Skip to content

Latest commit

 

History

History
108 lines (74 loc) · 3.92 KB

note.md

File metadata and controls

108 lines (74 loc) · 3.92 KB

#PRN: Epoch30: loss: 0.0111 - templateLoss: 0.0547 - mean_absolute_error: 0.0564 - val_loss: 0.0152 - val_templateLoss: 0.0748 - val_mean_absolute_error: 0.0646 Epoch5: loss: 0.0386 - templateLoss: 0.1895 - mean_absolute_error: 0.1131 - val_loss: 0.0412 - val_templateLoss: 0.2016 - val_mean_absolute_error: 0.1195

#APRN: l=1e-3(encode64) Epoch30: loss: 0.3177 - templateLoss: 0.0444 - mean_absolute_error: 0.0569 - val_loss: 0.1090 - val_templateLoss: 0.0781 - val_mean_absolute_error: 0.0720 Epoch5: loss: 1.1835 - templateLoss: 0.1690 - mean_absolute_error: 0.1151 - val_loss: 0.9979 - val_templateLoss: 0.1608 - val_mean_absolute_error: 0.1098

#parameter: init:13353618 initmy:13352633 prnmy:32127209 ###the number of BN parameters are not the same Total params: 13,372,445 Trainable params: 13,360,555 Non-trainable params: 11,890

running log

(with wightmask*16)
(init)
    momentum0.01: 10.9 15:20 (gpu1  tb 输入zscore了   3.76%  30/50epoch)
    momentum0.5: 10.9 17:50  (gpu1 init   输入进行了z-score normalize)
    momentum0.01: 10.9 18:12 (gpu1 train 输入进行了z-score normalize  )
             0.5  10.9 20:00  gpu4 qua   3.88%  40epoch
             
    momentum0.5 MCG03 train  2019-10-11-8:00  normalized  tanh (主要针对负的posmap的问题)  [get:3.72]
    
                MCG03  10-15-20:34  train    zeroz
    
                [gpu07 train]2019-10-16-9-8-46+2019-10-17-10-19-52  initprn2 尝试复现结果 [get3.72]
    
 (qua)
    m0.5  MCG03 qua quaternion  lossrate 0 :1 :255:500 2019-10-13-15:00
        momentum0.5 [MG03 qua] 10-12-17:30 quaternion loss 0:1:500:500
     
(Attention)    
    修改了erase方式  tanh 
    momentum0.5 [attention] 2019-10-13-03:00  normalized tanh attention   attention的训练  no clip  attentionlossrante=0.03  单卡 [get3.72  epoch32]
    
    momentum0.5 [train] 2019-10-13-15:30   normalized tanh attention  lossrate=1  [get3.75]
    
    momentum0.5 [attention3] 2019-10-13-15:49 lossrate1 no clip   [get bad]

    m0.5 [attention2]  10-15-13:28+2019-10-17-10-29-16  lossrate0.1  noclip  l2rate=0.0001 [get3.68]
    


10.28 lr5e-4 batchsize 32比16略好   10轮下降0.1不可取

600blocks:
10.31 lr1e-5warmup siam
11.1 9:55 1r1e-5warmup init

630blocks
11.1 14:29 lr1e-5 init 
11.3 23:00 attentionbatch16/48   10:35 siam


11.5
晚上三组 attention batch48 l2=0.0001 lossrate=0.1
         完全体  visible
         attention batch48 l2=0.0001 lossrate=0.5
         
         
 visible1 batchsize48 比visible稍微下降0.01基本没问题
 visible2 finalposerate 0.01 性能下降
 
 
11.11上午两组SDN
(visible2)
    self.criterion0 = getLossFunction('fwrse')(0.1)  # final pos
    self.criterion1 = getLossFunction('fwrse')(0.5)  # offset
    self.criterion2 = getLossFunction('fwrse')(1)  # kpt
    self.criterion3 = getLossFunction('bce')(0.1)  # attention
    self.criterion4 = getLossFunction('smooth')(0.)
    self.metrics0 = getLossFunction('nme')(1.)
    self.metrics1 = getLossFunction('frse')(1.)
    self.metrics2 = getLossFunction('kptc')(1.)
    self.metrics3 = getLossFunction('mae')(1.)
 
(visible1)
    decay 0.0002
    
11.12
    decay 0.0002 smooth0.0025
    

2.5
smooth0.1 SDRN   20:32
smooth0.25 RT    22:15
smooth0.025 qua  22:17

2.25
加了新的augmentation
finetune  2e-5 1e-3学习率
SDRNv2 二阶导当loss 取消整体loss
SDRN 在mcg03上 把kptloss改成只计算kpt处



2.27
v2 SDRN  /data1/rzy/disk/APRN/savedmodel/temp_best_model/2020-2-27-8-28-5SDRN
light lightSDRN  /data1/rzy/disk/APRN/savedmodel/temp_best_model/2020-2-27-10-59-10SDRNv2