loc_loss 误差这么小 为什么推理 跟踪突变这么大 718-》837

[wavelet2008]

2020-11-19 19:12:20,716-rk0-train.py#196] epoch: 14
[2020-11-19 19:12:20,716-rk0-train.py#201] epoch 14 lr 0.0009026601969102485
[2020-11-19 19:12:20,716-rk0-train.py#201] epoch 14 lr 0.0009026601969102485
[2020-11-19 19:12:25,057-rk0-train.py#249] Epoch: [14][15/625] lr: 0.001805
batch_time: 0.311095 (0.310400) data_time: 0.241359 (0.240799)
cls_loss: 0.002425 (0.008269) loc_loss: 0.116004 (0.124908)

0.93641865
x1=718 , y1=287 x2=76 , y2=109
269.2549228668213milliseconds

0.93656695
x1=837 , y1=337 x2=89 , y2=136

[ZhiyuanChen]

所以你的问题是？

[wavelet2008]

我加上自己数据，训练效果很差，跟踪效果很差，完全不对，rpn_head.loc.1.running_var 系数比大神提供的差距比较多。目前训练数据ytbb 中的train0000-train0019 加上自己数据，用 siamrpn_alex_dwxcorr 训练的 求大神指点

[ZhiyuanChen]

我无法从你所提供的信息中获得任何有价值的内容，比如说你的数据集是什么样的？训练使用了什么参数？是从什么开始训练的？跟踪效果怎么差了？

[wavelet2008]

大神好啊 训练参数脚本 /pysot/experiments/siamrpn_alex_dwxcorr_16gpu/config.yaml
META_ARC: "siamrpn_alex_dwxcorr"

BACKBONE:
TYPE: "alexnet"
KWARGS:
width_mult: 1.0
PRETRAINED: 'alexnet-bn.pth'
#PRETRAINED: 'checkpoint_e80.pth'
TRAIN_LAYERS: ['layer4', 'layer5']
TRAIN_EPOCH: 5
LAYERS_LR: 1.0

ADJUST:
ADJUST: False

RPN:
TYPE: 'DepthwiseRPN'
KWARGS:
anchor_num: 5
in_channels: 256
out_channels: 256

MASK:
MASK: False

ANCHOR:
STRIDE: 8
RATIOS: [0.33, 0.5, 1, 2, 3]
SCALES: [8]
ANCHOR_NUM: 5

TRACK:
TYPE: 'SiamRPNTracker'
PENALTY_K: 0.16
WINDOW_INFLUENCE: 0.40
LR: 0.30
EXEMPLAR_SIZE: 127
INSTANCE_SIZE: 287
BASE_SIZE: 0
CONTEXT_AMOUNT: 0.5

TRAIN:
EPOCH: 30
START_EPOCH: 0
BATCH_SIZE: 128
BASE_SIZE: 0
OUTPUT_SIZE: 17
BASE_LR: 0.005
CLS_WEIGHT: 10.
LOC_WEIGHT: 15.2
RESUME: ''
#RESUME: 'snapshot/checkpoint_e80.pth'
LR:
TYPE: 'log'
KWARGS:
start_lr: 0.01
end_lr: 0.0005
LR_WARMUP:
TYPE: 'step'
EPOCH: 5
KWARGS:
start_lr: 0.005
end_lr: 0.01
step: 1

DATASET:
NAMES:
#- 'VID'
- 'YOUTUBEBB'
#- 'COCO'
#- 'DET'

TEMPLATE:
    SHIFT: 4
    SCALE: 0.05
    BLUR: 0.0
    FLIP: 0.0
    COLOR: 1.0

SEARCH:
    SHIFT: 64
    SCALE: 0.18
    BLUR: 0.2
    FLIP: 0.0
    COLOR: 1.0

NEG: 0.05
GRAY: 0.0

训练数据 只用 ytbb 中的train0000-train0019
[2020-11-26 08:48:49,259-rk0-train.py#249] Epoch: [19][1970/3125] lr: 0.001974
batch_time: 2.171178 (2.176625) data_time: 1.866371 (1.904743)
cls_loss: 0.052419 (0.049440) loc_loss: 0.222597 (0.217926)
total_loss: 3.907659 (3.806876)
[2020-11-26 08:48:49,259-rk0-log_helper.py#105] Progress: 58220 / 93750 [62%], Speed: 2.177 s/iter, ETA 0:21:28 (D:H:M)

训练效果：
(pysot) ubuntu@ubuntu-X10SRA:~/pysot$ python tools/demo.py --config experiments/siamrpn_alex_dwxcorr_16gpu/config.yaml --snapshot experiments/siamrpn_alex_dwxcorr_16gpu/snapshot/checkpoint_e18.pth --video demo/21_head_video00035
[494, 126, 63, 85] 第一帧就错了很离谱
0.92190355
[858, 240, 63, 85]
0.96316814
[909, 65, 63, 85]
0.08671292
[929, 162, 60, 85]
0.9830604
[929, 5, 60, 85]
0.14619353
[929, 5, 60, 85]
0.9291736
[929, 5, 60, 85]
0.9297365
[929, 5, 60, 85]
0.9297072
[929, 5, 60, 85]
0.93052185
[929, 7, 60, 85]
0.92811483
[929, 8, 60, 85]
0.93009555
[929, 12, 60, 85]
0.9249779
[929, 10, 60, 85]
0.9232683
[929, 10, 60, 85]
0.92663795

求小样本训练方法与策略

[ZhiyuanChen]

……你至少给一个gif动图吧，给几个框我也没法想象你在跟啥啊