Training from scratch using TAO for maskrcnn

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

• Hardware (A30)
• Network Type (Mask_rcnn)
• TLT Version (/tao-toolkit-tf:v3.22.05-tf1.15.5-py3)
• Training spec file(If have, please share here)
maskrcnn_train_resnet18.txt (2.1 KB)
I have two queries.
(1)If I like to train mrcnn from scratch. What do I need to change in config file?
(2)My images are fisheye images and like to train for mrcnn.
If I train using pretrained model, what are the config parameters to take care for good accuracy?

fisheye
fisheye1920×1080 153 KB

3
  1. Can set checkpoint to “”.
  2. Can refer to Poor metric results after retraining maskrcnn using TLT notebook - #17 by Morganh
4

Sure thanks. I’ll train scratch first

5

I am using the following config file. What parameters should I change?
The training accuracy is quite bad. I am training from scratch for fisheye images.

maskrcnn_train_resnet18.txt (2.2 KB)

I used same dataset and trained from scratch using mmdetection’s maskrcnn, the detection accuracy is quite good.

DLL 2023-01-04 11:17:25.504965 - Iteration: 4830 Validation Iteration: 4830  AP : 0.00856037437915802
DLL 2023-01-04 11:17:25.505199 - Iteration: 4830 Validation Iteration: 4830  AP50 : 0.020551249384880066
DLL 2023-01-04 11:17:25.505258 - Iteration: 4830 Validation Iteration: 4830  AP75 : 0.00629538856446743
DLL 2023-01-04 11:17:25.505304 - Iteration: 4830 Validation Iteration: 4830  APs : 0.0
DLL 2023-01-04 11:17:25.505350 - Iteration: 4830 Validation Iteration: 4830  APm : 0.0002463551063556224
DLL 2023-01-04 11:17:25.505398 - Iteration: 4830 Validation Iteration: 4830  APl : 0.009035307914018631
DLL 2023-01-04 11:17:25.505441 - Iteration: 4830 Validation Iteration: 4830  ARmax1 : 0.04031337797641754
DLL 2023-01-04 11:17:25.505482 - Iteration: 4830 Validation Iteration: 4830  ARmax10 : 0.09632998704910278
DLL 2023-01-04 11:17:25.505522 - Iteration: 4830 Validation Iteration: 4830  ARmax100 : 0.10416806489229202
DLL 2023-01-04 11:17:25.505565 - Iteration: 4830 Validation Iteration: 4830  ARs : 0.0
DLL 2023-01-04 11:17:25.505605 - Iteration: 4830 Validation Iteration: 4830  ARm : 0.0010416667209938169
DLL 2023-01-04 11:17:25.505664 - Iteration: 4830 Validation Iteration: 4830  ARl : 0.11147881299257278
DLL 2023-01-04 11:17:25.505706 - Iteration: 4830 Validation Iteration: 4830  mask_AP : 0.0009729214361868799
DLL 2023-01-04 11:17:25.505747 - Iteration: 4830 Validation Iteration: 4830  mask_AP50 : 0.003719929838553071
DLL 2023-01-04 11:17:25.505787 - Iteration: 4830 Validation Iteration: 4830  mask_AP75 : 9.244621651305351e-06
DLL 2023-01-04 11:17:25.505832 - Iteration: 4830 Validation Iteration: 4830  mask_APs : 0.0
DLL 2023-01-04 11:17:25.505879 - Iteration: 4830 Validation Iteration: 4830  mask_APm : 0.0001544429687783122
DLL 2023-01-04 11:17:25.505924 - Iteration: 4830 Validation Iteration: 4830  mask_APl : 0.0009968809317797422
DLL 2023-01-04 11:17:25.505964 - Iteration: 4830 Validation Iteration: 4830  mask_ARmax1 : 0.0026150394696742296
DLL 2023-01-04 11:17:25.506003 - Iteration: 4830 Validation Iteration: 4830  mask_ARmax10 : 0.006770129315555096
DLL 2023-01-04 11:17:25.506047 - Iteration: 4830 Validation Iteration: 4830  mask_ARmax100 : 0.007635500282049179
DLL 2023-01-04 11:17:25.506087 - Iteration: 4830 Validation Iteration: 4830  mask_ARs : 0.0
DLL 2023-01-04 11:17:25.506124 - Iteration: 4830 Validation Iteration: 4830  mask_ARm : 0.00244633830152452
DLL 2023-01-04 11:17:25.506164 - Iteration: 4830 Validation Iteration: 4830  mask_ARl : 0.007796799764037132
6

You can check the log mentioned in Poor metric results after retraining maskrcnn using TLT notebook - #20 by Morganh. It was training against the COCO dataset with pretrained model.

7

Yes that log is way much better than my log. Which parameter should i change? I am training from scratch.

8

The part of corresponding spec file is Poor metric results after retraining maskrcnn using TLT notebook - #17 by Morganh

10

I am training from scratch. so I shouldn’t freeze layers and batchnorm, right?
These two

#freeze_bn: True
 #freeze_blocks: "[0,1]"
11

Yes, that is right.

12

More, it is suggested to run comparable experiments.

  • Training from scratch
  • Train with pretrained model
13

Now my result getting better in training from scratch.
I’ll discuss what I did when completed.

14

The attached file is my config file and training from scratch for fisheye images.
Achieved good accuracy in training.
maskrcnn_train_resnet50.txt (2.4 KB)

1 Like
15

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.