2019.7.15
avg + att_avg :在resnet後分別進行batch_normal後直接進行數學相加,
mAP:82.3
r1:93.4
r5:98.8
att_avg:將avg替換成att_avg:
mAP:76.5
r1:90.9
r5:96.8
att_avg, avg作爲兩個獨立分支分別求loss:
mAP:86.1
r1:94.3
r5:98.9
從resnet50截出layer2輸出,進行att_avg作爲兩個獨立的分支分別求loss:
mAP:79.1
r1:92.7
r5:96.9
2019 7.16
19:35分
突然發現原來有之前的代碼只有encode部分沒有decode,然後瘋狂修改,終於能跑通了。但願mAP能破90
直接將decode輸出的分類分數與baseline最終輸出分類分數相加(使用同一層batch_normal):
mAP:85.2
r1:94.1
r5:98.0
att_avg, avg作爲兩個獨立分支分別求loss(使用不同的batch_normal)::
mAP:85.1
r1:93.5
r5:98.8
下一步工作雙向GRU(運行速度過慢);使用可學習參數相加特徵,或者l2_normalization(速度過慢)
self_attention
下一步
將交叉熵損失換成triple loss
換成ADL處理方式與第一個mask版本進行比較
gloab_attention + part_attention + l2_normal:
實驗單支gloab_attention效果
實驗attention與resnet avg 相加
g(x)(1 + M(x))
cam_loss test前加上batch_normal
tricks | mAP | R1 | R5 |
---|---|---|---|
只使用part_attention | 87.0 | 94.9 | 98.3 |
使用part_at,使用lnn(最後一個維度) | 79.5 | 92.3: | 97.1 |
使用part_att,gloab_att,使用ln(最後一個維度) | 85.6 | 94.3 | 98.2 |
使用part_att,gloab_att,使用ln(最後兩個維度) | 86.0 | 94.8 | 98.1 |
使用part_att,gloab_att,使用ln(最後兩個維度) | 86.0 | 94.8 | 98.1 |
使用part_att,gloab_att,使用ln(最後兩個維度) 權值不共享,l2 | 80.0 | – | – |
使用part_att,gloab_att,權值不共享,l2 | 80.2 | – | – |
權值不共享 weight*gloal_weight: | 84.9 | 94.1 | – |
weight*gloal_weight: | 87.6 | 94.9最高95 | 98 |
調參backbone*0.1: | 不收斂______________ | ———————— | ——————————– |
layer0,1 * 0.5 layer2 * 0.7layer3 * 0.9layer4 * 1other * 1.2: | 不收斂 | – | – |
將part_q-global_q 做交叉熵損失求co_loss與 avg_loss相加:50*co_loss + avg_loss + [max(q) - max(p)] | 68.2 | 85.6 | 95.6 |
將part_q-global_q 做交叉熵損失求co_loss與 avg_loss相加:10*co_loss + avg_loss + [max(q-p] | 72.8 | 88.4 | 96.3 |
將part_q-global_q 做交叉熵損失求co_loss與 avg_loss相加;div_sqrt(k), diff用fusionconv處理:co_loss + avg_loss | 87.6 | 94.9 | 98.3 |
weight*gloab_weight改變weight生成維度由一維變爲與k維度相同 | 80.0 | 92.4 | 97.7 |
將part_q-global_q 做交叉熵損失求co_loss與 avg_loss相加;div_sqrt(k), diff用fusionconv處理:co_loss + avg_loss,q處理同上 | 86.7(200epoch) | 94.8 | 98.5 |
將part_q-global_q 做交叉熵損失求co_loss*1.6與 avg_loss相加;weight後加bn;div_sqrt(k), diff用fusionconv處理:co_loss + avg_loss | 86.7(epoch200) | 94.7 | 98.5 |
weight * gloal_weight 特徵mean+max獨立分支求loss:co_loss*1.6 weight: | 88.3 | 95.1 | 98.5 |
weight * gloal_weight獨立分支求loss:co_loss*1.6 | 88.5 | 95.3 | 98.5 |
將part_q-global_q 做交叉熵損失求co_loss*1.6與 avg_loss相加;div_sqrt(k), diff用fusionconv處理:co_loss + avg_loss | 87.3 | 94.9 | 98.4 |
weight * gloal_weight獨立分支求loss:co_loss*1.6;mask_att(0.75, 0.5) | 86.7 | 94.4 | 98.3 |
weight * gloal_weight獨立分支求loss:co_loss*1.6;mask_att自學習參數 | 86.3 | 94.4 | 98.5 |
weight * gloal_weight獨立分支求loss:co_loss*1.6; test時cat(feat, diff) | 89.1 | 95.5 | 98.6 |
weight * gloal_weight獨立分支求loss:co_loss*1.6,同上兩個weight相乘後再做softmax | 88.4 | 95.3 | 98.5 |
同上+cam; cam_loss*1.6 | 88.9 | 95.6 | 98.5 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1) | 89.5 | 95.5 | 98.4 |
weight * gloal_weight獨立分支求loss:co_loss*1.6; test時cat(feat, diff),加上mask(thr=0.25, down=0.6, up=1.2) | 88.6 | 95.1 | 98.4 |
weight * gloal_weight獨立分支求loss:co _loss1.6;(mask後加softmax)cam_loss0.75,加上mask(thr=0.5, down=0.6, up=1.2) | 89.3(240) | 95.6 | 98.6 |
weight * gloal_weight獨立分支求loss:co_loss * 1.6 ,cam_loss * 0.75 | 89.3(200) | 95.3 | 98.6 |
weight * gloal_weight獨立分支求triple_loss(係數爲1, 0.875) | 89.1(200) | 95.5 | 98.5 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1)(mask在dim=2進行 standard_normal sigmoid)mask(thr=0.5, down=0.6, up=1.2) | 89.0(200) | 95.6 | 98.6 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1)(mask在max-min normal)mask(thr=0.5, down=0.6, up=1.2) | 89.5 | 95.5 | 98.8 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1)(mask在max-min normal)mask(thr=0.5, down=0.2, up=1.2) | 89.3 | 95.5 | 98.7 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1)(mask在max-min normal)mask(thr=0.75, down=0.2, up=1.2) | 89.5 | 95.5 | () |
weight * gloal_weight獨立分支求triple_loss(係數都爲1)(mask在max-min normal)mask(thr=0.75, down=0.2, up=1.6) | 89.2 | 95.1 | 98.4 |
weight * gloal_weight獨立分支求triple_loss + cam_loss(係數爲1,1, 0.17) | 89.4 | 95.4 | 98.5 |
weight * gloal_weight獨立分支求triple_loss + cam_loss(係數爲1,1, 0.17)(cam增加捲積層,test前將att_cls進行bn) | 89.3 | 95.5 | 98.3 |
weight * gloal_weight獨立分支求triple_loss + cam_loss(係數都爲1, 0.17)(mask在max-min normal)mask(thr=0.25, down=0.6, up=1.2) | 89.1(200) | 95.3 | 98.6 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1)(故障)(mask在max-min normal)mask(thr=0.25, down=0.6, up=1.2) | |||
取消self.training,weight * gloal_weight獨立分支求triple_loss(係數都爲1) (thr=0.5,de_rate=0.6, in_rate=0, change_rate=0.75 ) | (220)89.3 | 95.2 | 98.6 |
取消self.training,weight * gloal_weight獨立分支求triple_loss(係數都爲1) (thr=0.5,de_rate=1, in_rate=0, change_rate=0.75 ) | 89.2 | 95.4 | 98.5 |
取消self.training,weight * gloal_weight獨立分支求triple_loss(係數都爲1) (thr=0.75,de_rate=1, in_rate=0, change_rate=0.75 ) | 89.5 | 95.3 | 98.6 |
取消self.training,weight * gloal_weight獨立分支求triple_loss(係數都爲1) (thr=0.75,de_rate=0.6, in_rate=1.2, change_rate=1) | 89.3 | 95.4 | 98.6 |
取消self.training, weight * gloal_weight獨立分支求triple_loss(係數都爲1) (thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=0.75 ) | 89.3 | 95.7 | 98.5 |
weight * gloal_weight獨立分支求triple_loss(係數都爲1) (thr=0.75,de_rate=1, in_rate=0, change_rate=0.75 ) | 89.5 | 95.3 | 98.6 |
加入gloab_att+mask(thr=0.75,de_rate=1, in_rate=0, change_rate=0.75 ) | 90.2 | 95.8 | 98.5 |
加入gloab_att+mask(thr=0.5,de_rate=1, in_rate=0, change_rate=0.75 ) | 90.1 | 95.9 | 98.5 |
gloab_att,part_att每一塊求loss | 效果差 | ||
加入gloab_att不進行mask | 90.0 | 95.4 | 98.7 |
加入gloab_att+mask(thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=1 ) | 90.2 | 96.0 | 98.4 |
加入gloab_att(正則化係數爲0.01) | 89.9 | 95.8 | 98.6 |
加入gloab_att(正則化係數爲0.1) | 90.2 | 95.9 | 98.4(收斂慢300epoch) |
加入gloab_att(正則化係數爲0.0005) | |||
加入gloab_att(正則化係數爲0.001) | 90.2 | 95.8 | 98.7(收斂快200epoch) |
加入gloab_att(正則化係數爲0.002) | |||
加入gloab_att(正則化係數爲0.004) | 90.0 | 95.6 | 98.4 |
加入gloab_att(正則化係數爲0.006) | 90.1 | 95.9 | 98.6 |
加入gloab_att(正則化係數爲0.008) | 90.2 | 95.7 | 98.6 |
加入gloab_att(正則化係數爲0.008partreg) | 90.1 | 95.6 | 98.6 |
加入gloab_att, two_part_att | 82.8 | 92.5 | 97.3 |
加入gloab_att+mask(thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=1 ), two_part_att+mask(thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=1 ) | |||
(dukemtmc)加入gloab_att, mask(thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=0.75 ) | 80.4 | 89.9 | 95.4 |
(cuhk03-detected)加入gloab_att, mask(thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=1 ) | 73.5 | 74.9 | 89.1 |
(cuhk03-detected)加入gloab_att, mask(thr=0.5,de_rate=0.6, in_rate=1.2, change_rate=1 ) | 76.9 | 78.9 | 90.7 |
base no (random erasing ;lable smoooth center) | 84.6 | 94.5 | 97.9 |