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
def mse(o, y):
return nd.square(o - y).sum() * 0.5 / self.X.shape[0], nd.argmax(
o, axis=1)
def grad_grad_op(x):
return x/nd.sqrt((nd.square(x)+1)**3)
def forward(self, fg, bg, pred, mask, merg):
c = fg * pred + (1 - pred) * bg
dis = mask * (c - merg)
l = nd.sqrt(self.eps + nd.square(dis).sum(0))
return l
def score(gradient, v, f):
if 2 * f + 2 > v.shape[1]:
f = int(math.floor((v.shape[1] - 2) / 2.0))
num_neighbours = v.shape[1] - 2 - f
sorted_distance = nd.square(v - gradient).sum(axis=0).sort()
return nd.sum(sorted_distance[1:(1 + num_neighbours)]).asscalar()
def forward(self, x):
x = nd.sigmoid(nd.sum(nd.square(x), 1))
return x
def garchLLH(y, par):
h = garchSim(nd.square(y), par)
T = y.shape[0]
llh = -0.5 * (T - 1) * math.log(
2 * math.pi) - 0.5 * nd.sum(nd.log(h) + (y / nd.sqrt(h))**2)
return llh.asscalar()
def unlabeled_train_op_mmd_combine(self, update_enc=True):
'''
Trains the MMD model
'''
batch_size = self.args['batch_size']
model_ctx = self.model_ctx
eps = 1e-10
# Retrieve data
docs = self.data.get_documents(key='train')
y_true = np.random.dirichlet(np.ones(self.ndim_y) *
self.args['dirich_alpha'],
size=batch_size)
y_true = nd.array(y_true, ctx=model_ctx)
with autograd.record():
### reconstruction phase ###
y_onehot_u = self.Enc(docs)
y_onehot_u_softmax = nd.softmax(y_onehot_u)
if self.args['latent_noise'] > 0:
y_noise = np.random.dirichlet(np.ones(self.ndim_y) *
self.args['dirich_alpha'],
size=batch_size)
y_noise = nd.array(y_noise, ctx=model_ctx)
y_onehot_u_softmax = (
1 - self.args['latent_noise']
) * y_onehot_u_softmax + self.args['latent_noise'] * y_noise
x_reconstruction_u = self.Dec(y_onehot_u_softmax)
logits = nd.log_softmax(x_reconstruction_u)
loss_reconstruction = nd.mean(nd.sum(-docs * logits, axis=1))
loss_total = loss_reconstruction * self.args['recon_alpha']
### mmd phase ###
if self.args['adverse']:
y_fake = self.Enc(docs)
y_fake = nd.softmax(y_fake)
loss_mmd = mmd_loss(y_true,
y_fake,
ctx_model=model_ctx,
t=self.args['kernel_alpha'])
loss_total = loss_total + loss_mmd
if self.args['l2_alpha'] > 0:
loss_total = loss_total + self.args['l2_alpha'] * nd.mean(
nd.sum(nd.square(y_onehot_u), axis=1))
loss_total.backward()
self.optimizer_enc.step(1)
self.optimizer_dec.step(1) # self.m.args['batch_size']
latent_max = nd.zeros(self.args['ndim_y'], ctx=model_ctx)
for max_ind in nd.argmax(y_onehot_u, axis=1):
latent_max[max_ind] += 1.0
latent_max /= batch_size
latent_entropy = nd.mean(
nd.sum(-y_onehot_u_softmax * nd.log(y_onehot_u_softmax + eps),
axis=1))
latent_v = nd.mean(y_onehot_u_softmax, axis=0)
dirich_entropy = nd.mean(nd.sum(-y_true * nd.log(y_true + eps),
axis=1))
if self.args['adverse']:
loss_mmd_return = loss_mmd.asscalar()
else:
loss_mmd_return = 0.0
return nd.mean(loss_reconstruction).asscalar(
), loss_mmd_return, latent_max.asnumpy(), latent_entropy.asscalar(
), latent_v.asnumpy(), dirich_entropy.asscalar()
even_split=False)
angs_list = mx.gluon.utils.split_and_load(angs,
ctx_list=devices,
even_split=False)
cate_list = mx.gluon.utils.split_and_load(cate,
ctx_list=devices,
even_split=False)
loss_list = []
with mx.autograd.record():
for data, lmks, angs, cate in zip(data_list, lmks_list,
angs_list, cate_list):
lmks_regs = net(data)
lmks_regs = nd.Flatten(lmks_regs)
lmks_loss = nd.square(lmks_regs - lmks)
lmks_loss = nd.sum(lmks_loss, axis=1)
#angs_loss = 1 - mx.nd.cos((angs_regs - angs))
#angs_loss = mx.nd.sum(angs_loss, axis=1)
loss = lmks_loss
#if with_angle:
# loss = loss * angs_loss
if with_category:
loss = loss * cate
loss_list.append(loss)
def RecLoss(rec_x, x):
x_reshape = x.reshape((0, -1))
diff = nd.square(x_reshape - rec_x)
return nd.mean(nd.sum(diff, axis=1))
def flowdr(dem_fill,NoData,rows,cols,ctx,switch):
ingrid = np.indices((rows, cols))
ingrid[0] # row indices
ingrid[1] # column indices
ingridxmx=nd.array(ingrid[1],ctx[0]).reshape((1,1,rows, cols))
ingridymx=nd.array(ingrid[0],ctx[0]).reshape((1,1,rows, cols))
dem_fillmx=nd.array(dem_fill,ctx[0])
demmx=dem_fillmx.reshape((1,1,rows, cols))
res=1
l=[0,1,2,3,4,5,6,7,0]
direct=[1,2,4,8,16,32,64,128]
direct_d=[[1,3],[2,6],[4,12],[8,24],[16,48],[32,96],[64,192],[128,129]]
weight=[None]*8
weight1=[None]*8
convx=[None]*8
convy=[None]*8
convz=[None]*8
runlen=[1,ma.pow(2,0.5),1,ma.pow(2,0.5),1,ma.pow(2,0.5),1,ma.pow(2,0.5)]*res
n = [[[] for x in range(3)] for x in range(8)]#create list to store normal vectors for each facet
s = [None]*8
d = [None]*8
weight[0] = nd.array([[0, 0, 0], [0, 1, -1], [0, 0, 0]], gpu(0))
weight[1] = nd.array([[0, 0, -1], [0, 1, 0], [0, 0, 0]], gpu(0))
weight[2] = nd.array([[0, -1, 0], [0, 1, 0], [0, 0, 0]], gpu(0))
weight[3] = nd.array([[-1, 0, 0], [0, 1, 0], [0, 0, 0]], gpu(0))
weight[4] = nd.array([[0, 0, 0], [-1, 1, 0], [0, 0, 0]], gpu(0))
weight[5] = nd.array([[0, 0, 0], [0, 1, 0], [-1, 0, 0]], gpu(0))
weight[6] = nd.array([[0, 0, 0], [0, 1, 0], [0, -1, 0]], gpu(0))
weight[7] = nd.array([[0, 0, 0], [0, 1, 0], [0, 0, -1]], gpu(0))
weight1[0] = nd.array([[0, 0, 0], [0, 1, -10], [0, 0, 0]], gpu(0))
weight1[1] = nd.array([[0, 0, -10], [0, 1, 0], [0, 0, 0]], gpu(0))
weight1[2] = nd.array([[0, -10, 0], [0, 1, 0], [0, 0, 0]], gpu(0))
weight1[3] = nd.array([[-10, 0, 0], [0, 1, 0], [0, 0, 0]], gpu(0))
weight1[4] = nd.array([[0, 0, 0], [-10, 1, 0], [0, 0, 0]], gpu(0))
weight1[5] = nd.array([[0, 0, 0], [0, 1, 0], [-10, 0, 0]], gpu(0))
weight1[6] = nd.array([[0, 0, 0], [0, 1, 0], [0, -10, 0]], gpu(0))
weight1[7] = nd.array([[0, 0, 0], [0, 1, 0], [0, 0, -10]], gpu(0))
d0=nd.zeros((rows, cols),ctx[0],dtype='float32')
dd=nd.zeros((rows, cols),ctx[0],dtype='float32')
d_flat=nd.zeros((rows, cols),ctx[0],dtype='float32')
flat=nd.zeros((rows, cols),ctx[0],dtype='float32')
dep=nd.zeros((rows, cols),ctx[0],dtype='float32')
high=nd.zeros((rows, cols),ctx[0],dtype='float32')
fd=nd.zeros((rows, cols),ctx[0],dtype='float32')-999
d_compact=nd.zeros((rows, cols),ctx[0],dtype='float32')-1
for i in range(0,8):
w=weight[i].reshape((1, 1, 3, 3))
convz[i] = nd.Convolution(data=demmx, weight=w, kernel=(3,3), no_bias=True, num_filter=1,pad=(1,1),cudnn_tune='off')
convz[i]=convz[i][0,0,:,:]
if switch==1 or 3:
convx[i] = nd.Convolution(data=ingridxmx, weight=w, kernel=(3,3), no_bias=True, num_filter=1,pad=(1,1),cudnn_tune='off')
convy[i] = nd.Convolution(data=ingridymx, weight=w, kernel=(3,3), no_bias=True, num_filter=1,pad=(1,1),cudnn_tune='off')
convx[i]=convx[i][0,0,:,:]
convy[i]=convy[i][0,0,:,:]
if switch==1 or 3:
for p in range(0,8):#8 facets from N-NE clockwise
l0=l[p]
l1=l[p+1]
d[l0]=d0-999#Nodata value
dmax=d0-999
smax=d0-999
n[l0][0]= convz[l0]*convy[l1]-convz[l1]*convy[l0]#nx
n[l0][1]= convz[l0]*convx[l1]-convz[l1]*convx[l0]#ny
n[l0][2]= convy[l0]*convx[l1]-convy[l1]*convx[l0]#nz
#make boolean mask to determine direction d and slope s
d[l0]=nd.where(condition=((n[l0][0]==0)*(n[l0][1]>=0)),x=d0,y=d[l0])
d[l0]=nd.where(condition=((n[l0][0]==0)*(n[l0][1])<0),x=d0+ma.pi,y=d[l0])
d[l0]=nd.where(condition=(n[l0][0]>0),x=ma.pi/2-nd.arctan(n[l0][1]/n[l0][0]),y=d[l0])
d[l0]=nd.where(condition=(n[l0][0]<0),x=3*ma.pi/2-nd.arctan(n[l0][1]/n[l0][0]),y=d[l0])
d[l0]=nd.where(condition=((convz[l0]<=0)*(convz[l1]<=0)),x=dmax,y=d[l0])
s[l0]=-nd.tan(nd.arccos(n[l0][2]/(nd.sqrt(nd.square(n[l0][0])+nd.square(n[l0][1])+nd.square(n[l0][2])))))#slope of the triangular facet
s[l0]=nd.where(condition=((convz[l0]<=0)*(convz[l1]<=0)),x=smax,y=s[l0])
#Modify the scenario when the steepest slope is outside the 45 range of each facet
dmax=nd.where(condition=((convz[l0]/runlen[l0]>=convz[l1]/runlen[l0])*(convz[l0]>0)),x=d0+ma.pi*l0/4,y=dmax)
dmax=nd.where(condition=((convz[l0]/runlen[l0]<convz[l1]/runlen[l0])*(convz[l1]>0)),x=d0+ma.pi*(l0+1)/4,y=dmax)
smax=nd.where(condition=((convz[l0]>=convz[l1])*(convz[l0]>0)),x=convz[l0]/runlen[l0],y=smax)
smax=nd.where(condition=((convz[l0]<convz[l1])*(convz[l1]>0)),x=convz[l1]/runlen[l1],y=smax)
d[l0]=nd.where(condition=((d[l0]<ma.pi*l0/4)+(d[l0]>ma.pi*l1/4)),x=dmax,y=d[l0])
s[l0]=nd.where(condition=((d[l0]<ma.pi*l0/4)+(d[l0]>ma.pi*l1/4)),x=smax,y=s[l0])
if switch==1:
#flat and depressions indicator grid
flat=(convz[l0]==0)+flat
dep=(convz[l0]<0)+dep
high=(convz[l0]>0)+high
for q in range(0,8):#check if the 45 degree range angles need to be maintaied, otherwise delete (set to NoData)
l0=l[q]
l1=l[q+1]
l2=l[q-1]
dmax=d0-999
if q==0:
dmax=nd.where(condition=(d[0]==d[1]),x=d[0],y=dmax)
dmax=nd.where(condition=(d[0]==d[7]),x=d[0],y=dmax)
d[0]=nd.where(condition=((d[0]==ma.pi*l0/4)+(d[0]==ma.pi*l1/4)),x=dmax,y=d[0])
else:
dmax=nd.where(condition=(d[l0]==d[l1]),x=d[l0],y=dmax)
dmax=nd.where(condition=(d[l0]==d[l2]),x=d[l0],y=dmax)
d[l0]=nd.where(condition=((d[l0]==ma.pi*l0/4)+(d[l0]==ma.pi*l1/4)),x=dmax,y=d[l0])
#Check if flat or surface depression area. then lable with -1 or -10 respectively
if switch==1:
fd=nd.where(condition=(flat==8),x=d0-2,y=fd)#flats
fd=nd.where(condition=(dep>=1)*(high==0),x=d0-3,y=fd)#high edge
high_zero=nd.where(condition=(high==0),x=d0+1,y=d0)
for j in range (0,8):
if switch==1 or switch==2:
d_flat=nd.where(condition=(convz[j]==0),x=d0+direct[j],y=d0)+d_flat
if switch==1:
flat_near=nd.where(condition=(convz[j]==0),x=d0+5,y=d0)
dd1=high_zero+flat_near
w=weight1[j].reshape((1, 1, 3, 3))
dd1=dd1.reshape((1,1,rows, cols))
conv_near= nd.Convolution(data=dd1, weight=w, kernel=(3,3), no_bias=True, num_filter=1,pad=(1,1),cudnn_tune='off')
conv_near= conv_near[0,0,:,:]
dd=nd.where(condition=(conv_near==-5)+(conv_near==-59)+(conv_near==-54)+(conv_near==-4),x=d0+1,y=d0)+dd
if switch==1 or switch==3:
d_compact=nd.where(condition=(d[j]==ma.pi*j/4),x=d0+direct_d[j][0],y=d_compact)
d_compact=nd.where(condition=(d[j]>j*ma.pi/4)*(d[j]<(j+1)*ma.pi/4),x=d0+direct_d[j][1],y=d_compact)
if switch==1 or switch==3:
d_compact=nd.where(condition=(dem_fillmx==d0+NoData),x=d0-999,y=d_compact)#NoData
if switch==1:
fd=nd.where(condition=(dd>=1)*(high>=1),x=d0-1,y=fd)#low edge
fd=nd.where(condition=(dep==8),x=d0-10,y=fd)#lowest points in depressions
return (fd.asnumpy(),d_compact.asnumpy(),d_flat.asnumpy())
if switch==2:
return (d_flat.asnumpy())
if switch==3:
return (d_compact.asnumpy())
def Norm(x):
return nd.sqrt(nd.sum(nd.square(x), axis=1, keepdims=True))
def _variance(a: nd.NDArray) -> nd.NDArray: """Compute variance of a of shape [n_samples, ...].""" mean = nd.mean(a, 0, keepdims=True) return nd.mean(nd.square(a - mean), 0)
def CapLoss(y_pred, y_true):
L = y_true * nd.square(nd.maximum(0., 0.9 - y_pred)) + \
0.5 * (1 - y_true) * nd.square(nd.maximum(0., y_pred - 0.1))
return nd.mean(nd.sum(L, 1))
def total_variation_loss(x):
""" regularize convolutional masks (not currently in use) """
a = nd.square(x[:, :, :-1, :-1] - x[:, :, 1:, :-1])
b = nd.square(x[:, :, :-1, :-1] - x[:, :, :-1, 1:])
return nd.sum(nd.mean(nd.power(a + b, 1.25), axis=(2,3)))
def diffusion_kernel(a, tmpt, dim):
# return (4 * np.pi * tmpt)**(-dim / 2) * nd.exp(- nd.square(nd.arccos(a)) / tmpt)
return nd.exp(-nd.square(nd.arccos(a)) / tmpt)
def loss(predictions, targets):
# return -nd.mean(targets * nd.log(predictions))
# return -nd.mean((targets * nd.log(predictions)) + ((1 - targets) * nd.log(1 - predictions)))
return nd.mean(nd.square(predictions - targets))
def loss_fn(y_pred, y):
return nd.mean(nd.square(y_pred - y))
def fit(self, num_steps=1):
"""
Fit the models
Returns:
Loss Functions (Q1-mse, Q2-mse, alpha-entropy, Policy-kl)
"""
logger_data = {k: [] for k in ["LossPi", "LossQ1", "LossQ2", "LossV"]}
for step in range(num_steps):
# sample a batch from memory
minibatch = self.memory.sample(self.batch_size)
obs = nd.array(minibatch["obs"], self.ctx)
acts = nd.array(minibatch["act"], self.ctx)
rewards = nd.array(minibatch["rew"], self.ctx)
next_obs = nd.array(minibatch["next_obs"], self.ctx)
nonterm = nd.array(minibatch["nt"], self.ctx)
lr = self.lr(self.steps) * self.lrmult
# update the policy function
with autograd.record():
_mu, _pi, _logp_pi = self.policy(obs)
_obspi = nd.concat(obs, _pi, dim=-1)
_q1_pi = self.qfn1(_obspi)
pi_loss = nd.mean(self.alpha * _logp_pi - _q1_pi)
pi_loss.backward()
self.mu.update(lr)
self.logstd.update(lr)
self.policy_base.update(lr)
# update the value functions
logp_pi = nd.stop_gradient(_logp_pi)
obspi = nd.stop_gradient(_obspi)
obsact = nd.concat(obs, acts, dim=-1)
q1_pi = self.qfn1(obspi)
q2_pi = self.qfn2(obspi)
min_q_pi = nd.minimum(q1_pi, q2_pi)
v_targ = self.vfn_targ(next_obs)
q_backup = nd.stop_gradient(rewards +
self.gamma * nonterm * v_targ)
v_backup = nd.stop_gradient(min_q_pi - self.alpha * logp_pi)
with autograd.record():
_q1 = self.qfn1(obsact)
_q2 = self.qfn2(obsact)
_v = self.vfn(obs)
q1_loss = 0.5 * nd.mean(nd.square(q_backup - _q1))
q2_loss = 0.5 * nd.mean(nd.square(q_backup - _q2))
v_loss = 0.5 * nd.mean(nd.square(v_backup - _v))
total_loss = q1_loss + q2_loss + v_loss
total_loss.backward()
self.qfn1.update(lr)
self.qfn2.update(lr)
self.vfn.update(lr)
# update the target network
for i in range(len(self.vfn.weights)):
self.vfn_targ.weights[i][:] = \
self.polyak * self.vfn_targ.weights[i][:] + \
(1 - self.polyak) * self.vfn.weights[i][:]
logger_data["LossPi"].append(pi_loss.asnumpy()[0])
logger_data["LossQ1"].append(q1_loss.asnumpy()[0])
logger_data["LossQ2"].append(q2_loss.asnumpy()[0])
logger_data["LossV"].append(v_loss.asnumpy()[0])
return logger_data
def forward(self, x):
x = nd.sqrt(nd.sum(nd.square(x), 1))
return x
def neglogp(action, mean, logstd):
assert (mean.shape[-1] == logstd.shape[-1])
std = nd.exp(logstd) + 1e-8
return 0.5 * nd.sum(nd.square((action - mean) / std), axis=-1) \
+ 0.5 * np.log(2.0 * np.pi) * action.shape[-1] \
+ nd.sum(logstd, axis=-1)
if len(grads_list) >= args.nworkers:
accumulate_param = 0
accumulate_grad = 0
if model_idx != len(params_prev_list) - 1:
grads_prev = grads_list[-1]
params_prev = params_prev_list[-1]
else:
grads_prev = grads_list[-2]
params_prev = params_prev_list[-2]
for param, param_prev, grad_prev in zip(
net.collect_params().values(), params_prev,
grads_prev):
if param.grad_req != 'null':
grad_current = param.grad()
param_current = param.data()
accumulate_param = accumulate_param + nd.square(
param_current - param_prev).sum()
accumulate_grad = accumulate_grad + nd.square(
grad_current - grad_prev).sum()
lips = math.sqrt(accumulate_grad.asscalar()) / math.sqrt(
accumulate_param.asscalar())
if lips <= np.quantile(lips_list, quantile_q):
byz_flag = False
accept_counter = accept_counter + 1
nd.waitall()
else:
byz_flag = False
accept_counter = accept_counter + 1
elif args.byz_test == 'zeno++':
zeno_max_delay = args.zeno_delay
zeno_rho = args.rho
zeno_epsilon = args.epsilon
def squash(x, axis):
s_squared_norm = nd.sum(nd.square(x), axis, keepdims=True)
scale = s_squared_norm / (1 + s_squared_norm) / nd.sqrt(s_squared_norm +
1e-5)
return scale * x
def _update_params(self, accumulated_grads):
# scale gradients by lot size, add noise, and update the parameters
for param_name, param in self._params.items():
# average the clipped gradients and then add noise to each averaged gradient
param_grad_update = (accumulated_grads[param_name] / self._hyperparams['lot_size']) + \
mx.random.normal(0, self._hyperparams['sigma'], param.shape, ctx=self._model_ctx)
# update biased first moment estimate
self._m[param_name] = self._hyperparams['beta_1'] * self._m[param_name] + (1 - self._hyperparams['beta_1']) * param_grad_update
# update biased second raw moment estimate
self._v[param_name] = self._hyperparams['beta_2'] * self._v[param_name] + (1 - self._hyperparams['beta_2']) * nd.square(param_grad_update)
# compute bias-corrected first moment estimate
m_hat = self._m[param_name] / (1 - nd.power(self._hyperparams['beta_1'], self._step + 1))
# compute bias-corrected second raw moment estimate
v_hat = self._v[param_name] / (1 - nd.power(self._hyperparams['beta_2'], self._step + 1))
# update params with ADAM
param[:] = param - self._hyperparams['lr'] * m_hat / (nd.sqrt(v_hat) + 1e-8)
def Squash(vector, axis):
norm = nd.sum(nd.square(vector), axis, keepdims=True)
v_j = norm / (1 + norm) / nd.sqrt(norm, keepdims=True) * vector
return v_j
def my_l2_loss(X, Y):
num_instances = X.shape[0]
return nd.sum(nd.square(X - Y)) / (2 * num_instances)
def forward(self, x):
out = nd.sqrt(nd.sum(nd.square(x), self.axis))
return out
def forward(self, cls_pred, ori_pred, box_pred, cls_target, ori_target,
box_target):
"""Compute loss in entire batch across devices."""
# require results across different devices at this time
cls_pred, ori_pred, box_pred, cls_target, box_target = [_as_list(x) \
for x in (cls_pred, ori_pred, box_pred, cls_target, box_target)]
# cross device reduction to obtain positive samples in entire batch
num_pos = []
for cp, op, bp, ct, bt in zip(
*[cls_pred, ori_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]
ori_losses = [nd.sum(op * 0) for op in ori_pred]
box_losses = [nd.sum(bp * 0) for bp in box_pred]
sum_losses = [
nd.sum(cp * 0) + nd.sum(op * 0) + nd.sum(bp * 0)
for cp, op, bp in zip(cls_pred, ori_pred, box_pred)
]
return sum_losses, cls_losses, ori_losses, box_losses
# compute element-wise cross entropy loss and sort, then perform negative mining
cls_losses = []
ori_losses = []
box_losses = []
sum_losses = []
for cp, op, bp, ct, ot, bt in zip(
*
[cls_pred, ori_pred, box_pred, cls_target, ori_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))
pred = nd.log_softmax(op, axis=-1)
pos = ot > 0
ori_loss = -nd.pick(pred, ot, axis=-1, keepdims=False)
rank = (ori_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
ori_loss = nd.where((pos + hard_negative) > 0, ori_loss,
nd.zeros_like(ori_loss))
ori_losses.append(
nd.sum(ori_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])
sum_losses.append(ori_losses[-1] + self._lambd * box_losses[-1])
return sum_losses, cls_losses, ori_losses, box_losses
def forward(self, cls_pred, box_pred, coef_center_pred, coef_pred,
cls_target, box_target, coef_center_target, coef_target):
"""Compute loss in entire batch across devices."""
# require results across different devices at this time
# print(cls_pred[0].shape, box_pred[0].shape, coef_center_pred[0].shape)
cls_pred, box_pred, coef_center_pred, coef_pred, cls_target, box_target, coef_center_target, coef_target = [_as_list(x) \
for x in (cls_pred, box_pred, coef_center_pred, coef_pred, cls_target, box_target, coef_center_target, coef_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, )), nd.zeros(
(1, )), nd.zeros((1, ))
# compute element-wise cross entropy loss and sort, then perform negative mining
cls_losses = []
box_losses = []
coef_center_losses = []
coef_losses = []
sum_losses = []
for cp, bp, coefcp, coefp, ct, bt, coefct, coeft in zip(*[
cls_pred, box_pred, coef_center_pred, coef_pred, cls_target,
box_target, coef_center_target, coef_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)
# print(bp.shape, bt.shape)
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)
coefcp = _reshape_like(nd, coefcp, coefct)
coef_center_loss = nd.abs(coefcp - coefct)
coef_center_loss = nd.where(coef_center_loss > self._rho,
coef_center_loss - 0.5 * self._rho,
(0.5 / self._rho) *
nd.square(coef_center_loss))
coef_center_loss = coef_center_loss * pos.expand_dims(axis=-1)
coef_center_losses.append(
nd.sum(coef_center_loss, axis=0, exclude=True) / num_pos_all)
coefp = _reshape_like(nd, coefp, coeft)
coef_loss = nd.abs(coefp, coeft)
coef_loss = nd.where(coef_loss > self._rho,
coef_loss - 0.5 * self._rho,
(0.5 / self._rho) * nd.square(coef_loss))
coef_loss = coef_loss * pos.expand_dims(axis=-1)
coef_losses.append(
nd.sum(coef_loss, axis=0, exclude=True) / num_pos_all)
sum_losses.append(cls_losses[-1] + self._lambd * box_losses[-1] +
coef_losses[-1] + coef_center_losses[-1])
return sum_losses, cls_losses, box_losses, coef_center_losses, coef_losses