for triplet loss

[ssbilakeri]

is the python 2.7 and pytorch 1.0.0 are not supportive for tripletloss ??

[huanghoujing]

Hi, python 2.7 and pytorch 1.0.0 are absolutely ok to run this code.

By the way, do you run the program in Linux? I ran it in Ubuntu.

[ssbilakeri]

Hello sir, I too ran the code on ubuntu python 2.7 and pytorch 1.0.0

The code works fine with ID and PS loss but when triplet is added it give error . Why is it so ?

On Sun, 13 Dec 2020, 1:22 pm Houjing Huang, notifications@github.com wrote:

Hi, python 2.7 and pytorch 1.0.0 are absolutely ok to run this code.

By the way, do you run the program in Linux? I ran it in Ubuntu.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-743969812, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBK4QGEBTE73W7PTKB3SURXCJANCNFSM4UQGVKRA .

[ssbilakeri]

I'm getting below error when included triplet loss. Help me to resolve this issue

File "/home/padmashree/anaconda3/envs/myenv/lib/python2.7/runpy.py", line 174, in _run_module_as_main "main", fname, loader, pkg_name) File "/home/padmashree/anaconda3/envs/myenv/lib/python2.7/runpy.py", line 72, in _run_code exec code in run_globals File "/home/padmashree/project_dir/EANet2/package/optim/eanet_trainer.py", line 135, in trainer.train_phases() File "/home/padmashree/project_dir/EANet2/package/optim/eanet_trainer.py", line 126, in train_phases self.train() File "package/optim/reid_trainer.py", line 338, in train self.trainer.train_one_epoch(trial_run_steps=3 if cfg.trial_run else None) File "package/optim/trainer.py", line 36, in train_one_epoch self.train_one_step(batch) File "package/optim/trainer.py", line 24, in train_one_step pred = self.train_forward(batch) File "/home/padmashree/project_dir/EANet2/package/optim/eanet_trainer.py", line 102, in train_forward loss += self.loss_funcs[loss_cfg.name](reid_batch, pred, step=self.trainer.current_step)['loss'] File "package/loss/triplet_loss.py", line 124, in call res3 = self.calculate(torch.stack(pred['feat_list']), batch['label'], hard_type=hard_type) File "package/loss/triplet_loss.py", line 107, in calculate dist_mat = compute_dist(feat, feat, dist_type=cfg.dist_type) File "package/eval/torch_distance.py", line 49, in compute_dist dist = euclidean_dist(array1, array2) File "package/eval/torch_distance.py", line 25, in euclidean_dist xx = torch.pow(x, 2).sum(1, keepdim=True).expand(m, n) RuntimeError: expand(torch.cuda.FloatTensor{[9, 1, 256]}, size=[9, 9]): the number of sizes provided (2) must be greater or equal to the number of dimensions in the tensor (3)

On Sun, 13 Dec 2020, 2:08 pm Shavantrevva Bilakeri, ssbilakeri@gmail.com wrote:

Hello sir, I too ran the code on ubuntu python 2.7 and pytorch 1.0.0

The code works fine with ID and PS loss but when triplet is added it give error . Why is it so ?

On Sun, 13 Dec 2020, 1:22 pm Houjing Huang, notifications@github.com wrote:

Hi, python 2.7 and pytorch 1.0.0 are absolutely ok to run this code.

By the way, do you run the program in Linux? I ran it in Ubuntu.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-743969812, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBK4QGEBTE73W7PTKB3SURXCJANCNFSM4UQGVKRA .

[huanghoujing]

Hi, thank you for your feedback.

To run triplet loss, we have to

1. Use PK sampling for batch construction
2. Increase training epochs

I provide a setting file paper_configs/PAP_S_PS_Triplet_Loss_Market1501.txt, and a script script/exp/train_PAP_S_PS_Triplet_Loss_Market1501.sh, to train with both PS loss and triplet loss on Market1501.

Besides, I also fixed a mistake in part index when feeding feature to triplet loss. You can find the detail in this commit.

Now, you can run the script by

bash script/exp/train_PAP_S_PS_Triplet_Loss_Market1501.sh

The result I obtained is

M -> M      [mAP:  86.0%], [cmc1:  95.0%], [cmc5:  98.0%], [cmc10:  98.8%]
M -> C      [mAP:  11.0%], [cmc1:  12.1%], [cmc5:  24.1%], [cmc10:  31.6%]
M -> D      [mAP:  29.2%], [cmc1:  46.7%], [cmc5:  63.2%], [cmc10:  69.5%]

I hope it helps.

[ssbilakeri]

Thank you so much for ur informative response.

On Thu, 17 Dec 2020, 12:25 pm Houjing Huang, notifications@github.com wrote:

Hi, thank you for your feedback.

To run triplet loss, we have to

1. Use PK sampling for batch construction
2. Increase training epochs

I provide a setting file paper_configs/PAP_S_PS_Triplet_Loss_Market1501.txt, and a script script/exp/train_PAP_S_PS_Triplet_Loss_Market1501.sh, to train with both PS loss and triplet loss on Market1501.

Besides, I also fixed a mistake in part index when feeding feature to triplet loss. You can find the detail in this commit https://github.com/huanghoujing/EANet/commit/e46d49528428fbf0c54b93e557e009012c9a34b4 .

Now, you can run the script by

bash script/exp/train_PAP_S_PS_Triplet_Loss_Market1501.sh

The result I obtained is

M -> M [mAP: 86.0%], [cmc1: 95.0%], [cmc5: 98.0%], [cmc10: 98.8%] M -> C [mAP: 11.0%], [cmc1: 12.1%], [cmc5: 24.1%], [cmc10: 31.6%] M -> D [mAP: 29.2%], [cmc1: 46.7%], [cmc5: 63.2%], [cmc10: 69.5%]

I hope it helps.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-747250377, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBO2TJRUMOOQRQWC23TSVGTOLANCNFSM4UQGVKRA .

[ssbilakeri]

Hi sir, In your EANet paper did u train the model including triplet loss ?? Kindly reply

On Thu, 17 Dec 2020, 9:15 pm Shavantrevva Bilakeri, ssbilakeri@gmail.com wrote:

Thank you so much for ur informative response.

On Thu, 17 Dec 2020, 12:25 pm Houjing Huang, notifications@github.com wrote:

Hi, thank you for your feedback.

To run triplet loss, we have to

1. Use PK sampling for batch construction
2. Increase training epochs

I provide a setting file paper_configs/PAP_S_PS_Triplet_Loss_Market1501.txt, and a script script/exp/train_PAP_S_PS_Triplet_Loss_Market1501.sh, to train with both PS loss and triplet loss on Market1501.

Besides, I also fixed a mistake in part index when feeding feature to triplet loss. You can find the detail in this commit https://github.com/huanghoujing/EANet/commit/e46d49528428fbf0c54b93e557e009012c9a34b4 .

Now, you can run the script by

bash script/exp/train_PAP_S_PS_Triplet_Loss_Market1501.sh

The result I obtained is

M -> M [mAP: 86.0%], [cmc1: 95.0%], [cmc5: 98.0%], [cmc10: 98.8%] M -> C [mAP: 11.0%], [cmc1: 12.1%], [cmc5: 24.1%], [cmc10: 31.6%] M -> D [mAP: 29.2%], [cmc1: 46.7%], [cmc5: 63.2%], [cmc10: 69.5%]

I hope it helps.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-747250377, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBO2TJRUMOOQRQWC23TSVGTOLANCNFSM4UQGVKRA .

[huanghoujing]

Hi ssbilakeri, I did not use triplet loss in the paper.

[ssbilakeri]

Hello sir, I'm trying to re-produce your paper result with re-ranking unfortunately I'm not able to do it.Since I need your paper result with re-ranking applied to compare with my work .could u please do it for me. It will be a great help Thank you

On Mon, 21 Dec 2020, 9:33 am Houjing Huang, notifications@github.com wrote:

Hi ssbilakeri, I did not use triplet loss in the paper.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-748744501, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBOTSE45VUNQTKSVK5LSV3CIDANCNFSM4UQGVKRA .

[huanghoujing]

Hi, ssbilakeri, for which Table of the paper do you need the re-ranking score?

[ssbilakeri]

I need it for PAP_S_PS (where ID loss and segmentation loss is used). Kindly help in in this regard. I will be looking forward to your response. Thank you.

On Sun, 27 Dec 2020, 5:17 pm Houjing Huang, notifications@github.com wrote:

Hi, ssbilakeri, for which Table of the paper do you need the re-ranking score?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-751457930, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBKMJFZ7S2TPJ2AKWXDSW4NE3ANCNFSM4UQGVKRA .

[ssbilakeri]

Hi sir, when I run your code with re-ranking getting below results. kindly suggest what could be the problem.

I have attached the code with this mail. please help me . Thank you.

Extract Feature: 0%| | 0/106 [00:00<?, ? batches/s]Loaded pickle file /home/padmashree/project_dir/dataset/market1501/im_path_to_kpt.pklLoaded pickle file /home/padmashree/project_dir/dataset/market1501/im_path_to_kpt.pkl

Extract Feature: 100%|##########################################################| 106/106 [00:13<00:00, 7.75 batches/s] Extract Feature: 0%| | 0/498 [00:00<?, ? batches/s]Loaded pickle file /home/padmashree/project_dir/dataset/market1501/im_path_to_kpt.pkl Loaded pickle file /home/padmashree/project_dir/dataset/market1501/im_path_to_kpt.pkl Extract Feature: 100%|##########################################################| 498/498 [01:04<00:00, 7.75 batches/s] => Eval Statistics: dic.keys(): ['g_feat', 'q_feat', 'q_visible', 'q_label', 'q_cam', 'g_visible', 'g_label', 'g_cam'] dic['q_feat'].shape: (3368, 2304) dic['q_label'].shape: (3368,) dic['q_cam'].shape: (3368,) dic['g_feat'].shape: (15913, 2304) dic['g_label'].shape: (15913,) dic['g_cam'].shape: (15913,) M -> M [mAP: 1.5%], [cmc1: 7.3%], [cmc5: 14.5%], [cmc10: 19.3%] Extract Feature: 0%| | 0/44 [00:00<?, ? batches/s]Loaded pickle file /home/padmashree/project_dir/dataset/cuhk03_np_detected_jpg/im_path_to_kpt.pkl Loaded pickle file /home/padmashree/project_dir/dataset/cuhk03_np_detected_jpg/im_path_to_kpt.pkl Extract Feature: 100%|############################################################| 44/44 [00:05<00:00, 7.79 batches/s] Extract Feature: 0%| | 0/167 [00:00<?, ? batches/s]Loaded pickle file /home/padmashree/project_dir/dataset/cuhk03_np_detected_jpg/im_path_to_kpt.pkl Loaded pickle file /home/padmashree/project_dir/dataset/cuhk03_np_detected_jpg/im_path_to_kpt.pkl Extract Feature: 100%|##########################################################| 167/167 [00:20<00:00, 8.10 batches/s] => Eval Statistics: dic.keys(): ['g_feat', 'q_feat', 'q_visible', 'q_label', 'q_cam', 'g_visible', 'g_label', 'g_cam'] dic['q_feat'].shape: (1400, 2304) dic['q_label'].shape: (1400,) dic['q_cam'].shape: (1400,) dic['g_feat'].shape: (5332, 2304) dic['g_label'].shape: (5332,) dic['g_cam'].shape: (5332,) M -> C [mAP: 0.2%], [cmc1: 0.1%], [cmc5: 0.5%], [cmc10: 1.7%] Extract Feature: 0%| | 0/70 [00:00<?, ? batches/s]Loaded pickle file /home/padmashree/project_dir/dataset/duke/im_path_to_kpt.pkl Loaded pickle file /home/padmashree/project_dir/dataset/duke/im_path_to_kpt.pkl Extract Feature: 100%|############################################################| 70/70 [00:08<00:00, 7.90 batches/s] Extract Feature: 0%| | 0/552 [00:00<?, ? batches/s]Loaded pickle file /home/padmashree/project_dir/dataset/duke/im_path_to_kpt.pkl Loaded pickle file /home/padmashree/project_dir/dataset/duke/im_path_to_kpt.pkl Extract Feature: 100%|##########################################################| 552/552 [01:07<00:00, 8.15 batches/s] => Eval Statistics: dic.keys(): ['g_feat', 'q_feat', 'q_visible', 'q_label', 'q_cam', 'g_visible', 'g_label', 'g_cam'] dic['q_feat'].shape: (2228, 2304) dic['q_label'].shape: (2228,) dic['q_cam'].shape: (2228,) dic['g_feat'].shape: (17661, 2304) dic['g_label'].shape: (17661,) dic['g_cam'].shape: (17661,) M -> D [mAP: 0.3%], [cmc1: 1.2%], [cmc5: 2.8%], [cmc10: 4.3%]

On Sun, 27 Dec 2020, 6:38 pm Shavantrevva Bilakeri, ssbilakeri@gmail.com wrote:

I need it for PAP_S_PS (where ID loss and segmentation loss is used). Kindly help in in this regard. I will be looking forward to your response. Thank you.

On Sun, 27 Dec 2020, 5:17 pm Houjing Huang, notifications@github.com wrote:

Hi, ssbilakeri, for which Table of the paper do you need the re-ranking score?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-751457930, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBKMJFZ7S2TPJ2AKWXDSW4NE3ANCNFSM4UQGVKRA .

[huanghoujing]

It seems that the trained model weight is not loaded.

[ssbilakeri]

Could you please check with your code hope you have trained weights

On Mon, 28 Dec 2020, 8:26 pm Houjing Huang, notifications@github.com wrote:

It seems that the trained model weight is not loaded.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-751738171, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBIQXJJ6RGPQ2AOOLUTSXCMAXANCNFSM4UQGVKRA .

[huanghoujing]

Hi, ssbilakeri. I have tested the re-ranking results for PAP_S_PS models (the script to run this is script/exp/test_PAP_S_PS_reranking.sh).

The original scores, as well as the re-ranking scores are as follows.

	mAP	Rank-1	Rank-5	Rank-10
M -> M	85.6	94.6	98.2	99.0
ReRank M -> M	93.5	95.7	97.5	98.3
M -> C	12.8	14.2	28.1	35.4
ReRank M -> C	19.4	17.6	28.1	35.9
M -> D	31.7	51.4	67.2	72.5
ReRank M -> D	47.6	57.6	67.9	73.4
C -> M	33.3	59.4	73.7	78.7
ReRank C -> M	47.3	64.0	72.0	76.1
C -> C	66.7	72.5	86.1	91.3
ReRank C -> C	80.8	80.1	86.9	92.2
C -> D	22.0	39.3	54.4	60.3
ReRank C -> D	36.1	47.7	57.5	61.8
D -> M	32.8	61.7	77.2	83.0
ReRank D -> M	48.0	65.6	74.1	78.8
D -> C	9.6	11.4	22.7	28.9
ReRank D -> C	15.4	14.4	22.1	28.7
D -> D	74.6	87.5	93.4	95.3
ReRank D -> D	85.5	89.7	93.6	95.2

I update the code, so that it can test with re-ranking now. Please refer to this commit. You can run it for yourself, by setting cfg.eval.rerank to True in package/config/default.py.

[ssbilakeri]

Thank you very much for your valuable response. Thanks a lot.

On Tue, Dec 29, 2020 at 9:10 AM Houjing Huang notifications@github.com wrote:

Hi, ssbilakeri. I have tested the re-ranking results for PAP_S_PS models (the script to run this is script/exp/test_PAP_S_PS_reranking.sh).

The original scores, as well as the re-ranking scores are as follows. mAP Rank-1 Rank-5 Rank-10 M -> M 85.6 94.6 98.2 99.0 ReRank M -> M 93.5 95.7 97.5 98.3 M -> C 12.8 14.2 28.1 35.4 ReRank M -> C 19.4 17.6 28.1 35.9 M -> D 31.7 51.4 67.2 72.5 ReRank M -> D 47.6 57.6 67.9 73.4 C -> M 33.3 59.4 73.7 78.7 ReRank C -> M 47.3 64.0 72.0 76.1 C -> C 66.7 72.5 86.1 91.3 ReRank C -> C 80.8 80.1 86.9 92.2 C -> D 22.0 39.3 54.4 60.3 ReRank C -> D 36.1 47.7 57.5 61.8 D -> M 32.8 61.7 77.2 83.0 ReRank D -> M 48.0 65.6 74.1 78.8 D -> C 9.6 11.4 22.7 28.9 ReRank D -> C 15.4 14.4 22.1 28.7 D -> D 74.6 87.5 93.4 95.3 ReRank D -> D 85.5 89.7 93.6 95.2

I update the code, so that it can test with re-ranking now. Please refer to this commit https://github.com/huanghoujing/EANet/commit/a38f12477e3edd625699f5a1beae92181e2c6b62. You can run it for yourself, by setting cfg.eval.rerank to True in package/config/default.py.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-751935474, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBMU2NOGZZEGCDXCHZTSXFFR3ANCNFSM4UQGVKRA .

[ssbilakeri]

Hello sir, when I train your model considering only segmentation loss it gives so less accuracy. have you run with only segmentation loss ? then what accuracy did you get?

for me result is like this Epoch 60 M->M: 6.8 ( 1.4), M->C: 0.1 ( 0.2), M->D: 5.7 ( 1.2)

kindly respond.

On Tue, Dec 29, 2020 at 11:08 AM Shavantrevva Bilakeri ssbilakeri@gmail.com wrote:

Thank you very much for your valuable response. Thanks a lot.

On Tue, Dec 29, 2020 at 9:10 AM Houjing Huang notifications@github.com wrote:

Hi, ssbilakeri. I have tested the re-ranking results for PAP_S_PS models (the script to run this is script/exp/test_PAP_S_PS_reranking.sh).

The original scores, as well as the re-ranking scores are as follows. mAP Rank-1 Rank-5 Rank-10 M -> M 85.6 94.6 98.2 99.0 ReRank M -> M 93.5 95.7 97.5 98.3 M -> C 12.8 14.2 28.1 35.4 ReRank M -> C 19.4 17.6 28.1 35.9 M -> D 31.7 51.4 67.2 72.5 ReRank M -> D 47.6 57.6 67.9 73.4 C -> M 33.3 59.4 73.7 78.7 ReRank C -> M 47.3 64.0 72.0 76.1 C -> C 66.7 72.5 86.1 91.3 ReRank C -> C 80.8 80.1 86.9 92.2 C -> D 22.0 39.3 54.4 60.3 ReRank C -> D 36.1 47.7 57.5 61.8 D -> M 32.8 61.7 77.2 83.0 ReRank D -> M 48.0 65.6 74.1 78.8 D -> C 9.6 11.4 22.7 28.9 ReRank D -> C 15.4 14.4 22.1 28.7 D -> D 74.6 87.5 93.4 95.3 ReRank D -> D 85.5 89.7 93.6 95.2

I update the code, so that it can test with re-ranking now. Please refer to this commit https://github.com/huanghoujing/EANet/commit/a38f12477e3edd625699f5a1beae92181e2c6b62. You can run it for yourself, by setting cfg.eval.rerank to True in package/config/default.py.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-751935474, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBMU2NOGZZEGCDXCHZTSXFFR3ANCNFSM4UQGVKRA .

[huanghoujing]

Hi, ssbilakeri. That's normal. Because segmentation does not learn anything about person re-identification. You have to train with at least one kind of re-identification loss, i.e. id loss or triplet loss.

[ssbilakeri]

I was not knowing that. Thank you for the information.

On Wed, 30 Dec 2020, 12:00 pm Houjing Huang, notifications@github.com wrote:

Hi, ssbilakeri. That's normal. Because segmentation does not learn anything about person re-identification. You have to train with at least one kind of re-identification loss, i.e. id loss or triplet loss.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-752345030, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBMEAE7AHM4VTN3HT2DSXLCHTANCNFSM4UQGVKRA .

[ssbilakeri]

Hi sir, how did you partition the feature map with keypoint delimitation. Which part of the code does that? Help me to understand Thank you.

On Wed, Dec 30, 2020 at 12:22 PM Shavantrevva Bilakeri ssbilakeri@gmail.com wrote:

I was not knowing that. Thank you for the information.

On Wed, 30 Dec 2020, 12:00 pm Houjing Huang, notifications@github.com wrote:

Hi, ssbilakeri. That's normal. Because segmentation does not learn anything about person re-identification. You have to train with at least one kind of re-identification loss, i.e. id loss or triplet loss.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/huanghoujing/EANet/issues/35#issuecomment-752345030, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOOYJBMEAE7AHM4VTN3HT2DSXLCHTANCNFSM4UQGVKRA .

[huanghoujing]

Hi, package/data/kpt_to_pap_mask.py is the code that partitions body into regions using keypoints.