How to fine-tune from a provided checkpoint ?

[ghost]

I would like to fine tune the provided pretrained model for Faster-RCNN trained on MSCOCO on some other dataset.

How can I do this fine-tuning ? To the best of my knowledge there is no such documentation on this aspect.

[joaqo]

Hi! You'll need to do the following to fine-tune on other datasets:

• If your dataset is not one of the supported datasets (COCO, PASCAL, OpenImages) you'll need to create a custom data reader for you data. Take a look at the PASCAL reader for an example of how to do it. Then you'll need to convert your data to a TF Recordd using this data reader.
• You'll need to create a custom config file which should point to the pretrained model's folder (keep in mind the default models are saved to your $HOME folder), so training knows to resume from this checkpoint, take a look here.

Let me know how it goes, Cheers!

[mauricio-meli]

When you say

point to the pretrained model's folder

you mean the job_dir in the train section right? Should it work with lumi cloud gc? When I try to retrain the model pointing to the old job_dir, a new job with a different config.yml is created in google cloud. It downloads the resnet_v1_101.ckpt checkpoint instead of using my previously trained checkpoint.

[giuliafonts]

@mauricio-meli did you find a solution?