bbox_reg[:, :, 1] * output_height
w = bbox_reg[:, :, 2] * output_width
h = bbox_reg[:, :, 3] * output_height
x1 = x_c - 0.5 * w
x2 = x_c + 0.5 * w
y1 = y_c - 0.5 * h
y2 = y_c + 0.5 * h
xt = torch.reshape(
torch.range(start=0, end=output_width, dtype=torch.float32),
[1, 1, 1, -1])
xt = torch.reshape(
torch.tile(xt, [batch_size, num_bboxes, output_height, 1]),
[batch_size, num_bboxes, -1])
yt = torch.reshape(
torch.range(start=0, end=output_height, dtype=torch.float32),
[1, 1, 1, -1])
yt = torch.reshape(
torch.tile(yt, [batch_size, num_bboxes, 1, output_width]),
[batch_size, num_bboxes, -1])
x1_diff = torch.reshape(
xt - x1, [batch_size, num_bboxes, output_height, output_width, 1])
y1_diff = torch.reshape(
yt - y1, [batch_size, num_bboxes, output_height, output_width, 1])
x2_diff = torch.reshape(
x2 - xt, [batch_size, num_bboxes, output_height, output_width, 1])
y2_diff = torch.reshape(
y2 - yt, [batch_size, num_bboxes, output_height, output_width, 1])
x1_line = self.relu(1.0 - torch.abs(x1_diff)) * torch.minimum(
self.relu(y1_diff), 1.0) * torch.minimum(self.relu(y2_diff), 1.0)
print(x1_line.shape)
print(x1_line)
0 / 0
def hinge_adv_loss(real_fake_logits_real, real_fake_logits_fake, device):
"""
the hinge version of the adversarial loss
:param real_fake_logits_real: ``Tensor([1, 5, 5])``
:param real_fake_logits_fake: ``Tensor([1, 5, 5])``
:param device: torch device
:return: ``float``, discriminator loss
"""
threshold = torch.Tensor([0.0]).to(device)
real_loss = -1 * torch.mean(torch.minimum(threshold, -1 + real_fake_logits_real))
fake_loss = -1 * torch.mean(torch.minimum(threshold, -1 - real_fake_logits_fake))
return real_loss + fake_loss
def bbox_overlaps_ciou(bboxes1, bboxes2):
bboxes1 = convert_box(bboxes1)
bboxes2 = convert_box(bboxes2)
rows = bboxes1.shape[0]
cols = bboxes2.shape[0]
cious = torch.zeros((rows, cols))
if rows * cols == 0:
return cious
exchange = False
if bboxes1.shape[0] > bboxes2.shape[0]:
bboxes1, bboxes2 = bboxes2, bboxes1
cious = torch.zeros((cols, rows))
exchange = True
w1 = bboxes1[:, 2] - bboxes1[:, 0]
h1 = bboxes1[:, 3] - bboxes1[:, 1]
w2 = bboxes2[:, 2] - bboxes2[:, 0]
h2 = bboxes2[:, 3] - bboxes2[:, 1]
area1 = w1 * h1
area2 = w2 * h2
center_x1 = (bboxes1[:, 2] + bboxes1[:, 0]) / 2
center_y1 = (bboxes1[:, 3] + bboxes1[:, 1]) / 2
center_x2 = (bboxes2[:, 2] + bboxes2[:, 0]) / 2
center_y2 = (bboxes2[:, 3] + bboxes2[:, 1]) / 2
inter_max_xy = torch.minimum(bboxes1[:, 2:], bboxes2[:, 2:])
inter_min_xy = torch.maximum(bboxes1[:, :2], bboxes2[:, :2])
out_max_xy = torch.maximum(bboxes1[:, 2:], bboxes2[:, 2:])
out_min_xy = torch.minimum(bboxes1[:, :2], bboxes2[:, :2])
inter = torch.clamp((inter_max_xy - inter_min_xy), min=0)
inter_area = inter[:, 0] * inter[:, 1]
inter_diag = (center_x2 - center_x1)**2 + (center_y2 - center_y1)**2
outer = torch.clamp((out_max_xy - out_min_xy), min=0)
outer_diag = (outer[:, 0]**2) + (outer[:, 1]**2)
union = area1 + area2 - inter_area
u = (inter_diag) / outer_diag
iou = inter_area / union
with torch.no_grad():
arctan = torch.atan(w2 / h2) - torch.atan(w1 / h1)
v = (4 / (math.pi**2)) * torch.pow(
(torch.atan(w2 / h2) - torch.atan(w1 / h1)), 2)
S = 1 - iou
alpha = v / (S + v)
w_temp = 2 * w1
ar = (8 / (math.pi**2)) * arctan * ((w1 - w_temp) * h1)
cious = iou - (u + alpha * ar)
cious = torch.clamp(cious, min=-1.0, max=1.0)
if exchange:
cious = cious.T
return cious
def forward(self, discriminator_prediction_real: torch.Tensor,
discriminator_prediction_fake: torch.Tensor,
**kwargs) -> torch.Tensor:
"""
Forward pass.
:param discriminator_prediction_real: (torch.Tensor) Raw discriminator prediction for real samples
:param discriminator_prediction_fake: (torch.Tensor) Raw discriminator predictions for fake samples
:return: (torch.Tensor) Hinge discriminator GAN loss
"""
return - torch.minimum(torch.tensor(0., dtype=torch.float, device=discriminator_prediction_real.device),
discriminator_prediction_real - 1.).mean() \
- torch.minimum(torch.tensor(0., dtype=torch.float, device=discriminator_prediction_fake.device),
- discriminator_prediction_fake - 1.).mean()
def forward(ctx, X, rank: int = 100):
U, S, V = torch.svd(X, compute_uv=True, some=False)
S = torch.diag(S[0:(rank - 1)])
U = torch.matmul(U[:, 0:(rank - 1)], S)
V = torch.transpose(V, 0, 1)[0:(rank - 1), :]
x, y = X.shape
Unew = U[:, 0]
Vnew = V[0, :]
__U = torch.where(torch.less(torch.min(V[0, :]), torch.min(-V[0, :])),
-(Unew.view(x, 1)), Unew.view(x, 1))
__V = torch.where(torch.less(torch.min(V[0, :]), torch.min(-V[0, :])),
-(Vnew.view(1, y)), Vnew.view(1, y))
if rank > 2:
for i in range(1, rank - 1):
Unew = Unew.view(x, 1)
Vnew = Vnew.view(1, y)
__U = torch.where(
torch.less(torch.min(V[0, :]), torch.min(-V[0, :])),
torch.cat((__U, -Unew), dim=1),
torch.cat((__U, Unew), dim=1))
__V = torch.where(
torch.less(torch.min(V[0, :]), torch.min(-V[0, :])),
torch.cat((__V, -Vnew), dim=0),
torch.cat((__V, Vnew), dim=0))
if rank == 2:
A = torch.cat((U, -U), dim=1)
else:
Un = torch.transpose(-(torch.sum(U, dim=1)), 0, -1).view(x, 1)
A = torch.cat((U, Un), dim=1)
B = torch.cat((V, torch.zeros((1, y))), dim=0)
if rank >= 3:
b, _ = torch.min(V, dim=0)
B = torch.subtract(B, torch.minimum(torch.tensor(0.), b))
else:
B = torch.subtract(B, torch.minimum(torch.tensor(0.), V))
x = torch.tensor(x)
y = torch.tensor(y)
normalize = torch.sqrt(torch.multiply(x, y).type(torch.FloatTensor))
norm = torch.norm(A)
return torch.multiply(torch.div(A, norm),
normalize), torch.div(torch.multiply(B, norm),
normalize)
def _train(self, BATCH):
q1 = self.critic(BATCH.obs, begin_mask=BATCH.begin_mask) # [T, B, A]
q2 = self.critic2(BATCH.obs, begin_mask=BATCH.begin_mask) # [T, B, A]
q1_eval = (q1 * BATCH.action).sum(-1, keepdim=True) # [T, B, 1]
q2_eval = (q2 * BATCH.action).sum(-1, keepdim=True) # [T, B, 1]
q1_log_probs = (q1 / (self.alpha + th.finfo().eps)).log_softmax(-1) # [T, B, A]
q1_entropy = -(q1_log_probs.exp() * q1_log_probs).sum(-1, keepdim=True).mean() # 1
q1_target = self.critic.t(BATCH.obs_, begin_mask=BATCH.begin_mask) # [T, B, A]
q2_target = self.critic2.t(BATCH.obs_, begin_mask=BATCH.begin_mask) # [T, B, A]
q1_target_max = q1_target.max(-1, keepdim=True)[0] # [T, B, 1]
q1_target_log_probs = (q1_target / (self.alpha + th.finfo().eps)).log_softmax(-1) # [T, B, A]
q1_target_entropy = -(q1_target_log_probs.exp() * q1_target_log_probs).sum(-1, keepdim=True) # [T, B, 1]
q2_target_max = q2_target.max(-1, keepdim=True)[0] # [T, B, 1]
# q2_target_log_probs = q2_target.log_softmax(-1)
# q2_target_log_max = q2_target_log_probs.max(1, keepdim=True)[0]
q_target = th.minimum(q1_target_max, q2_target_max) + self.alpha * q1_target_entropy # [T, B, 1]
dc_r = n_step_return(BATCH.reward,
self.gamma,
BATCH.done,
q_target,
BATCH.begin_mask).detach() # [T, B, 1]
td_error1 = q1_eval - dc_r # [T, B, 1]
td_error2 = q2_eval - dc_r # [T, B, 1]
q1_loss = (td_error1.square() * BATCH.get('isw', 1.0)).mean() # 1
q2_loss = (td_error2.square() * BATCH.get('isw', 1.0)).mean() # 1
loss = 0.5 * (q1_loss + q2_loss)
self.critic_oplr.optimize(loss)
summaries = {
'LEARNING_RATE/critic_lr': self.critic_oplr.lr,
'LOSS/loss': loss,
'Statistics/log_alpha': self.log_alpha,
'Statistics/alpha': self.alpha,
'Statistics/q1_entropy': q1_entropy,
'Statistics/q_min': th.minimum(q1, q2).mean(),
'Statistics/q_mean': q1.mean(),
'Statistics/q_max': th.maximum(q1, q2).mean()
}
if self.auto_adaption:
alpha_loss = -(self.alpha * (self.target_entropy - q1_entropy).detach()).mean()
self.alpha_oplr.optimize(alpha_loss)
summaries.update({
'LOSS/alpha_loss': alpha_loss,
'LEARNING_RATE/alpha_lr': self.alpha_oplr.lr
})
return (td_error1 + td_error2) / 2, summaries
def calculate_iou(best_pred, preds, areas):
# Remove bboxes with IOU >= Thresh
max_min_x = torch.maximum(best_pred[1], preds[1:, 1])
max_min_y = torch.maximum(best_pred[2], preds[1:, 2])
min_max_x = torch.minimum(best_pred[3], preds[1:, 3])
min_max_y = torch.minimum(best_pred[4], preds[1:, 4])
intersection_x = min_max_x - max_min_x
intersection_y = min_max_y - max_min_y
intersection_area = intersection_x * intersection_y
iou = intersection_area / (areas[0] + areas[1:] -
intersection_area)
return iou
def reduce_relu(self, nodes):
w = torch.exp(self.w)
R = torch.clamp(self.R, 0.000001, 0.999999)
msg = w * nodes.mailbox['m'] + self.b
fsum = torch.sum(torch.maximum(msg, R * msg), dim=1)
out_h = (torch.minimum(fsum, fsum / R) - self.b) / w
return {'sum_sigma_h': out_h}
def comparison_ops(self):
a = torch.randn(4)
b = torch.randn(4)
return (
torch.allclose(a, b),
torch.argsort(a),
torch.eq(a, b),
torch.equal(a, b),
torch.ge(a, b),
torch.greater_equal(a, b),
torch.gt(a, b),
torch.greater(a, b),
torch.isclose(a, b),
torch.isfinite(a),
torch.isin(a, b),
torch.isinf(a),
torch.isposinf(a),
torch.isneginf(a),
torch.isnan(a),
torch.isreal(a),
torch.kthvalue(a, 1),
torch.le(a, b),
torch.less_equal(a, b),
torch.lt(a, b),
torch.less(a, b),
torch.maximum(a, b),
torch.minimum(a, b),
torch.fmax(a, b),
torch.fmin(a, b),
torch.ne(a, b),
torch.not_equal(a, b),
torch.sort(a),
torch.topk(a, 1),
torch.msort(a),
)
def orthogonalized_raised_cosines(cls,
dt,
last_time_peak,
n,
b,
a=1e0,
weight=None):
range_locs = torch.log(torch.tensor([0, last_time_peak]) + b)
delta = (range_locs[1] - range_locs[0]) / (n - 1)
locs = torch.linspace(range_locs[0], range_locs[1], n)
last_time = torch.exp(range_locs[1] + 2 * delta / a) - b
t = torch.arange(0, last_time, dt)
support = torch.tensor([t[0], t[-1] + dt])
pi_torch = torch.tensor([pi])
raised_cosines = torch.minimum(
a * (torch.log(t[:, None] + b) - locs[None, :]) * pi / delta / 2,
pi_torch)
raised_cosines = (
1 + torch.cos(torch.maximum(-pi_torch, raised_cosines))) / 2
raised_cosines = raised_cosines / torch.sqrt(
torch.sum(raised_cosines**2, 0))
u, s, v = torch.linalg.svd(raised_cosines)
basis = u[:, :n]
return cls(basis=basis, support=support, weight=weight)
def _get_slate_size(self, state: rlt.FeatureData) -> torch.Tensor:
"""Get the actual size (ignore all padded items) of each slate by summing item masks."""
mask = self._get_item_mask(state)
return torch.minimum(
mask.sum(1, keepdim=True),
torch.tensor([self.slate_size], device=mask.device),
)
def step(self, closure=None):
"""Performs a single optimization step.
Args:
closure (callable, optional): A closure that reevaluates the model and returns the loss.
"""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
device = self.param_groups[0]['params'][0].device
one_tensor = torch.tensor(1.0, device=device) # because torch.where doesn't handle scalars correctly
for group in self.param_groups:
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
trust_coeff = group['trust_coeff']
eps = group['eps']
for p in group['params']:
if p.grad is None:
continue
grad = p.grad
# apply LARS LR adaptation, LARC clipping, weight decay
# ref: https://github.com/NVIDIA/apex/blob/master/apex/parallel/LARC.py
if weight_decay != 0 or group['always_adapt']:
w_norm = p.norm(2.0)
g_norm = grad.norm(2.0)
trust_ratio = trust_coeff * w_norm / (g_norm + w_norm * weight_decay + eps)
# FIXME nested where required since logical and/or not working in PT XLA
trust_ratio = torch.where(
w_norm > 0,
torch.where(g_norm > 0, trust_ratio, one_tensor),
one_tensor,
)
if group['trust_clip']:
trust_ratio = torch.minimum(trust_ratio / group['lr'], one_tensor)
grad.add(p, alpha=weight_decay)
grad.mul_(trust_ratio)
# apply SGD update https://github.com/pytorch/pytorch/blob/1.7/torch/optim/sgd.py#L100
if momentum != 0:
param_state = self.state[p]
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = torch.clone(grad).detach()
else:
buf = param_state['momentum_buffer']
buf.mul_(momentum).add_(grad, alpha=1. - dampening)
if nesterov:
grad = grad.add(buf, alpha=momentum)
else:
grad = buf
p.add_(grad, alpha=-group['lr'])
return loss
def boundaries_detection(self, inputs):
y_fwd, y_bwd, seq_len_y, *_ = self.forward(inputs)
seq_mask = compute_mask(y_fwd,
seq_len_y,
batch_axis=0,
sequence_axis=-1)
return torch.minimum(y_fwd * seq_mask, y_bwd * seq_mask), seq_len_y
def cdf_tail(self,
z: torch.Tensor,
left_tail: bool = True) -> torch.Tensor:
r"""
Computes the quantile level alpha_tilde such that
alpha_tilde
= q^{-1}(z) if z is in the tail region
= qk_x_l or qk_x_r if z is in the non-tail region
Parameters
----------
z
Observation, shape = (*batch_shape,)
left_tail
If True, compute alpha_tilde for the left tail
Otherwise, compute alpha_tilde for the right tail
Returns
-------
alpha_tilde
Corresponding quantile level, shape = (*batch_shape,)
"""
if left_tail:
tail_a, tail_b, qk_x = self.tail_al, self.tail_bl, self.qk_x_l
else:
tail_a, tail_b, qk_x = self.tail_ar, self.tail_br, 1 - self.qk_x_r
log_alpha_tilde = torch.minimum((z - tail_b) / tail_a, torch.log(qk_x))
alpha_tilde = torch.exp(log_alpha_tilde)
return alpha_tilde if left_tail else 1 - alpha_tilde
def quantile_spline(
self,
alpha: torch.Tensor,
dim: Optional[int] = None,
) -> torch.Tensor:
# Refer to the description in quantile_internal
qk_y = self.qk_y
sk_x, delta_sk_x, delta_sk_y = (
self.sk_x,
self.delta_sk_x,
self.delta_sk_y,
)
if dim is not None:
qk_y = qk_y.unsqueeze(dim=0 if dim == 0 else -1)
sk_x = sk_x.unsqueeze(dim=dim)
delta_sk_x = delta_sk_x.unsqueeze(dim=dim)
delta_sk_y = delta_sk_y.unsqueeze(dim=dim)
if dim is None or dim == 0:
alpha = alpha.unsqueeze(dim=-1)
alpha = alpha.unsqueeze(dim=-1)
spline_val = (alpha - sk_x) / delta_sk_x
spline_val = torch.maximum(
torch.minimum(spline_val, torch.ones_like(spline_val)),
torch.zeros_like(spline_val),
)
return qk_y + torch.sum(spline_val * delta_sk_y, dim=-1)
def streaming_perception(self,
d,
cd=None,
ece=None,
ce=None,
x=None,
rgb=None,
**args):
if self.training:
w_pow_d, w_pow_e = self.prepare_weights()
else:
w_pow_d, w_pow_e = self.weights
if cd is None:
cd = (d > 0).float()
outs1 = self.inner_model.streaming_perception(d, cd, ece, ce, x, rgb,
**args)
mng = plt.get_current_fig_manager()
mng.window.state('zoomed')
plt.tight_layout()
cd = cd * torch.pow(
torch.minimum(outs1['d'], d) / (d + self.eps), w_pow_d)
e1 = outs1['e']
ce1 = torch.pow(outs1['ce'], w_pow_e)
return self.inner_model.streaming_perception(d, cd, e1 * ce1, ce1, x,
rgb, **args)
def _policy_update(self, advantage):
for epoch in range(self._iteration_op_policy):
loss = torch.zeros(1)
len_data = 0
for n, traj in enumerate(self._data_buffer):
len_data += len(traj)
# per-batch normalization
advantage_ = (advantage[n, :] - torch.mean(
advantage[n, :])) / torch.std(advantage[n, :])
for t, experience in enumerate(traj):
log_p = self._policy_network.log_prob(
Tensor(experience.observation),
Tensor(experience.action))
log_p_old = self._policy_old.log_prob(
Tensor(experience.observation),
Tensor(experience.action)).detach()
ratio = torch.exp(log_p - log_p_old)
tmp = torch.minimum(
ratio,
torch.clip(ratio,
max=1 + self._eps_policy_clip,
min=1 - self._eps_policy_clip))
loss += -tmp * advantage_[t]
loss /= len_data
print(
f"policy loss {epoch + 1}/{self._iteration_op_policy}: {len_data * loss.detach()}"
)
self._optimizer_policy.zero_grad()
loss.backward()
self._optimizer_policy.step()
# copy old policy
self._policy_old.load_state_dict(self._policy_network.state_dict())
def sample_cdf(z_vals, weights, det=False):
bins = 0.5 * (z_vals[..., 1:] + z_vals[..., :-1])
weights = weights + 1e-5
pdf = weights / torch.sum(weights, -1, keepdim=True)
cdf = torch.cumsum(pdf, -1)
cdf = torch.cat([torch.zeros_like(cdf[..., :1]), cdf], -1)
if det:
u = torch.linspace(0., 1., config.N_samples)
u = torch.broadcast_to(u, list(cdf.shape[:-1]) + [config.N_samples])
else:
u = torch.rand(cdf.shape[0], config.N_samples)
u = u.cuda(cdf.device)
idxs = torch.searchsorted(cdf, u, right=True)
below = torch.maximum(torch.zeros_like(idxs), idxs - 1).long()
above = torch.minimum(torch.ones_like(idxs) * (cdf.shape[-1] - 1),
idxs).long()
# idxs_g = torch.stack([below, above], -1)
cdf_below = torch.gather(cdf, dim=1, index=below)
cdf_above = torch.gather(cdf, dim=1, index=above)
bin_below = torch.gather(bins, dim=1, index=below)
bin_above = torch.gather(bins, dim=1, index=above)
denom = cdf_above - cdf_below
denom = torch.clamp(denom, 1e-5, 99999)
t = (u - cdf_below) / denom
samples = bin_below + t * (bin_above - bin_below)
return samples
def forward(self, x, *args, **kwargs):
b, device = x.shape[0], x.device
D = np.cumprod(x.shape[1:])[-1]
shape = x.shape
t = torch.randint(1, self.T - self.tao, (b, ), device=device).long()
# print(t.device, self.sqrt_cumprod_alpha.device, self.betas.device)
alpha_n = extract(self.sqrt_cumprod_alpha, t, shape)
beta_np1 = 1. - (
extract(self.sqrt_cumprod_alpha, t + self.tao, shape) / alpha_n)**2
delta_n = torch.sqrt(1. - alpha_n**2)
noise = torch.randn_like(x)
xn = alpha_n * x + delta_n * noise
denoise_par = self.get_denoise_par(alpha_n, *args, **kwargs)
epsilon_theta = self.denoise_fn(xn, *denoise_par)
beta_n = torch.minimum(delta_n**2, beta_np1) * reshape(
self.sigma_phi(xn), delta_n.shape)
Cn = 0.25 * torch.log(
delta_n**2 / beta_n) + 0.5 * (beta_n / delta_n**2 - 1.)
# print(delta_n.shape, beta_n.shape, noise.shape, epsilon_theta.shape, beta_np1.shape, self.sigma_phi(xn).shape)
term1 = 0.5 / (delta_n**2 - beta_n) * (
delta_n * noise - beta_n / delta_n * epsilon_theta)**2
# print(Cn.shape, term1.shape)
loss1 = torch.mean(Cn.squeeze())
# loss2 = torch.mean(torch.sum(term1, dim=[-i for i in range(1, len(shape))]))
loss2 = torch.mean(term1)
# return torch.mean(Cn.squeeze() + torch.sum(term1, dim=[-i for i in range(1, len(shape))]))
return loss1, loss2