Train squeezedet for own dataset for 1 class

[Shashikant255]

Hi,
I am wondering how to do transfer learning for squeezedet using pretrained model. I also want to reduce number of classes to 1.

[daisukelab]

Hi, I could successfully train my own.
You can simply edit squeeze.config created by python ../../main/config/create_config.py.
Check it first, it is easy to customize.
Regarding pretrained model, you can find following instruction in README.md. Change # of GPU to match to your environment.

python ../../scripts/train.py --gpus 2 --init ../../main/model/imagenet.h5

Hope this helps.

[Shashikant255]

@daisukelab Thanks for your help.
I have one more doubt regarding label (bounding box information about object i.e x, y, w, h) file. If any image doesn't have any class or object then its corresponding label file should be empty?

[daisukelab]

Hi, training code doesn't allow to have empty label, then you need to remove images which have zero class label from your train/valid lists. (I already experienced that training stopped with an error, then removed these entries to solve this issue.)

[Caliber2]

"CLASSES": 1,
"CLASS_NAMES": ["my_object"] is this correct way top modify squeeze.config for single object detection? or background needs to be added in CLASS_NAMES and CLASSES = 2?