trackformer icon indicating copy to clipboard operation
trackformer copied to clipboard

Smaller Backbone Models

Open XAVILLA opened this issue 4 years ago • 4 comments

Hi Tim Meinhardt thanks for the great work!

I'm trying to adapt trackformer in my own research, however I don't have enough GPU memory. Do you have private mode models with smaller backbones(resnet18/34) pretrained on coco/crowdedhuman detection task or scripts for coco detection pretraining? Thanks!

XAVILLA avatar Jun 03 '21 18:06 XAVILLA

I am sorry but we never trained smaller models than the provided ResNet50.

timmeinhardt avatar Jun 04 '21 11:06 timmeinhardt

Okay. Do you have instructions/training scripts for coco pretraining then? I can also try to train it with smaller models myself. Thanks!

XAVILLA avatar Jun 04 '21 22:06 XAVILLA

After downloading and unpacking the COCO dataset into the data directory you should be able to train a model on COCO detection by running something like this:

python src/train.py with \
    deformable \
    backbone=resnet18 \
    output_dir=models/coco_resnet18

But this is not tested so make yourself familiar with the codebase to understand the configuration parameters in cfgs/train.yaml.

timmeinhardt avatar Jun 05 '21 13:06 timmeinhardt

Great! Thanks a lot!

XAVILLA avatar Jun 06 '21 20:06 XAVILLA