low mAP when doing transfer learning on a custom dataset?

[amaanda]

Hi! Thank you for sharing your work :)

I've been attempting to do transfer learning with a class that is not present on the coco dataset. I've read the issues in this repo and followed the instructions about creating a custom dataset and running transfer learning with yolact++. I changed the config.py file with the information about my dataset, used COCO Annotator to check my annotations and it all seemed fine.

However, I'm not sure if the results I'm getting so far are correct.

[ 50]   10550 || B: 0.378 | C: 0.735 | M: 0.717 | S: 0.065 | I: 0.043 | T: 1.939 || ETA: 6 days, 12:19:33 || timer: 0.647

Calculating mAP...

       |  all  |  .50  |  .55  |  .60  |  .65  |  .70  |  .75  |  .80  |  .85  |  .90  |  .95  |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+
   box | 24.89 | 30.36 | 30.36 | 30.28 | 30.15 | 29.78 | 29.35 | 28.27 | 24.70 | 14.16 |  1.44 |
  mask | 28.09 | 30.60 | 30.28 | 30.25 | 30.25 | 30.25 | 30.20 | 29.92 | 29.07 | 26.92 | 13.22 |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+


[204]   42430 || B: 0.380 | C: 0.751 | M: 0.656 | S: 0.065 | I: 0.032 | T: 1.883 || ETA: 7 days, 7:51:03 || timer: 0.763

Calculating mAP...

       |  all  |  .50  |  .55  |  .60  |  .65  |  .70  |  .75  |  .80  |  .85  |  .90  |  .95  |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+
   box | 24.50 | 30.48 | 30.45 | 30.43 | 30.01 | 30.01 | 29.61 | 27.86 | 22.88 | 12.74 |  0.58 |
  mask | 27.87 | 30.22 | 30.18 | 30.16 | 30.14 | 29.86 | 29.82 | 29.50 | 28.75 | 26.59 | 13.48 |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+


[420]   87140 || B: 0.377 | C: 0.683 | M: 0.573 | S: 0.052 | I: 0.030 | T: 1.715 || ETA: 8 days, 9:55:49 || timer: 0.801

Calculating mAP...

       |  all  |  .50  |  .55  |  .60  |  .65  |  .70  |  .75  |  .80  |  .85  |  .90  |  .95  |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+
   box | 24.03 | 30.63 | 30.63 | 30.63 | 30.62 | 30.59 | 29.85 | 27.90 | 20.63 |  8.16 |  0.64 |
  mask | 28.06 | 30.31 | 30.29 | 30.29 | 30.29 | 30.29 | 30.25 | 29.86 | 29.02 | 26.91 | 13.11 |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+

Thanks for you help!

[kidpaul94]

@amaanda May I ask how many dataset you had and how long you trained to get this mAP? Did you change learning rate or max_iteration? I'm doing a similar stuff and my mAP gets stuck around 14 for both box and mask.

[shivani21998]

I'm facing the same issue. Did you find out what was wrong?

[xuanzhiliu]

I have the same problem, I have 40 imgs for training and 20imgs for val,but I only have on class