init pretrained weights
HanwenCao opened this issue · 16 comments
In training, the pretrained weights of resnet50 seem to be loaded to resnet101. Here are the logs:
Initializing model: resnet101
Downloading: "https://download.pytorch.org/models/resnet50-19c8e357.pth"
I was wondering why not use resnet101's pretrained weights to initialize resnet101? Thank you!
Hi Hanwen, I thought it a strategy for training. We might release specific illustrations in another repo in the future. Thanks!
Thank you, Kevin. I only achieve mAP: 20.4%
and Rank-1 : 34.4%
after training the Re-id model. (almost the same for epoch 9-25) Am I doing right? Thank you
Hi, I further load the weights you provided in this repo and it can achieve mAP: 82.4%
and Rank-1 : 92.2%
, which is much better than my weights.
Hi Hanwen,
I thought mAP: 20.4% and Rank-1: 34.4% seems to have some problems. Now that the hosts of AiCity can successfully reproduce our training procedure, I assume some problems happened in your training procedure. I think it a great idea to try our provided model first. BTW, a star can give us more motivation to continue releasing the code.
Sorry, just notice you have tried our provided model. Then, did u make any modifications to the training script that we provide?
Sorry, just notice you have tried our provided model. Then, did u make any modifications to the training script that we provide?
Oh yes. Due to the memory limitation of my machine, I use a smaller train_bath_size, i.e. 64, rather than 128.
I will create a new project and try to use the exact same training script that you provided to train, including the batch size, to see if that can help. I will let you know if I am done. Thank you for your reply!
Hi Kevin,
calculate_mean_and_std is slow. Can I comment it out?
Besides, I have noticed that in build_transforms fixed values are placed. So looks like commenting the above line won't influence.
I further spend some time to calculate_mean_and_std
. I then got mean and std: tensor([-0.4599, -0.2843, 0.1362]) tensor([1.0227, 1.0484, 0.9768])
. But the fixed values in your code are not equal to the mean and std calculated by calculate_mean_and_std
. Does it matter?
Hi Kevin,
This time I use the same training code and args with the modification listed below:
- I comment calculate_mean_and_std as talked before
- I download resnet50-19c8e357.pth to a local path and modify the line at here by
pretrain_dict = model_zoo.load_url(model_url,model_dir='/model/caohw9/track3_model')
. The goal is trying to prevent downloading by using the local weights file. Not sure if it is correct though. - no I think that's all...
But I still cannot achieve a high mAP and rank-1. Here is the log.
Currently using GPU 0
Initializing image data manager
=> Initializing TRAIN (source) datasets
=> Initializing TEST datasets
**************** Summary ****************
train names : ['Aic']
# train datasets : 1
# train ids : 521
# train images : 206059
# train cameras : 36
test names : Aic
*****************************************
Initializing model: resnet101
Model size: 42.500 M
=> Start training
Epoch: [1][100/1609] Time 0.535 (0.958) Data 0.0008 (0.3947) Xent 5.9572 (6.0629) Htri 0.1422 (0.2513) Acc 2.34 (1.87)
Epoch: [1][200/1609] Time 0.929 (0.904) Data 0.4190 (0.3739) Xent 5.9288 (6.0237) Htri 0.1243 (0.1930) Acc 2.34 (1.98)
Epoch: [1][300/1609] Time 0.568 (0.889) Data 0.0517 (0.3701) Xent 5.8334 (5.9870) Htri 0.1691 (0.1704) Acc 1.56 (2.15)
Epoch: [1][400/1609] Time 0.797 (0.882) Data 0.3114 (0.3688) Xent 4.8599 (5.8658) Htri 0.2285 (0.1915) Acc 11.72 (2.77)
Epoch: [1][500/1609] Time 0.496 (0.885) Data 0.0011 (0.3762) Xent 3.9280 (5.5799) Htri 0.2541 (0.2335) Acc 12.50 (4.42)
Epoch: [1][600/1609] Time 0.471 (0.884) Data 0.0005 (0.3783) Xent 3.2735 (5.2284) Htri 0.4535 (0.2697) Acc 21.09 (7.15)
Epoch: [1][700/1609] Time 0.489 (0.880) Data 0.0002 (0.3770) Xent 2.3923 (4.8827) Htri 0.3877 (0.2877) Acc 37.50 (10.67)
Epoch: [1][800/1609] Time 0.479 (0.888) Data 0.0005 (0.3876) Xent 2.1850 (4.5608) Htri 0.2014 (0.2936) Acc 41.41 (14.51)
Epoch: [1][900/1609] Time 0.471 (0.893) Data 0.0002 (0.3944) Xent 1.8294 (4.2687) Htri 0.4545 (0.2962) Acc 53.12 (18.39)
Epoch: [1][1000/1609] Time 0.476 (0.893) Data 0.0002 (0.3958) Xent 1.6893 (4.0037) Htri 0.2295 (0.2933) Acc 59.38 (22.28)
Epoch: [1][1100/1609] Time 0.480 (0.888) Data 0.0002 (0.3924) Xent 1.3831 (3.7635) Htri 0.1999 (0.2868) Acc 61.72 (26.02)
Epoch: [1][1200/1609] Time 0.476 (0.882) Data 0.0005 (0.3880) Xent 1.0735 (3.5479) Htri 0.2118 (0.2791) Acc 71.88 (29.57)
Epoch: [1][1300/1609] Time 0.475 (0.876) Data 0.0002 (0.3831) Xent 0.9743 (3.3559) Htri 0.1204 (0.2721) Acc 71.88 (32.82)
Epoch: [1][1400/1609] Time 0.469 (0.874) Data 0.0002 (0.3815) Xent 1.0831 (3.1823) Htri 0.2948 (0.2645) Acc 73.44 (35.83)
Epoch: [1][1500/1609] Time 0.491 (0.873) Data 0.0004 (0.3808) Xent 0.8309 (3.0257) Htri 0.3563 (0.2576) Acc 77.34 (38.61)
Epoch: [1][1600/1609] Time 0.473 (0.869) Data 0.0002 (0.3774) Xent 0.7550 (2.8854) Htri 0.2194 (0.2505) Acc 76.56 (41.11)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.101/256
Computing CMC and mAP
Results ----------
mAP: 11.6%
CMC curve
Rank-1 : 18.9%
Rank-5 : 30.0%
Rank-10 : 34.7%
Rank-20 : 43.1%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-1"
Epoch: [2][100/1609] Time 0.494 (0.491) Data 0.0016 (0.0090) Xent 0.7822 (0.6875) Htri 0.1342 (0.1303) Acc 79.69 (81.63)
Epoch: [2][200/1609] Time 0.474 (0.484) Data 0.0007 (0.0048) Xent 0.7326 (0.6587) Htri 0.0875 (0.1248) Acc 81.25 (82.33)
Epoch: [2][300/1609] Time 0.478 (0.482) Data 0.0007 (0.0034) Xent 0.5381 (0.6413) Htri 0.1281 (0.1220) Acc 85.94 (82.78)
Epoch: [2][400/1609] Time 0.468 (0.480) Data 0.0005 (0.0027) Xent 0.5125 (0.6274) Htri 0.0469 (0.1209) Acc 85.94 (83.20)
Epoch: [2][500/1609] Time 0.458 (0.480) Data 0.0005 (0.0023) Xent 0.5749 (0.6053) Htri 0.1293 (0.1156) Acc 85.16 (83.85)
Epoch: [2][600/1609] Time 0.473 (0.479) Data 0.0006 (0.0020) Xent 0.4393 (0.5843) Htri 0.1207 (0.1106) Acc 89.06 (84.45)
Epoch: [2][700/1609] Time 0.467 (0.479) Data 0.0007 (0.0018) Xent 0.3883 (0.5684) Htri 0.0429 (0.1071) Acc 89.06 (84.85)
Epoch: [2][800/1609] Time 0.467 (0.479) Data 0.0005 (0.0017) Xent 0.4700 (0.5511) Htri 0.3527 (0.1044) Acc 89.06 (85.33)
Epoch: [2][900/1609] Time 0.471 (0.479) Data 0.0005 (0.0016) Xent 0.4654 (0.5371) Htri 0.0309 (0.1025) Acc 87.50 (85.77)
Epoch: [2][1000/1609] Time 0.467 (0.479) Data 0.0004 (0.0020) Xent 0.3293 (0.5261) Htri 0.0931 (0.1004) Acc 92.19 (86.06)
Epoch: [2][1100/1609] Time 0.478 (0.479) Data 0.0004 (0.0018) Xent 0.4022 (0.5144) Htri 0.0634 (0.0978) Acc 89.84 (86.41)
Epoch: [2][1200/1609] Time 0.481 (0.479) Data 0.0007 (0.0017) Xent 0.3909 (0.4990) Htri 0.1587 (0.0945) Acc 90.62 (86.87)
Epoch: [2][1300/1609] Time 0.462 (0.479) Data 0.0006 (0.0016) Xent 0.5039 (0.4885) Htri 0.1012 (0.0930) Acc 91.41 (87.17)
Epoch: [2][1400/1609] Time 0.471 (0.479) Data 0.0004 (0.0016) Xent 0.2532 (0.4781) Htri 0.0221 (0.0903) Acc 96.09 (87.46)
Epoch: [2][1500/1609] Time 0.482 (0.479) Data 0.0005 (0.0015) Xent 0.2285 (0.4668) Htri 0.1213 (0.0880) Acc 95.31 (87.79)
Epoch: [2][1600/1609] Time 0.476 (0.479) Data 0.0005 (0.0015) Xent 0.2717 (0.4564) Htri 0.1202 (0.0858) Acc 94.53 (88.09)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.065/256
Computing CMC and mAP
Results ----------
mAP: 16.7%
CMC curve
Rank-1 : 27.1%
Rank-5 : 37.8%
Rank-10 : 44.9%
Rank-20 : 51.6%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-2"
Epoch: [3][100/1609] Time 0.472 (0.485) Data 0.0005 (0.0083) Xent 0.4068 (0.2817) Htri 0.1156 (0.0521) Acc 87.50 (93.03)
Epoch: [3][200/1609] Time 0.474 (0.480) Data 0.0005 (0.0047) Xent 0.2807 (0.2778) Htri 0.0407 (0.0474) Acc 92.19 (93.11)
Epoch: [3][300/1609] Time 0.470 (0.479) Data 0.0004 (0.0034) Xent 0.2790 (0.2835) Htri 0.0600 (0.0518) Acc 96.09 (93.01)
Epoch: [3][400/1609] Time 0.469 (0.478) Data 0.0005 (0.0028) Xent 0.2447 (0.2847) Htri 0.0557 (0.0494) Acc 93.75 (93.03)
Epoch: [3][500/1609] Time 0.473 (0.478) Data 0.0006 (0.0024) Xent 0.2447 (0.2837) Htri 0.0168 (0.0493) Acc 92.19 (93.08)
Epoch: [3][600/1609] Time 0.484 (0.478) Data 0.0007 (0.0021) Xent 0.1971 (0.2825) Htri 0.0637 (0.0497) Acc 96.09 (93.10)
Epoch: [3][700/1609] Time 0.483 (0.478) Data 0.0005 (0.0020) Xent 0.2601 (0.2805) Htri 0.0575 (0.0504) Acc 94.53 (93.13)
Epoch: [3][800/1609] Time 0.480 (0.478) Data 0.0015 (0.0018) Xent 0.2517 (0.2750) Htri 0.0000 (0.0492) Acc 93.75 (93.28)
Epoch: [3][900/1609] Time 0.472 (0.478) Data 0.0004 (0.0017) Xent 0.2213 (0.2711) Htri 0.0061 (0.0478) Acc 94.53 (93.39)
Epoch: [3][1000/1609] Time 0.483 (0.478) Data 0.0005 (0.0017) Xent 0.2835 (0.2680) Htri 0.0141 (0.0470) Acc 92.19 (93.47)
Epoch: [3][1100/1609] Time 0.486 (0.478) Data 0.0007 (0.0016) Xent 0.3539 (0.2652) Htri 0.0309 (0.0463) Acc 86.72 (93.55)
Epoch: [3][1200/1609] Time 0.473 (0.478) Data 0.0007 (0.0015) Xent 0.2219 (0.2626) Htri 0.0154 (0.0463) Acc 92.97 (93.63)
Epoch: [3][1300/1609] Time 0.481 (0.478) Data 0.0013 (0.0015) Xent 0.1932 (0.2603) Htri 0.0737 (0.0456) Acc 96.09 (93.70)
Epoch: [3][1400/1609] Time 0.477 (0.478) Data 0.0011 (0.0014) Xent 0.2690 (0.2577) Htri 0.0260 (0.0451) Acc 92.97 (93.77)
Epoch: [3][1500/1609] Time 0.487 (0.477) Data 0.0007 (0.0014) Xent 0.2305 (0.2546) Htri 0.0673 (0.0442) Acc 93.75 (93.86)
Epoch: [3][1600/1609] Time 0.478 (0.477) Data 0.0016 (0.0014) Xent 0.3577 (0.2523) Htri 0.0081 (0.0439) Acc 92.19 (93.92)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.066/256
Computing CMC and mAP
Results ----------
mAP: 17.6%
CMC curve
Rank-1 : 28.9%
Rank-5 : 39.3%
Rank-10 : 47.6%
Rank-20 : 54.4%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-3"
Epoch: [4][100/1609] Time 0.468 (0.481) Data 0.0006 (0.0078) Xent 0.2021 (0.2164) Htri 0.0000 (0.0372) Acc 94.53 (94.84)
Epoch: [4][200/1609] Time 0.469 (0.478) Data 0.0007 (0.0043) Xent 0.1921 (0.1994) Htri 0.0000 (0.0316) Acc 96.09 (95.30)
Epoch: [4][300/1609] Time 0.471 (0.476) Data 0.0006 (0.0031) Xent 0.1502 (0.1921) Htri 0.0155 (0.0306) Acc 97.66 (95.54)
Epoch: [4][400/1609] Time 0.479 (0.476) Data 0.0017 (0.0025) Xent 0.1193 (0.1932) Htri 0.0147 (0.0315) Acc 96.88 (95.51)
Epoch: [4][500/1609] Time 0.473 (0.476) Data 0.0008 (0.0022) Xent 0.1620 (0.1916) Htri 0.0000 (0.0314) Acc 95.31 (95.58)
Epoch: [4][600/1609] Time 0.461 (0.475) Data 0.0006 (0.0020) Xent 0.2531 (0.1926) Htri 0.1143 (0.0315) Acc 95.31 (95.55)
Epoch: [4][700/1609] Time 0.474 (0.475) Data 0.0014 (0.0018) Xent 0.2256 (0.1954) Htri 0.0000 (0.0314) Acc 95.31 (95.48)
Epoch: [4][800/1609] Time 0.468 (0.475) Data 0.0005 (0.0017) Xent 0.1945 (0.1961) Htri 0.0109 (0.0314) Acc 93.75 (95.45)
Epoch: [4][900/1609] Time 0.479 (0.475) Data 0.0008 (0.0016) Xent 0.1585 (0.1965) Htri 0.0000 (0.0312) Acc 98.44 (95.45)
Epoch: [4][1000/1609] Time 0.476 (0.475) Data 0.0009 (0.0015) Xent 0.2160 (0.1945) Htri 0.0146 (0.0308) Acc 95.31 (95.52)
Epoch: [4][1100/1609] Time 0.481 (0.475) Data 0.0007 (0.0015) Xent 0.1940 (0.1922) Htri 0.0095 (0.0304) Acc 93.75 (95.59)
Epoch: [4][1200/1609] Time 0.469 (0.475) Data 0.0006 (0.0014) Xent 0.1885 (0.1898) Htri 0.0202 (0.0297) Acc 96.09 (95.66)
Epoch: [4][1300/1609] Time 0.498 (0.475) Data 0.0010 (0.0014) Xent 0.1664 (0.1871) Htri 0.0937 (0.0288) Acc 95.31 (95.73)
Epoch: [4][1400/1609] Time 0.470 (0.475) Data 0.0004 (0.0014) Xent 0.1845 (0.1849) Htri 0.0693 (0.0283) Acc 96.88 (95.81)
Epoch: [4][1500/1609] Time 0.471 (0.475) Data 0.0005 (0.0013) Xent 0.1982 (0.1829) Htri 0.0058 (0.0279) Acc 96.09 (95.87)
Epoch: [4][1600/1609] Time 0.457 (0.475) Data 0.0007 (0.0013) Xent 0.2275 (0.1832) Htri 0.0244 (0.0277) Acc 96.09 (95.86)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.066/256
Computing CMC and mAP
Results ----------
mAP: 15.7%
CMC curve
Rank-1 : 23.3%
Rank-5 : 36.4%
Rank-10 : 43.8%
Rank-20 : 51.8%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-4"
Epoch: [5][100/1609] Time 0.493 (0.483) Data 0.0006 (0.0090) Xent 0.2329 (0.1396) Htri 0.0873 (0.0192) Acc 96.88 (97.27)
Epoch: [5][200/1609] Time 0.465 (0.478) Data 0.0008 (0.0049) Xent 0.1902 (0.1435) Htri 0.0069 (0.0209) Acc 93.75 (97.09)
Epoch: [5][300/1609] Time 0.480 (0.476) Data 0.0016 (0.0035) Xent 0.1616 (0.1495) Htri 0.0021 (0.0216) Acc 97.66 (96.93)
Epoch: [5][400/1609] Time 0.481 (0.475) Data 0.0007 (0.0029) Xent 0.1752 (0.1549) Htri 0.0020 (0.0220) Acc 96.88 (96.79)
Epoch: [5][500/1609] Time 0.474 (0.476) Data 0.0015 (0.0025) Xent 0.1421 (0.1529) Htri 0.0000 (0.0213) Acc 95.31 (96.81)
Epoch: [5][600/1609] Time 0.463 (0.475) Data 0.0007 (0.0022) Xent 0.1611 (0.1503) Htri 0.0000 (0.0202) Acc 95.31 (96.82)
Epoch: [5][700/1609] Time 0.478 (0.475) Data 0.0005 (0.0020) Xent 0.2193 (0.1515) Htri 0.0058 (0.0208) Acc 94.53 (96.79)
Epoch: [5][800/1609] Time 0.461 (0.475) Data 0.0007 (0.0019) Xent 0.1331 (0.1519) Htri 0.0000 (0.0208) Acc 98.44 (96.77)
Epoch: [5][900/1609] Time 0.471 (0.475) Data 0.0009 (0.0018) Xent 0.1158 (0.1545) Htri 0.0337 (0.0213) Acc 100.00 (96.71)
Epoch: [5][1000/1609] Time 0.476 (0.475) Data 0.0007 (0.0017) Xent 0.1043 (0.1586) Htri 0.0021 (0.0225) Acc 99.22 (96.59)
Epoch: [5][1100/1609] Time 0.456 (0.475) Data 0.0006 (0.0016) Xent 0.0808 (0.1576) Htri 0.0242 (0.0222) Acc 97.66 (96.60)
Epoch: [5][1200/1609] Time 0.469 (0.475) Data 0.0007 (0.0015) Xent 0.1015 (0.1570) Htri 0.0148 (0.0221) Acc 99.22 (96.62)
Epoch: [5][1300/1609] Time 0.485 (0.475) Data 0.0011 (0.0015) Xent 0.1049 (0.1544) Htri 0.0035 (0.0214) Acc 98.44 (96.70)
Epoch: [5][1400/1609] Time 0.495 (0.475) Data 0.0009 (0.0014) Xent 0.0804 (0.1526) Htri 0.0000 (0.0210) Acc 99.22 (96.74)
Epoch: [5][1500/1609] Time 0.488 (0.478) Data 0.0008 (0.0030) Xent 0.1842 (0.1524) Htri 0.0184 (0.0209) Acc 95.31 (96.76)
Epoch: [5][1600/1609] Time 0.478 (0.479) Data 0.0004 (0.0028) Xent 0.1615 (0.1519) Htri 0.0000 (0.0207) Acc 98.44 (96.77)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.058/256
Computing CMC and mAP
Results ----------
mAP: 17.5%
CMC curve
Rank-1 : 31.1%
Rank-5 : 41.1%
Rank-10 : 47.8%
Rank-20 : 55.1%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-5"
Epoch: [6][100/1609] Time 0.480 (0.513) Data 0.0008 (0.0288) Xent 0.1906 (0.1312) Htri 0.0000 (0.0140) Acc 96.09 (97.11)
Epoch: [6][200/1609] Time 0.496 (0.500) Data 0.0015 (0.0149) Xent 0.1652 (0.1264) Htri 0.0000 (0.0140) Acc 95.31 (97.36)
Epoch: [6][300/1609] Time 0.468 (0.497) Data 0.0006 (0.0102) Xent 0.0757 (0.1267) Htri 0.0000 (0.0139) Acc 97.66 (97.34)
Epoch: [6][400/1609] Time 0.478 (0.498) Data 0.0008 (0.0086) Xent 0.0886 (0.1294) Htri 0.0000 (0.0157) Acc 99.22 (97.29)
Epoch: [6][500/1609] Time 0.492 (0.496) Data 0.0007 (0.0070) Xent 0.1583 (0.1292) Htri 0.0885 (0.0156) Acc 96.88 (97.31)
Epoch: [6][600/1609] Time 0.477 (0.498) Data 0.0009 (0.0088) Xent 0.1911 (0.1343) Htri 0.0000 (0.0174) Acc 94.53 (97.17)
Epoch: [6][700/1609] Time 0.481 (0.497) Data 0.0008 (0.0077) Xent 0.1532 (0.1385) Htri 0.0003 (0.0181) Acc 96.88 (97.07)
Epoch: [6][800/1609] Time 0.488 (0.496) Data 0.0010 (0.0068) Xent 0.1321 (0.1380) Htri 0.0115 (0.0184) Acc 96.88 (97.10)
Epoch: [6][900/1609] Time 0.471 (0.497) Data 0.0008 (0.0081) Xent 0.1602 (0.1420) Htri 0.0024 (0.0191) Acc 97.66 (96.99)
Epoch: [6][1000/1609] Time 0.483 (0.496) Data 0.0009 (0.0074) Xent 0.2067 (0.1429) Htri 0.0051 (0.0196) Acc 95.31 (96.94)
Epoch: [6][1100/1609] Time 0.497 (0.495) Data 0.0013 (0.0068) Xent 0.1295 (0.1423) Htri 0.0000 (0.0194) Acc 97.66 (96.97)
Epoch: [6][1200/1609] Time 0.484 (0.494) Data 0.0012 (0.0063) Xent 0.1279 (0.1419) Htri 0.0000 (0.0196) Acc 97.66 (96.99)
Epoch: [6][1300/1609] Time 0.483 (0.494) Data 0.0005 (0.0059) Xent 0.1385 (0.1429) Htri 0.0609 (0.0196) Acc 98.44 (96.97)
Epoch: [6][1400/1609] Time 0.473 (0.493) Data 0.0006 (0.0055) Xent 0.1856 (0.1440) Htri 0.0277 (0.0202) Acc 95.31 (96.94)
Epoch: [6][1500/1609] Time 0.477 (0.493) Data 0.0008 (0.0052) Xent 0.0706 (0.1445) Htri 0.0000 (0.0201) Acc 100.00 (96.92)
Epoch: [6][1600/1609] Time 0.479 (0.493) Data 0.0006 (0.0049) Xent 0.1654 (0.1431) Htri 0.0049 (0.0199) Acc 95.31 (96.96)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.060/256
Computing CMC and mAP
Results ----------
mAP: 20.6%
CMC curve
Rank-1 : 32.7%
Rank-5 : 43.8%
Rank-10 : 50.9%
Rank-20 : 57.6%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-6"
Epoch: [7][100/1609] Time 0.477 (0.522) Data 0.0007 (0.0379) Xent 0.1657 (0.1275) Htri 0.0000 (0.0193) Acc 97.66 (97.34)
Epoch: [7][200/1609] Time 0.504 (0.513) Data 0.0008 (0.0263) Xent 0.1211 (0.1179) Htri 0.0108 (0.0164) Acc 97.66 (97.54)
Epoch: [7][300/1609] Time 0.479 (0.504) Data 0.0008 (0.0180) Xent 0.1161 (0.1131) Htri 0.0510 (0.0151) Acc 97.66 (97.67)
Epoch: [7][400/1609] Time 0.483 (0.501) Data 0.0008 (0.0137) Xent 0.0833 (0.1153) Htri 0.0000 (0.0151) Acc 97.66 (97.64)
Epoch: [7][500/1609] Time 0.481 (0.498) Data 0.0008 (0.0111) Xent 0.0466 (0.1123) Htri 0.0000 (0.0141) Acc 100.00 (97.72)
Epoch: [7][600/1609] Time 0.482 (0.496) Data 0.0008 (0.0094) Xent 0.1146 (0.1115) Htri 0.0000 (0.0143) Acc 97.66 (97.74)
Epoch: [7][700/1609] Time 0.501 (0.495) Data 0.0008 (0.0082) Xent 0.1027 (0.1168) Htri 0.0103 (0.0156) Acc 97.66 (97.65)
Epoch: [7][800/1609] Time 0.486 (0.494) Data 0.0011 (0.0073) Xent 0.1421 (0.1178) Htri 0.0080 (0.0158) Acc 96.88 (97.66)
Epoch: [7][900/1609] Time 0.479 (0.493) Data 0.0006 (0.0065) Xent 0.1317 (0.1210) Htri 0.0046 (0.0159) Acc 96.88 (97.59)
Epoch: [7][1000/1609] Time 0.540 (0.493) Data 0.0009 (0.0060) Xent 0.2222 (0.1244) Htri 0.0946 (0.0167) Acc 93.75 (97.50)
Epoch: [7][1100/1609] Time 0.481 (0.493) Data 0.0008 (0.0061) Xent 0.1009 (0.1264) Htri 0.0000 (0.0168) Acc 97.66 (97.45)
Epoch: [7][1200/1609] Time 0.486 (0.493) Data 0.0005 (0.0057) Xent 0.0915 (0.1251) Htri 0.0000 (0.0163) Acc 99.22 (97.48)
Epoch: [7][1300/1609] Time 0.479 (0.493) Data 0.0009 (0.0053) Xent 0.0919 (0.1233) Htri 0.0592 (0.0158) Acc 98.44 (97.53)
Epoch: [7][1400/1609] Time 0.491 (0.493) Data 0.0010 (0.0050) Xent 0.1308 (0.1236) Htri 0.0576 (0.0159) Acc 96.09 (97.52)
Epoch: [7][1500/1609] Time 0.491 (0.493) Data 0.0007 (0.0057) Xent 0.0793 (0.1234) Htri 0.0000 (0.0159) Acc 97.66 (97.53)
Epoch: [7][1600/1609] Time 0.473 (0.495) Data 0.0008 (0.0071) Xent 0.1075 (0.1234) Htri 0.0217 (0.0159) Acc 98.44 (97.52)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.062/256
Computing CMC and mAP
Results ----------
mAP: 19.9%
CMC curve
Rank-1 : 35.6%
Rank-5 : 46.0%
Rank-10 : 54.0%
Rank-20 : 62.4%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-7"
Epoch: [8][100/1609] Time 0.496 (0.508) Data 0.0012 (0.0221) Xent 0.1181 (0.1140) Htri 0.0055 (0.0133) Acc 97.66 (97.71)
Epoch: [8][200/1609] Time 0.477 (0.497) Data 0.0008 (0.0115) Xent 0.0810 (0.1008) Htri 0.0000 (0.0093) Acc 99.22 (98.06)
Epoch: [8][300/1609] Time 0.478 (0.495) Data 0.0007 (0.0080) Xent 0.0794 (0.0953) Htri 0.0000 (0.0086) Acc 97.66 (98.18)
Epoch: [8][400/1609] Time 0.477 (0.515) Data 0.0006 (0.0283) Xent 0.1256 (0.1017) Htri 0.0253 (0.0099) Acc 99.22 (98.06)
Epoch: [8][500/1609] Time 0.491 (0.510) Data 0.0009 (0.0228) Xent 0.1538 (0.1026) Htri 0.0080 (0.0105) Acc 96.09 (98.08)
Epoch: [8][600/1609] Time 0.474 (0.506) Data 0.0008 (0.0192) Xent 0.1723 (0.1070) Htri 0.0183 (0.0113) Acc 96.88 (97.98)
Epoch: [8][700/1609] Time 0.467 (0.504) Data 0.0006 (0.0166) Xent 0.2116 (0.1138) Htri 0.0284 (0.0128) Acc 94.53 (97.83)
Epoch: [8][800/1609] Time 0.476 (0.502) Data 0.0007 (0.0151) Xent 0.1798 (0.1174) Htri 0.0219 (0.0136) Acc 95.31 (97.73)
Epoch: [8][900/1609] Time 0.490 (0.507) Data 0.0008 (0.0195) Xent 0.1005 (0.1194) Htri 0.0300 (0.0144) Acc 98.44 (97.66)
Epoch: [8][1000/1609] Time 0.477 (0.506) Data 0.0011 (0.0190) Xent 0.0948 (0.1220) Htri 0.0218 (0.0145) Acc 97.66 (97.58)
Epoch: [8][1100/1609] Time 0.486 (0.505) Data 0.0008 (0.0178) Xent 0.1707 (0.1233) Htri 0.0000 (0.0145) Acc 97.66 (97.55)
Epoch: [8][1200/1609] Time 0.482 (0.505) Data 0.0005 (0.0181) Xent 0.0493 (0.1213) Htri 0.0141 (0.0140) Acc 100.00 (97.60)
Epoch: [8][1300/1609] Time 0.489 (0.504) Data 0.0008 (0.0168) Xent 0.0993 (0.1189) Htri 0.0000 (0.0135) Acc 96.88 (97.65)
Epoch: [8][1400/1609] Time 0.487 (0.503) Data 0.0006 (0.0164) Xent 0.1146 (0.1171) Htri 0.0342 (0.0133) Acc 96.88 (97.69)
Epoch: [8][1500/1609] Time 0.482 (0.504) Data 0.0008 (0.0167) Xent 0.1066 (0.1172) Htri 0.0000 (0.0133) Acc 98.44 (97.70)
Epoch: [8][1600/1609] Time 0.482 (0.505) Data 0.0008 (0.0180) Xent 0.1300 (0.1160) Htri 0.0093 (0.0131) Acc 96.88 (97.73)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.063/256
Computing CMC and mAP
Results ----------
mAP: 19.4%
CMC curve
Rank-1 : 34.2%
Rank-5 : 46.0%
Rank-10 : 51.3%
Rank-20 : 56.7%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-8"
Epoch: [9][100/1609] Time 0.473 (0.508) Data 0.0008 (0.0218) Xent 0.0968 (0.1007) Htri 0.0000 (0.0136) Acc 97.66 (98.18)
Epoch: [9][200/1609] Time 0.479 (0.497) Data 0.0009 (0.0113) Xent 0.0776 (0.1053) Htri 0.0222 (0.0130) Acc 97.66 (98.03)
Epoch: [9][300/1609] Time 0.497 (0.494) Data 0.0007 (0.0079) Xent 0.0915 (0.1062) Htri 0.0000 (0.0135) Acc 100.00 (98.10)
Epoch: [9][400/1609] Time 0.481 (0.508) Data 0.0007 (0.0215) Xent 0.0957 (0.1061) Htri 0.0727 (0.0124) Acc 97.66 (98.08)
Epoch: [9][500/1609] Time 0.481 (0.529) Data 0.0006 (0.0424) Xent 0.0862 (0.1057) Htri 0.0145 (0.0122) Acc 99.22 (98.06)
Epoch: [9][600/1609] Time 0.516 (0.531) Data 0.0235 (0.0439) Xent 0.1029 (0.1085) Htri 0.0091 (0.0128) Acc 96.88 (97.99)
Epoch: [9][700/1609] Time 0.482 (0.543) Data 0.0009 (0.0560) Xent 0.1049 (0.1077) Htri 0.0000 (0.0122) Acc 98.44 (98.01)
Epoch: [9][800/1609] Time 0.479 (0.543) Data 0.0005 (0.0562) Xent 0.1502 (0.1066) Htri 0.0000 (0.0116) Acc 98.44 (98.02)
Epoch: [9][900/1609] Time 0.487 (0.545) Data 0.0005 (0.0587) Xent 0.0813 (0.1068) Htri 0.0000 (0.0119) Acc 97.66 (98.03)
Epoch: [9][1000/1609] Time 0.483 (0.551) Data 0.0008 (0.0638) Xent 0.1831 (0.1103) Htri 0.0475 (0.0126) Acc 97.66 (97.94)
Epoch: [9][1100/1609] Time 0.488 (0.555) Data 0.0007 (0.0681) Xent 0.1301 (0.1130) Htri 0.0000 (0.0129) Acc 96.09 (97.87)
Epoch: [9][1200/1609] Time 0.478 (0.562) Data 0.0007 (0.0754) Xent 0.1528 (0.1153) Htri 0.0045 (0.0133) Acc 94.53 (97.80)
Epoch: [9][1300/1609] Time 0.484 (0.567) Data 0.0007 (0.0809) Xent 0.0579 (0.1145) Htri 0.0000 (0.0130) Acc 99.22 (97.81)
Epoch: [9][1400/1609] Time 0.505 (0.575) Data 0.0008 (0.0889) Xent 0.1193 (0.1132) Htri 0.0000 (0.0128) Acc 96.88 (97.83)
Epoch: [9][1500/1609] Time 0.476 (0.582) Data 0.0005 (0.0953) Xent 0.0750 (0.1127) Htri 0.0000 (0.0126) Acc 99.22 (97.84)
Epoch: [9][1600/1609] Time 0.476 (0.585) Data 0.0009 (0.0990) Xent 0.0740 (0.1121) Htri 0.0000 (0.0124) Acc 98.44 (97.85)
=> Validation
Evaluating Aic ...
Extracted features for query set, obtained 3028-by-2048 matrix
Extracted features for gallery set, obtained 20494-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.060/256
Computing CMC and mAP
Results ----------
mAP: 20.4%
CMC curve
Rank-1 : 34.9%
Rank-5 : 46.9%
Rank-10 : 54.2%
Rank-20 : 59.1%
------------------
Checkpoint saved to "/output/models/resnet101-Aic/model.pth.tar-9"
Hi, Hanwen. Thanks for providing the training log! Hmm, that's strange then. I do not think the mean&std or training batch&learning rate will influence the model that large.
Since we did not preserve the log files after the challenge, I will go through it again and compare it with your log.
Hi, Hanwen. Sorry for the late reply. I find out the reason. The linked model is our submission model which is trained on trn+val.
The data set will then be:
**************** Summary ****************
train names : ['Aic']
# train datasets : 1
# train ids : 666
# train images : 226553
# train cameras : 40
test names : Aic
*****************************************
Sorry for the misunderstanding. I will update it in the ReadMe doc.
Wow. Thanks for your reply. I will try. Thanks for help me out
I have added a notice in the ReadMe doc. Please let me know if u can reproduce the results then.
No problem. I will let you know as soon as I make it
I have added a notice in the ReadMe doc. Please let me know if u can reproduce the results then.
Hi Kevin. Thank you for your help. This time it seems much better. At the 9th epoch, I achieve
mAP: 77.3%
CMC curve
Rank-1 : 88.9%
Rank-5 : 93.1%
Rank-10 : 95.6%
Rank-20 : 96.9%
Looks pretty close it shown in the ReadMe. But after epoch 10 or 11, the scores seem higher. Why did you decide to use weights pth.tar-9, rather than like tar-11?
Hi, Hanwen. That's because we fix the epoch through the experience we gained in trn/val experiments for the final submission model.