- faster rcnn训练可分为交替训练和多任务训练两种方式，后者损失的计算过程可参看train.py，
# RPN
# classification loss
rpn_cls_score = tf.reshape(self.net.get_output('rpn_cls_score_reshape'),[-1,2])
rpn_label = tf.reshape(self.net.get_output('rpn-data')[0],[-1])
rpn_cls_score = tf.reshape(tf.gather(rpn_cls_score,tf.where(tf.not_equal(rpn_label,-1))),[-1,2])
rpn_label = tf.reshape(tf.gather(rpn_label,tf.where(tf.not_equal(rpn_label,-1))),[-1])
rpn_cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=rpn_cls_score, labels=rpn_label))
# bounding box regression L1 loss
rpn_bbox_pred = self.net.get_output('rpn_bbox_pred')
rpn_bbox_targets = tf.transpose(self.net.get_output('rpn-data')[1],[0,2,3,1])
rpn_bbox_inside_weights = tf.transpose(self.net.get_output('rpn-data')[2],[0,2,3,1])
rpn_bbox_outside_weights = tf.transpose(self.net.get_output('rpn-data')[3],[0,2,3,1])
rpn_smooth_l1 = self._modified_smooth_l1(3.0, rpn_bbox_pred, rpn_bbox_targets, rpn_bbox_inside_weights, rpn_bbox_outside_weights)
rpn_loss_box = tf.reduce_mean(tf.reduce_sum(rpn_smooth_l1, reduction_indices=[1, 2, 3]))
# R-CNN
# classification loss
cls_score = self.net.get_output('cls_score')
label = tf.reshape(self.net.get_output('roi-data')[1],[-1])
cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=cls_score, labels=label))
# bounding box regression L1 loss
bbox_pred = self.net.get_output('bbox_pred')
bbox_targets = self.net.get_output('roi-data')[2]
bbox_inside_weights = self.net.get_output('roi-data')[3]
bbox_outside_weights = self.net.get_output('roi-data')[4]
smooth_l1 = self._modified_smooth_l1(1.0, bbox_pred, bbox_targets, bbox_inside_weights, bbox_outside_weights)
loss_box = tf.reduce_mean(tf.reduce_sum(smooth_l1, reduction_indices=[1]))
# final loss
loss = cross_entropy + loss_box + rpn_cross_entropy + rpn_loss_box- darknet损失的计算过程可参看network.c
void calc_network_cost(network *netp)
{
network net = *netp;
int i;
float sum = 0;
int count = 0;
for(i = 0; i < net.n; ++i){
if(net.layers[i].cost){
sum += net.layers[i].cost[0];
++count;
}
}
*net.cost = sum/count;
}- pytorch-yolo2的损失函数可参考该函数region_loss.py，也可参考yolo2-pytorch项目中的darknet.py(line216行for training部分)
def forward(self, output, target):
#output : BxAs*(4+1+num_classes)*H*W
t0 = time.time()
nB = output.data.size(0)
nA = self.num_anchors
nC = self.num_classes
nH = output.data.size(2)
nW = output.data.size(3)
output = output.view(nB, nA, (5+nC), nH, nW)#batch*anchor*(5+nc)*feature_map分辨率转换数据维度
x = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([0]))).view(nB, nA, nH, nW))
y = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([1]))).view(nB, nA, nH, nW))
w = output.index_select(2, Variable(torch.cuda.LongTensor([2]))).view(nB, nA, nH, nW)
h = output.index_select(2, Variable(torch.cuda.LongTensor([3]))).view(nB, nA, nH, nW)
conf = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([4]))).view(nB, nA, nH, nW)) #网络输出的检测参数的结果
cls = output.index_select(2, Variable(torch.linspace(5,5+nC-1,nC).long().cuda()))
cls = cls.view(nB*nA, nC, nH*nW).transpose(1,2).contiguous().view(nB*nA*nH*nW, nC) #识别类别
t1 = time.time()
pred_boxes = torch.cuda.FloatTensor(4, nB*nA*nH*nW)
grid_x = torch.linspace(0, nW-1, nW).repeat(nH,1).repeat(nB*nA, 1, 1).view(nB*nA*nH*nW).cuda()
grid_y = torch.linspace(0, nH-1, nH).repeat(nW,1).t().repeat(nB*nA, 1, 1).view(nB*nA*nH*nW).cuda()
anchor_w = torch.Tensor(self.anchors).view(nA, self.anchor_step).index_select(1, torch.LongTensor([0])).cuda()
anchor_h = torch.Tensor(self.anchors).view(nA, self.anchor_step).index_select(1, torch.LongTensor([1])).cuda()
anchor_w = anchor_w.repeat(nB, 1).repeat(1, 1, nH*nW).view(nB*nA*nH*nW)
anchor_h = anchor_h.repeat(nB, 1).repeat(1, 1, nH*nW).view(nB*nA*nH*nW)
pred_boxes[0] = x.data + grid_x
pred_boxes[1] = y.data + grid_y #相对输出feature map栅格的坐标偏移
pred_boxes[2] = torch.exp(w.data) * anchor_w
pred_boxes[3] = torch.exp(h.data) * anchor_h #相对于anchor的宽高比
pred_boxes = convert2cpu(pred_boxes.transpose(0,1).contiguous().view(-1,4)) #转换为nB*nA*nH*nW 乘以 4的二维向量
t2 = time.time()
nGT, nCorrect, coord_mask, conf_mask, cls_mask, tx, ty, tw, th, tconf,tcls = build_targets(pred_boxes, target.data, self.anchors, nA, nC, \
nH, nW, self.noobject_scale, self.object_scale, self.thresh, self.seen)
cls_mask = (cls_mask == 1)
nProposals = int((conf > 0.25).sum().data[0])
tx = Variable(tx.cuda())
ty = Variable(ty.cuda())
tw = Variable(tw.cuda())
th = Variable(th.cuda())
tconf = Variable(tconf.cuda())
tcls = Variable(tcls.view(-1)[cls_mask].long().cuda())
coord_mask = Variable(coord_mask.cuda())
conf_mask = Variable(conf_mask.cuda().sqrt())
cls_mask = Variable(cls_mask.view(-1, 1).repeat(1,nC).cuda())
cls = cls[cls_mask].view(-1, nC)
t3 = time.time()
loss_x = self.coord_scale * nn.MSELoss(size_average=False)(x*coord_mask, tx*coord_mask)/2.0
loss_y = self.coord_scale * nn.MSELoss(size_average=False)(y*coord_mask, ty*coord_mask)/2.0
loss_w = self.coord_scale * nn.MSELoss(size_average=False)(w*coord_mask, tw*coord_mask)/2.0
loss_h = self.coord_scale * nn.MSELoss(size_average=False)(h*coord_mask, th*coord_mask)/2.0
loss_conf = nn.MSELoss(size_average=False)(conf*conf_mask, tconf*conf_mask)/2.0 #边界框的置信度
loss_cls = self.class_scale * nn.CrossEntropyLoss(size_average=False)(cls, tcls) #分类结果
loss = loss_x + loss_y + loss_w + loss_h + loss_conf + loss_cls- PyTorch-YOLOv3的损失函数可参考models.py文件
# Get outputs
x = torch.sigmoid(prediction[..., 0]) # Center x
y = torch.sigmoid(prediction[..., 1]) # Center y
w = prediction[..., 2] # Width
h = prediction[..., 3] # Height
pred_conf = torch.sigmoid(prediction[..., 4]) # Conf
pred_cls = torch.sigmoid(prediction[..., 5:]) # Cls pred.
# Calculate offsets for each grid
grid_x = torch.arange(nG).repeat(nG, 1).view([1, 1, nG, nG]).type(FloatTensor)
grid_y = torch.arange(nG).repeat(nG, 1).t().view([1, 1, nG, nG]).type(FloatTensor)
scaled_anchors = FloatTensor([(a_w / stride, a_h / stride) for a_w, a_h in self.anchors])
anchor_w = scaled_anchors[:, 0:1].view((1, nA, 1, 1))
anchor_h = scaled_anchors[:, 1:2].view((1, nA, 1, 1))
# Add offset and scale with anchors
pred_boxes = FloatTensor(prediction[..., :4].shape)
pred_boxes[..., 0] = x.data + grid_x
pred_boxes[..., 1] = y.data + grid_y
pred_boxes[..., 2] = torch.exp(w.data) * anchor_w
pred_boxes[..., 3] = torch.exp(h.data) * anchor_h
# Training
if targets is not None:
if x.is_cuda:
self.mse_loss = self.mse_loss.cuda()
self.bce_loss = self.bce_loss.cuda()
self.ce_loss = self.ce_loss.cuda()
nGT, nCorrect, mask, conf_mask, tx, ty, tw, th, tconf, tcls = build_targets(
pred_boxes=pred_boxes.cpu().data,
pred_conf=pred_conf.cpu().data,
pred_cls=pred_cls.cpu().data,
target=targets.cpu().data,
anchors=scaled_anchors.cpu().data,
num_anchors=nA,
num_classes=self.num_classes,
grid_size=nG,
ignore_thres=self.ignore_thres,
img_dim=self.image_dim,
)
nProposals = int((pred_conf > 0.5).sum().item())
recall = float(nCorrect / nGT) if nGT else 1
precision = float(nCorrect / nProposals)
# Handle masks
mask = Variable(mask.type(ByteTensor))
conf_mask = Variable(conf_mask.type(ByteTensor))
# Handle target variables
tx = Variable(tx.type(FloatTensor), requires_grad=False)
ty = Variable(ty.type(FloatTensor), requires_grad=False)
tw = Variable(tw.type(FloatTensor), requires_grad=False)
th = Variable(th.type(FloatTensor), requires_grad=False)
tconf = Variable(tconf.type(FloatTensor), requires_grad=False)
tcls = Variable(tcls.type(LongTensor), requires_grad=False)
# Get conf mask where gt and where there is no gt
conf_mask_true = mask
conf_mask_false = conf_mask - mask
# Mask outputs to ignore non-existing objects
loss_x = self.mse_loss(x[mask], tx[mask])
loss_y = self.mse_loss(y[mask], ty[mask])
loss_w = self.mse_loss(w[mask], tw[mask])
loss_h = self.mse_loss(h[mask], th[mask])
loss_conf = self.bce_loss(pred_conf[conf_mask_false], tconf[conf_mask_false]) + self.bce_loss(
pred_conf[conf_mask_true], tconf[conf_mask_true]
)
loss_cls = (1 / nB) * self.ce_loss(pred_cls[mask], torch.argmax(tcls[mask], 1))
loss = loss_x + loss_y + loss_w + loss_h + loss_conf + loss_cls- ssd.pytorch的损失函数可参考文件multibox_loss.py
# Localization Loss (Smooth L1)
# Shape: [batch,num_priors,4]
pos_idx = pos.unsqueeze(pos.dim()).expand_as(loc_data)
loc_p = loc_data[pos_idx].view(-1, 4)
loc_t = loc_t[pos_idx].view(-1, 4)
loss_l = F.smooth_l1_loss(loc_p, loc_t, size_average=False)
# Compute max conf across batch for hard negative mining
batch_conf = conf_data.view(-1, self.num_classes)
loss_c = log_sum_exp(batch_conf) - batch_conf.gather(1, conf_t.view(-1, 1))
# Hard Negative Mining
loss_c[pos] = 0 # filter out pos boxes for now
loss_c = loss_c.view(num, -1)
_, loss_idx = loss_c.sort(1, descending=True)
_, idx_rank = loss_idx.sort(1)
num_pos = pos.long().sum(1, keepdim=True)
num_neg = torch.clamp(self.negpos_ratio*num_pos, max=pos.size(1)-1)
neg = idx_rank < num_neg.expand_as(idx_rank)
# Confidence Loss Including Positive and Negative Examples
pos_idx = pos.unsqueeze(2).expand_as(conf_data)
neg_idx = neg.unsqueeze(2).expand_as(conf_data)
conf_p = conf_data[(pos_idx+neg_idx).gt(0)].view(-1, self.num_classes)
targets_weighted = conf_t[(pos+neg).gt(0)]
loss_c = F.cross_entropy(conf_p, targets_weighted, size_average=False)
# Sum of losses: L(x,c,l,g) = (Lconf(x, c) + αLloc(x,l,g)) / N
N = num_pos.data.sum()
loss_l /= N
loss_c /= N
return loss_l, loss_c
参考文献：
6143
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
