Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

NAN map values after training with pkummd dataset #81

Open
fmthoker opened this issue Nov 25, 2018 · 1 comment
Open

NAN map values after training with pkummd dataset #81

fmthoker opened this issue Nov 25, 2018 · 1 comment

Comments

@fmthoker
Copy link

fmthoker commented Nov 25, 2018

@yjxiong
I sucessfully trained a new model with PKUMMD dataset. However during testing I am getting nan values for map. I have attached both training and testing logs below.

These are some logs at test time.
/home/thoker/detection/ssn/anet_toolkit/Evaluation/eval_detection.py:231: RuntimeWarning: invalid value encountered in true_divide
rec = this_tp / npos
/home/thoker/detection/ssn/anet_toolkit/Evaluation/eval_detection.py:231: RuntimeWarning: invalid value encountered in true_divide
rec = this_tp / npos
/home/thoker/detection/ssn/anet_toolkit/Evaluation/eval_detection.py:231: RuntimeWarning: invalid value encountered in true_divide
rec = this_tp / npos
/home/thoker/detection/ssn/anet_toolkit/Evaluation/eval_detection.py:231: RuntimeWarning: invalid value encountered in true_divide
rec = this_tp / npos
Evaluation done.
+Detection Performance on pkummd--+------+------+------+------+------+------+------+---------+
| IoU thresh | 0.50 | 0.55 | 0.60 | 0.65 | 0.70 | 0.75 | 0.80 | 0.85 | 0.90 | 0.95 | Average |
+------------+------+------+------+------+------+------+------+------+------+------+---------+
| mean AP | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
+------------+------+------+------+------+------+------+------+------+------+------+---------+

Just to make sure my model is converging here are the training logs after 60 epochs:

Act. FG 100.00 (84.90) Act. BG 99.93 (99.93) [76/1402]
Epoch: [61][980/1177], lr: 0.00010 Time 0.967 (1.166) Data 0.535 (0.800) Loss 0.1610 (0.2812) Act. Loss 0.128 ( 0.233) Comp. Loss 0.061 ( 0.250) Reg
. Loss 0.271 (0.228)
Act. FG 87.50 (84.88) Act. BG 99.94 (99.94)
Epoch: [61][1000/1177], lr: 0.00010 Time 1.311 (1.165) Data 0.876 (0.800) Loss 0.4598 (0.2808) Act. Loss 0.407 ( 0.233) Comp. Loss 0.204 ( 0.249) Reg
. Loss 0.322 (0.229)
Act. FG 75.00 (84.92) Act. BG 99.94 (99.94)
Epoch: [61][1020/1177], lr: 0.00010 Time 1.445 (1.163) Data 1.016 (0.799) Loss 0.2110 (0.2813) Act. Loss 0.167 ( 0.234) Comp. Loss 0.246 ( 0.249) Reg
. Loss 0.190 (0.228)
Act. FG 75.00 (84.89) Act. BG 99.94 (99.94)
Epoch: [61][1040/1177], lr: 0.00010 Time 1.517 (1.169) Data 1.090 (0.806) Loss 0.1750 (0.2805) Act. Loss 0.129 ( 0.233) Comp. Loss 0.218 ( 0.249) Reg
. Loss 0.241 (0.228)
Act. FG 87.50 (84.98) Act. BG 99.94 (99.94)
Epoch: [61][1060/1177], lr: 0.00010 Time 0.701 (1.171) Data 0.297 (0.807) Loss 0.8812 (0.2801) Act. Loss 0.837 ( 0.232) Comp. Loss 0.346 ( 0.248) Reg
. Loss 0.098 (0.228)
Act. FG 75.00 (85.04) Act. BG 99.94 (99.94)
Epoch: [61][1080/1177], lr: 0.00010 Time 0.620 (1.171) Data 0.001 (0.809) Loss 0.3701 (0.2809) Act. Loss 0.291 ( 0.233) Comp. Loss 0.116 ( 0.249) Reg
. Loss 0.674 (0.228)
Act. FG 75.00 (85.01) Act. BG 99.94 (99.94)
Epoch: [61][1100/1177], lr: 0.00010 Time 0.625 (1.170) Data 0.000 (0.807) Loss 0.2096 (0.2793) Act. Loss 0.152 ( 0.232) Comp. Loss 0.277 ( 0.248) Reg
. Loss 0.302 (0.229)
Act. FG 87.50 (85.13) Act. BG 99.94 (99.94)
Epoch: [61][1120/1177], lr: 0.00010 Time 0.618 (1.168) Data 0.058 (0.805) Loss 0.6727 (0.2788) Act. Loss 0.559 ( 0.231) Comp. Loss 0.634 ( 0.249) Reg
. Loss 0.506 (0.228)
Act. FG 62.50 (85.18) Act. BG 99.94 (99.94)
Epoch: [61][1140/1177], lr: 0.00010 Time 1.028 (1.166) Data 0.595 (0.803) Loss 0.2802 (0.2800) Act. Loss 0.196 ( 0.232) Comp. Loss 0.236 ( 0.249) Reg
. Loss 0.605 (0.228)
Act. FG 87.50 (85.10) Act. BG 99.95 (99.95)
Epoch: [61][1160/1177], lr: 0.00010 Time 1.290 (1.166) Data 0.865 (0.803) Loss 0.2358 (0.2799) Act. Loss 0.209 ( 0.232) Comp. Loss 0.136 ( 0.249) Reg
. Loss 0.133 (0.228)
Act. FG 62.50 (85.13) Act. BG 99.95 (99.95)
Test: [0/17] Time 4.963 (4.963) Loss 1.9900 (1.9900) Act. Loss 1.975 (1.975) Comp. Loss 0.070 (0.070) Act. Accuracy 75.00 (75.00) FG 87.50 BG 62.50 Reg. Loss 0.079 (0.
079)
Testing Results: Loss 4.05014 Activity Loss 4.006 Completeness Loss 0.207
Act Accuracy 56.06 FG Acc. 79.55 BG Acc. 32.58 Regression Loss 0.238

Epoch: [62][0/1177], lr: 0.00010 Time 7.005 (7.005) Data 6.562 (6.562) Loss 1.0257 (1.0257) Act. Loss 0.961 ( 0.961) Comp. Loss 0.544 ( 0.544) Reg
. Loss 0.099 (0.099)

@fmthoker fmthoker changed the title NAN map values after training with pkummd dataaset NAN map values after training with pkummd dataset Nov 25, 2018
@yjxiong
Copy link
Owner

yjxiong commented Nov 27, 2018

You can modify the code to see the per-class APs. Usually, this is due to an incorrect number of classes, where the missing class has no instance detected and thus a NaN ap.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants