search

dbolya / yolact

A simple, fully convolutional model for real-time instance segmentation.
MIT License
5.01k stars 1.32k forks source link

How to read the metrics? #631

Open SSITB opened 3 years ago

SSITB commented 3 years ago

Any idea for to read the result when i should stop training? example the below, what is the accuracy so far

[ 36] 3830 || B: 2.026 | C: 1.975 | M: 1.067 | S: 0.031 | T: 5.099 || ETA: 2 days, 20:34:07 || timer: 0.255 [ 36] 3840 || B: 2.027 | C: 1.985 | M: 1.047 | S: 0.030 | T: 5.090 || ETA: 2 days, 20:52:11 || timer: 0.348 [ 36] 3850 || B: 2.011 | C: 1.979 | M: 1.070 | S: 0.030 | T: 5.090 || ETA: 2 days, 21:01:24 || timer: 0.231 [ 36] 3860 || B: 2.016 | C: 1.972 | M: 1.070 | S: 0.030 | T: 5.089 || ETA: 2 days, 21:08:35 || timer: 0.264 [ 36] 3870 || B: 2.072 | C: 1.991 | M: 1.085 | S: 0.030 | T: 5.177 || ETA: 2 days, 19:33:48 || timer: 0.257 [ 36] 3880 || B: 2.076 | C: 2.013 | M: 1.108 | S: 0.031 | T: 5.227 || ETA: 2 days, 19:33:35 || timer: 0.275 [ 36] 3890 || B: 2.013 | C: 2.007 | M: 1.091 | S: 0.031 | T: 5.142 || ETA: 2 days, 19:36:01 || timer: 0.270 [ 36] 3900 || B: 2.020 | C: 2.022 | M: 1.089 | S: 0.032 | T: 5.162 || ETA: 2 days, 19:39:23 || timer: 0.211 [ 36] 3910 || B: 2.026 | C: 2.034 | M: 1.091 | S: 0.031 | T: 5.182 || ETA: 2 days, 19:43:30 || timer: 0.241 [ 36] 3920 || B: 2.054 | C: 2.043 | M: 1.094 | S: 0.031 | T: 5.222 || ETA: 2 days, 19:42:24 || timer: 0.193

Computing validation mAP (this may take a while)...

Calculating mAP...

   |  all  |  .50  |  .55  |  .60  |  .65  |  .70  |  .75  |  .80  |  .85  |  .90  |  .95  |

-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+ box | 3.06 | 7.88 | 6.37 | 5.47 | 4.12 | 3.08 | 2.02 | 1.08 | 0.39 | 0.14 | 0.02 | mask | 0.85 | 4.48 | 2.45 | 1.05 | 0.37 | 0.08 | 0.02 | 0.01 | 0.00 | 0.00 | 0.00 | -------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+

rusvagzur commented 3 years ago

Hello, depending on your task, my suggestion is Early Stopping (val loss vs train loss) or do a plot of the validation mAP and observe what is the iteration from which the network is not learning anymore (val mAP saturates).