def extract_multi_position_matrix_nd(bbox):
bbox = nd.transpose(bbox, axes=(1, 0, 2))
xmin, ymin, xmax, ymax = nd.split(data=bbox, num_outputs=4, axis=2)
# [num_fg_classes, num_boxes, 1]
bbox_width = xmax - xmin + 1.
bbox_height = ymax - ymin + 1.
center_x = 0.5 * (xmin + xmax)
center_y = 0.5 * (ymin + ymax)
# [num_fg_classes, num_boxes, num_boxes]
delta_x = nd.broadcast_minus(lhs=center_x,
rhs=nd.transpose(center_x, axes=(0, 2, 1)))
delta_x = nd.broadcast_div(delta_x, bbox_width)
delta_x = nd.log(nd.maximum(nd.abs(delta_x), 1e-3))
delta_y = nd.broadcast_minus(lhs=center_y,
rhs=nd.transpose(center_y, axes=(0, 2, 1)))
delta_y = nd.broadcast_div(delta_y, bbox_height)
delta_y = nd.log(nd.maximum(nd.abs(delta_y), 1e-3))
delta_width = nd.broadcast_div(lhs=bbox_width,
rhs=nd.transpose(bbox_width,
axes=(0, 2, 1)))
delta_width = nd.log(delta_width)
delta_height = nd.broadcast_div(lhs=bbox_height,
rhs=nd.transpose(bbox_height,
axes=(0, 2, 1)))
delta_height = nd.log(delta_height)
concat_list = [delta_x, delta_y, delta_width, delta_height]
for idx, sym in enumerate(concat_list):
concat_list[idx] = nd.expand_dims(sym, axis=3)
position_matrix = nd.concat(*concat_list, dim=3)
return position_matrix
def norm(tensor, order=2, axis=None):
"""Computes the l-`order` norm of tensor
Parameters
----------
tensor : ndarray
order : int
axis : int or tuple
Returns
-------
float or tensor
If `axis` is provided returns a tensor.
"""
# handle difference in default axis notation
if axis is None:
axis = ()
if order == 'inf':
res = nd.max(nd.abs(tensor), axis=axis)
elif order == 1:
res = nd.sum(nd.abs(tensor), axis=axis)
elif order == 2:
res = nd.sqrt(nd.sum(tensor**2, axis=axis))
else:
res = nd.sum(nd.abs(tensor)**order, axis=axis)**(1/order)
if res.shape == (1,):
return res.asscalar()
return res
def quantize(self, x):
max = nd.max(nd.abs(x))
if max != 0:
int_len = (nd.ceil(nd.log2(nd.max(nd.abs(x))))).astype('float32')
num_bit = self.num_bit.as_in_context(x.context)
frac_len = num_bit - int_len
f = (2**(frac_len)).astype('float32')
y = ((x * f)).floor() * (1 / f)
return y
return x
def int_inference(self, x):
max = nd.max(nd.abs(x))
if max != 0:
int_len = (nd.ceil(nd.log2(nd.max(nd.abs(x))))).astype('float32')
num_bit = self.num_bit.as_in_context(x.context)
frac_len = num_bit - int_len
f = (2**(frac_len)).astype('float32')
y = ((x * f)).round()
return y.astype('int8'), frac_len
return x.astype('int8'), 0
def int_quantize(self, x):
max = nd.max(nd.abs(x))
if max != 0:
int_len = (nd.ceil(nd.log2(nd.max(nd.abs(x))))).astype('float32')
num_bit = self.num_bit.as_in_context(x.context)
frac_len = num_bit - int_len
f = (2**(frac_len)).astype('float32')
y = ((x * f)).floor()
y = nd.clip(y, a_min=-128, a_max=127)
return y, frac_len
return x, 0
def update(self, data): pred = self.scaler.inverse_transform(data[self.pred_name]) label = self.scaler.inverse_transform(data[self.label_name]) mask = label >= 10 _cnt = nd.sum(mask).as_in_context(mx.cpu()) _loss = nd.sum(nd.abs(pred - label) / (nd.abs(pred) + nd.abs(label) + 1e-5) * mask).as_in_context(mx.cpu()) if self.cnt is None: self.cnt = self.loss = 0 self.cnt += _cnt self.loss += _loss
def forward(self, is_train, req, in_data, out_data, aux):
x = in_data[0]
y = out_data[0]
in_max = nd.max(nd.abs(x))
if in_max != 0:
int_len = (nd.ceil(nd.log2(nd.max(nd.abs(x))))).astype('float32')
num_bit = self.num_bit.as_in_context(x.context)
frac_len = num_bit - int_len
f = (2**(frac_len)).astype('float32')
y = ((x * f)).round() * (1 / f)
y = x
self.assign(out_data[0], req[0], mx.nd.array(y))
def bbox_iou(box1, box2, transform=True):
"""Calculate the IoU Error
"""
#Change to NDArray if not
if not isinstance(box1, nd.NDArray):
box1 = nd.array(box1)
if not isinstance(box2, nd.NDArray):
box2 = nd.array(box2)
#Make sure > 0
box1 = nd.abs(box1)
box2 = nd.abs(box2)
'''Calculate the IoU'''
if transform:
tmp_box1 = box1.copy()
tmp_box1[:, 0] = box1[:, 0] - box1[:, 2] / 2.0
tmp_box1[:, 1] = box1[:, 1] - box1[:, 3] / 2.0
tmp_box1[:, 2] = box1[:, 0] + box1[:, 2] / 2.0
tmp_box1[:, 3] = box1[:, 1] + box1[:, 3] / 2.0
box1 = tmp_box1
tmp_box2 = box2.copy()
tmp_box2[:, 0] = box2[:, 0] - box2[:, 2] / 2.0
tmp_box2[:, 1] = box2[:, 1] - box2[:, 3] / 2.0
tmp_box2[:, 2] = box2[:, 0] + box2[:, 2] / 2.0
tmp_box2[:, 3] = box2[:, 1] + box2[:, 3] / 2.0
box2 = tmp_box2
# Get the coordinates of bounding boxes (xStart,yStart,xEnd,yEnd)
b1_x1, b1_y1, b1_x2, b1_y2 = box1[:, 0], box1[:, 1], box1[:, 2], box1[:, 3]
b2_x1, b2_y1, b2_x2, b2_y2 = box2[:, 0], box2[:, 1], box2[:, 2], box2[:, 3]
# get the corrdinates of the intersection rectangle
inter_rect_x1 = nd.where(
b1_x1 > b2_x1, b1_x1, b2_x1
) #if b1_x1 > b2_x1 => x1 of the intersection rectangle must be b1_x1, otherwise it will be b2_x1. Basically it's just a max function!
inter_rect_y1 = nd.where(b1_y1 > b2_y1, b1_y1, b2_y1)
inter_rect_x2 = nd.where(b1_x2 < b2_x2, b1_x2, b2_x2)
inter_rect_y2 = nd.where(b1_y2 < b2_y2, b1_y2, b2_y2)
# Intersection area
inter_area = nd.clip(
inter_rect_x2 - inter_rect_x1 + 1, a_min=0, a_max=10000) * nd.clip(
inter_rect_y2 - inter_rect_y1 + 1, a_min=0, a_max=10000)
# Union Area
b1_area = (b1_x2 - b1_x1 + 1) * (b1_y2 - b1_y1 + 1)
b2_area = (b2_x2 - b2_x1 + 1) * (b2_y2 - b2_y1 + 1)
iou = inter_area / (b1_area + b2_area - inter_area)
return nd.clip(iou, 1e-5, 1. - 1e-5)
def test_513():
# 边缘检测小例
X = nd.ones((8, 8))
X[2:6, 2:6] = 0
print(X)
'''
原始图片，任务是进行边缘检测
[
[1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 0. 0. 0. 0. 1. 1.]
[1. 1. 0. 0. 0. 0. 1. 1.]
[1. 1. 0. 0. 0. 0. 1. 1.]
[1. 1. 0. 0. 0. 0. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1.]
]
'''
# 构造kernel
K1 = nd.array([[1, -1], [1, -1]])
Y1 = corr2d(X, K1)
print(Y1)
'''
检测纵向边缘
[[ 0. 0. 0. 0. 0. 0. 0.]
[ 0. 1. 0. 0. 0. -1. 0.]
[ 0. 2. 0. 0. 0. -2. 0.]
[ 0. 2. 0. 0. 0. -2. 0.]
[ 0. 2. 0. 0. 0. -2. 0.]
[ 0. 1. 0. 0. 0. -1. 0.]
[ 0. 0. 0. 0. 0. 0. 0.]]
'''
# 构造kernel
K2 = nd.array([[1, 1], [-1, -1]])
Y2 = corr2d(X, K2)
print(Y2)
'''
检测横向边缘
[[ 0. 0. 0. 0. 0. 0. 0.]
[ 0. 1. 2. 2. 2. 1. 0.]
[ 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0.]
[ 0. -1. -2. -2. -2. -1. 0.]
[ 0. 0. 0. 0. 0. 0. 0.]]
'''
Y3 = nd.add(nd.abs(Y1), nd.abs(Y2))
print(Y3)
'''
def bbox_iou(box1, box2, transform=True, ctx=None):
'''
判断预测盒子和实际盒子的重合度。>0.5是比较好的预测
'''
ctx = ctx
if not isinstance(box1, nd.NDArray):
box1 = nd.array(box1, ctx=ctx)
if not isinstance(box2, nd.NDArray):
box2 = nd.array(box2, ctx=ctx)
box1 = nd.abs(box1)
box2 = nd.abs(box2)
if transform:
tmp_box1 = box1.copy()
tmp_box1[:, 0] = box1[:, 0] - box1[:, 2] / 2.0
tmp_box1[:, 1] = box1[:, 1] - box1[:, 3] / 2.0
tmp_box1[:, 2] = box1[:, 0] + box1[:, 2] / 2.0
tmp_box1[:, 3] = box1[:, 1] + box1[:, 3] / 2.0
box1 = tmp_box1
tmp_box2 = box2.copy()
tmp_box2[:, 0] = box2[:, 0] - box2[:, 2] / 2.0
tmp_box2[:, 1] = box2[:, 1] - box2[:, 3] / 2.0
tmp_box2[:, 2] = box2[:, 0] + box2[:, 2] / 2.0
tmp_box2[:, 3] = box2[:, 1] + box2[:, 3] / 2.0
box2 = tmp_box2
# Get the coordinates of bounding boxes
b1_x1, b1_y1, b1_x2, b1_y2 = box1[:, 0], box1[:, 1], box1[:, 2], box1[:, 3]
b2_x1, b2_y1, b2_x2, b2_y2 = box2[:, 0], box2[:, 1], box2[:, 2], box2[:, 3]
# get the corrdinates of the intersection rectangle
inter_rect_x1 = nd.where(b1_x1 > b2_x1, b1_x1, b2_x1)
inter_rect_y1 = nd.where(b1_y1 > b2_y1, b1_y1, b2_y1)
inter_rect_x2 = nd.where(b1_x2 < b2_x2, b1_x2, b2_x2)
inter_rect_y2 = nd.where(b1_y2 < b2_y2, b1_y2, b2_y2)
# Intersection area
inter_area = nd.clip(
inter_rect_x2 - inter_rect_x1 + 1, a_min=0, a_max=10000) * nd.clip(
inter_rect_y2 - inter_rect_y1 + 1, a_min=0, a_max=10000)
# Union Area
b1_area = (b1_x2 - b1_x1 + 1) * (b1_y2 - b1_y1 + 1)
b2_area = (b2_x2 - b2_x1 + 1) * (b2_y2 - b2_y1 + 1)
iou = inter_area / (b1_area + b2_area - inter_area)
# iou[inter_area >= b1_area] = 0.8
# iou[inter_area >= b2_area] = 0.8
return nd.clip(iou, 1e-5, 1. - 1e-5)
def forward(self, output, mask, ind, target):
pred = _tranpose_and_gather_feat(output, ind)
pred = pred.swapaxes(dim1=0, dim2=1)
mask = mask.astype('float32')
loss = nd.abs(pred * mask - target * mask).sum()
loss = loss / (mask.sum() + 1e-4)
return loss
def accuracy(predictions, targets):
# predictions = nd.argmax(predictions, 1)
# targets = nd.argmax(targets, 1)
# return nd.mean(nd.equal(predictions, targets)).asscalar() * 100
predictions = nd.where(predictions > 0.5, nd.ones_like(predictions),
nd.zeros_like(predictions))
return 100 - nd.mean(nd.abs(predictions - targets)).asscalar() * 100
def acc(output, label):
min_acc = 1
for i in range(output.shape[0]):
x = nd.sum(nd.abs(nd.subtract(output[i], label[i])), axis=0)
if 1 - x / 96 < min_acc:
min_acc = 1 - x / 96
return min_acc.asscalar()
def add_split(x, leaf, p_tau):
center = leaf.parent['node'].center.data()
radius = leaf.parent['node'].radius.data()
tau = p_tau + nd.random.exponential(radius**-1)
while 1:
s = nd.random.normal(shape=(2, x.shape[-1]))
s = s / nd.norm(s, axis=-1, keepdims=True)
r = nd.random.uniform(low=nd.array([0]), high=radius)
r = r * nd.random.uniform()**(1 / 3)
if nd.sign(s[0][-1]) > 0:
weight = s[0]
bias = nd.dot(s[0], -1 * r * (s[1] + center))
y = nd.sign(nd.dot(x, weight) + bias)
if nd.abs(nd.sum(y)) != len(y):
break
split = Split(weight=weight,
bias=bias,
sharpness=3 / radius,
tau=tau,
decision=leaf.parent['decision'],
side=leaf.parent['side'])
tree.splits.add(split)
leaf.parent['node'].child['decision'] = split
leaf.parent['decision'] = split
def int_quantize_double(self, x, w):
max1 = nd.max(nd.abs(x))
max2 = nd.max(nd.abs(w))
if max1 > max2:
max = max1
else:
max = max2
if max != 0:
int_len = (nd.ceil(nd.log2(max))).astype('float32')
num_bit = self.num_bit.as_in_context(x.context)
frac_len = num_bit - int_len
f = (2**(frac_len)).astype('float32')
int_x = ((x * f)).floor()
int_w = ((w * f)).floor()
return int_x, int_w, frac_len
return x, w, 0
def nd_std(x, axis=-1):
""" Standard Deviation (SD)
Note: Do not try 'axis=0'
"""
return nd.sqrt(
nd.square(nd.abs(x - x.mean(axis=axis).expand_dims(axis=axis))).mean(
axis=axis))
def calc_loss(pred, label):
# Mean average loss
softmax_pred = stable_softmax(pred) + NEAR_0
expectation = idx_tensor * softmax_pred
expectation = nd.sum(expectation, 1) * 3 - 102
loss = nd.abs(expectation - label)
loss = nd.sum(loss) / len(loss)
def compute(self, input):
"""Compute the output of the neural net given the input.
The input has to be a ndarray of shape input_size or (input_size, N) where N is the batch_size."""
if len(input.shape) == 1:
input = input.reshape((input.shape[0], 1))
X = input[0]
Y = input[1]
for i in range(self.n):
_X = nd.cos(self.thetas[i]) * X + nd.sin(
self.thetas[i]) * Y # Sym / (cos(O)u_x + sin(O)u_y)
Y = -nd.sin(self.thetas[i]) * X + nd.cos(self.thetas[i]) * Y
X = nd.abs(_X + self.biases[i]) - self.biases[i]
if (False): # optimize
_X = nd.cos(self.thetas[i]) * X - nd.sin(
self.thetas[i]) * Y # Sym / (cos(O)u_x + sin(O)u_y)
Y = nd.sin(self.thetas[i]) * X + nd.cos(self.thetas[i]) * Y
X = _X
_X = nd.cos(self.thetas[self.n]) * X + nd.sin(self.thetas[self.n]) * Y
Y = -nd.sin(self.thetas[self.n]) * X + nd.cos(self.thetas[self.n]) * Y
return nd.sigmoid(Y - _X)
def test(ctx=mx.cpu()):
from mxboard import SummaryWriter
sw = SummaryWriter(logdir='sphere_dynamic', flush_secs=5)
net = nn.Sequential()
b1 = base_net(48,
3,
fun=special_conv,
kernel_size=(3, 3),
same_shape=False)
b2 = base_net(1,
48,
fun=special_conv,
kernel_size=(3, 3),
same_shape=False)
fc = nn.Dense(3, in_units=9)
net.add(b1, b2, fc)
init_s(net, ctx)
from mxnet import gluon, autograd
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.01})
for i in range(10000):
with autograd.record():
out = net(img)
loss = nd.sum(nd.abs(out - target))
loss.backward()
trainer.step(2)
sw.add_scalar(tag='loss', value=loss.asscalar(), global_step=i)
if i % 100 == 0:
print i, loss.asscalar()
sw.close()
def bbox_iou(box1, box2, transform=True):
"""
Returns the IoU of two bounding boxes
"""
box1 = nd.array(box1)
box2 = nd.array(box2)
if box1.size == 0 or box2.size == 0:
raise ValueError
box1 = nd.abs(box1)
box2 = nd.abs(box2)
if transform:
tmp_box1 = box1.copy()
tmp_box1[:, 0] = box1[:, 0] - box1[:, 2] / 2.0
tmp_box1[:, 1] = box1[:, 1] - box1[:, 3] / 2.0
tmp_box1[:, 2] = box1[:, 0] + box1[:, 2] / 2.0
tmp_box1[:, 3] = box1[:, 1] + box1[:, 3] / 2.0
box1 = tmp_box1
tmp_box2 = box2.copy()
tmp_box2[:, 0] = box2[:, 0] - box2[:, 2] / 2.0
tmp_box2[:, 1] = box2[:, 1] - box2[:, 3] / 2.0
tmp_box2[:, 2] = box2[:, 0] + box2[:, 2] / 2.0
tmp_box2[:, 3] = box2[:, 1] + box2[:, 3] / 2.0
box2 = tmp_box2
# Get the coordinates of bounding boxes
b1_x1, b1_y1, b1_x2, b1_y2 = box1[:, 0], box1[:, 1], box1[:, 2], box1[:, 3]
b2_x1, b2_y1, b2_x2, b2_y2 = box2[:, 0], box2[:, 1], box2[:, 2], box2[:, 3]
# get the corrdinates of the intersection rectangle
inter_rect_x1 = nd.where(b1_x1 > b2_x1, b1_x1, b2_x1)
inter_rect_y1 = nd.where(b1_y1 > b2_y1, b1_y1, b2_y1)
inter_rect_x2 = nd.where(b1_x2 < b2_x2, b1_x2, b2_x2)
inter_rect_y2 = nd.where(b1_y2 < b2_y2, b1_y2, b2_y2)
# Intersection area
inter_area = nd.clip(
inter_rect_x2 - inter_rect_x1 + 1, a_min=0, a_max=10000) * nd.clip(
inter_rect_y2 - inter_rect_y1 + 1, a_min=0, a_max=10000)
# Union Area
b1_area = (b1_x2 - b1_x1 + 1) * (b1_y2 - b1_y1 + 1)
b2_area = (b2_x2 - b2_x1 + 1) * (b2_y2 - b2_y1 + 1)
iou = inter_area / (b1_area + b2_area - inter_area)
# iou[inter_area >= b1_area] = 0.8
# iou[inter_area >= b2_area] = 0.8
# iou[inter_area >= b2_area] = 0.8
return nd.clip(iou, 1e-5, 1. - 1e-5)
def custom_loss(pred, label):
channel_weights = nd.arange(1, 1+label.shape[1], ctx=label.context).reshape(1, -1, 1, 1) * 0.5
# point_weights = 0.1 + label + (nd.clip(label, 0.5, 1.0) - 0.5) * 10
point_weights = nd.exp(label*4) - 0.8
tp = nd.abs(pred - label) * point_weights * channel_weights
cloud_ratio = nd.mean(label > 0.1, axis=(0, 1), exclude=True)
bc = nd.mean(tp, axis=(0, 1), exclude=True) * cloud_ratio
return nd.mean(bc, axis=0, exclude=True)
def forward(self, cls_pred, box_pred, cls_target, box_target):
"""Compute loss in entire batch across devices."""
# require results across different devices at this time
cls_pred, box_pred, cls_target, box_target = [
_as_list(x) for x in (cls_pred, box_pred, cls_target, box_target)]
# cross device reduction to obtain positive samples in entire batch
num_pos = []
for cp, bp, ct, bt in zip(
*[cls_pred, box_pred, cls_target, box_target]):
pos_samples = (ct > 0)
num_pos.append(pos_samples.sum())
num_pos_all = sum([p.asscalar() for p in num_pos])
if num_pos_all < 1:
# no positive samples found, return dummy losses
return nd.zeros((1,)), nd.zeros((1,)), nd.zeros((1,))
# compute element-wise cross entropy loss and sort, then perform
# negative mining
cls_losses = []
box_losses = []
sum_losses = []
for cp, bp, ct, bt in zip(
*[cls_pred, box_pred, cls_target, box_target]):
pred = nd.log_softmax(cp, axis=-1)
pos = ct > 0
cls_loss = -nd.pick(pred, ct, axis=-1, keepdims=False)
rank = (cls_loss * (pos - 1)).argsort(axis=1).argsort(axis=1)
hard_negative = rank < (
pos.sum(axis=1) * self._negative_mining_ratio).expand_dims(-1)
# mask out if not positive or negative
cls_loss = nd.where(
(pos + hard_negative) > 0,
cls_loss,
nd.zeros_like(cls_loss))
cls_losses.append(
nd.sum(
cls_loss,
axis=0,
exclude=True) /
num_pos_all)
bp = _reshape_like(nd, bp, bt)
box_loss = nd.abs(bp - bt)
box_loss = nd.where(
box_loss > self._rho,
box_loss - 0.5 * self._rho,
(0.5 / self._rho) * nd.square(box_loss))
# box loss only apply to positive samples
box_loss = box_loss * pos.expand_dims(axis=-1)
box_losses.append(
nd.sum(
box_loss,
axis=0,
exclude=True) /
num_pos_all)
sum_losses.append(cls_losses[-1] + self._lambd * box_losses[-1])
return sum_losses, cls_losses, box_losses
def fmap_lossfunc(fmap, focus_lbl):
"""
calculate loss among feature-map generated by focus branch
:param fmap: mxnet ndarray, feature map
:param focus_lbl: mxnet ndarray, label map generated by affine_fmap-n-gt
:return: loss.
"""
res = nd.abs(nd.sum(focus_lbl - fmap) / nd.sum(focus_lbl))
return res
def accuracy(y_hat, y):
sum = 0
yy = y_hat - y
yy = nd.abs(yy)
i = 0
for i, val in enumerate(yy):
sum = sum + equal(val)
return sum / i
def contrastive_loss(net,data,label):
label = label.reshape(-1, 1)
label_mat = nd.relu(-nd.abs(label - label.T) + 1).astype('float32')
vec = net(data)
vec = nd.Flatten(vec)
dist_self = nd.sum(nd.square(vec), axis=1, keepdims=True)
dist_mat = nd.broadcast_add(dist_self, dist_self.T) - 2 * nd.dot(vec, vec.T)
loss = label_mat * dist_mat + nd.relu((1.0 - dist_mat * (1 - label_mat))).astype('float32')
return loss
def norm(tensor, order=2, axis=None):
# handle difference in default axis notation
if axis is None:
axis = ()
if order == 'inf':
res = nd.max(nd.abs(tensor), axis=axis)
elif order == 1:
res = nd.sum(nd.abs(tensor), axis=axis)
elif order == 2:
res = nd.sqrt(nd.sum(tensor**2, axis=axis))
else:
res = nd.sum(nd.abs(tensor)**order, axis=axis)**(1 / order)
if res.shape == (1, ):
return res.asscalar()
return res
def forward(self, output, mask, ind, target):
pred = _tranpose_and_gather_feat(output, ind)
pred = pred.swapaxes(dim1=0, dim2=1)
mask = mask.expand_dims(axis=2).broadcast_like(pred).astype('float32')
pred = pred / (target + 1e-4)
target = target * 0 + 1
loss = nd.abs(pred * mask - target * mask).sum()
loss = loss / (mask.sum() + 1e-4)
return loss
def compare():
nparams = []
diff = 0.
print('p, b {} {}'.format(params[0][0],
BNs[0].params.get('running_mean').data()[0]))
for param, bn in zip(params, BNs):
nparam = bn.params.get('running_mean').data()
diff += nd.sum(nd.abs(nparam - param)).asscalar()
return diff
def quantize_to(x, bits=8):
max_v = nd.max(nd.abs(x))
if max_v == 0:
return x.astype(np.int8), 8
int_len = nd.ceil(nd.log2(max_v)).asscalar()
sb = bits - int_len
f = 2**sb
y = nd.floor(x * f)
y = nd.clip(y, a_min=-2**(bits - 1), a_max=2**(bits - 1) - 1)
return y, sb
def forward(self, cls_pred, box_pred, cls_target, box_target):
"""Compute loss in entire batch across devices."""
# require results across different devices at this time
cls_pred, box_pred, cls_target, box_target = [_as_list(x) \
for x in (cls_pred, box_pred, cls_target, box_target)]
# cross device reduction to obtain positive samples in entire batch
pos_ct = [ct > 0 for ct in cls_target]
num_pos = [ct.sum() for ct in pos_ct]
num_pos_all = sum([p.asscalar() for p in num_pos])
# print ('num_pos_all: {}'.format(num_pos_all))
if num_pos_all < 1 and self._min_hard_negatives < 1:
# no positive samples and no hard negatives, return dummy losses
cls_losses = [nd.sum(cp * 0) for cp in cls_pred]
box_losses = [nd.sum(bp * 0) for bp in box_pred]
sum_losses = [
nd.sum(cp * 0) + nd.sum(bp * 0)
for cp, bp in zip(cls_pred, box_pred)
]
return sum_losses, cls_losses, box_losses
# compute element-wise cross entropy loss and sort, then perform negative mining
cls_losses = []
box_losses = []
sum_losses = []
for cp, bp, ct, bt in zip(
*[cls_pred, box_pred, cls_target, box_target]):
# print ('cp shape: {}'.format(cp.shape))
# print ('bp shape: {}'.format(bp.shape))
# print ('ct shape: {}'.format(ct.shape))
# print ('bt shape: {}'.format(bt.shape))
pred = nd.log_softmax(cp, axis=-1)
pos = ct > 0
cls_loss = -nd.pick(pred, ct, axis=-1, keepdims=False)
rank = (cls_loss * (pos - 1)).argsort(axis=1).argsort(axis=1)
hard_negative = rank < nd.maximum(
self._min_hard_negatives,
pos.sum(axis=1) * self._negative_mining_ratio).expand_dims(-1)
# mask out if not positive or negative
cls_loss = nd.where((pos + hard_negative) > 0, cls_loss,
nd.zeros_like(cls_loss))
cls_losses.append(
nd.sum(cls_loss, axis=0, exclude=True) / max(1., num_pos_all))
bp = _reshape_like(nd, bp, bt)
box_loss = nd.abs(bp - bt)
box_loss = nd.where(box_loss > self._rho,
box_loss - 0.5 * self._rho,
(0.5 / self._rho) * nd.square(box_loss))
# box loss only apply to positive samples
box_loss = box_loss * pos.expand_dims(axis=-1)
box_losses.append(
nd.sum(box_loss, axis=0, exclude=True) / max(1., num_pos_all))
sum_losses.append(cls_losses[-1] + self._lambd * box_losses[-1])
return sum_losses, cls_losses, box_losses
def forward(self, cls_pred, box_pred, cls_target, box_target):
"""Compute loss in entire batch across devices."""
# require results across different devices at this time
cls_pred, box_pred, cls_target, box_target = [_as_list(x) \
for x in (cls_pred, box_pred, cls_target, box_target)]
# cross device reduction to obtain positive samples in entire batch
num_pos = []
for cp, bp, ct, bt in zip(*[cls_pred, box_pred, cls_target, box_target]):
pos_samples = (ct > 0)
num_pos.append(pos_samples.sum())
num_pos_all = sum([p.asscalar() for p in num_pos])
if num_pos_all < 1 and self._min_hard_negatives < 1:
# no positive samples and no hard negatives, return dummy losses
cls_losses = [nd.sum(cp * 0) for cp in cls_pred]
box_losses = [nd.sum(bp * 0) for bp in box_pred]
sum_losses = [nd.sum(cp * 0) + nd.sum(bp * 0) for cp, bp in zip(cls_pred, box_pred)]
return sum_losses, cls_losses, box_losses
# compute element-wise cross entropy loss and sort, then perform negative mining
cls_losses = []
box_losses = []
sum_losses = []
for cp, bp, ct, bt in zip(*[cls_pred, box_pred, cls_target, box_target]):
pred = nd.log_softmax(cp, axis=-1)
pos = ct > 0
cls_loss = -nd.pick(pred, ct, axis=-1, keepdims=False)
rank = (cls_loss * (pos - 1)).argsort(axis=1).argsort(axis=1)
hard_negative = rank < nd.maximum(self._min_hard_negatives, pos.sum(axis=1)
* self._negative_mining_ratio).expand_dims(-1)
# mask out if not positive or negative
cls_loss = nd.where((pos + hard_negative) > 0, cls_loss, nd.zeros_like(cls_loss))
cls_losses.append(nd.sum(cls_loss, axis=0, exclude=True) / max(1., num_pos_all))
bp = _reshape_like(nd, bp, bt)
box_loss = nd.abs(bp - bt)
box_loss = nd.where(box_loss > self._rho, box_loss - 0.5 * self._rho,
(0.5 / self._rho) * nd.square(box_loss))
# box loss only apply to positive samples
box_loss = box_loss * pos.expand_dims(axis=-1)
box_losses.append(nd.sum(box_loss, axis=0, exclude=True) / max(1., num_pos_all))
sum_losses.append(cls_losses[-1] + self._lambd * box_losses[-1])
return sum_losses, cls_losses, box_losses