[EfficientDet] Train a custom dataset with pretrained weights

[SadiaAfrinPurba]

Hello,

I have trained a custom dataset from scratch using EfficientDet-D0. The performance was not enough satisfactory.
Is there any way I can train the custom dataset with pretrained weights (COCO weights) and fine-tune the last layer?

[abhi-kumar]

In the code here,
https://github.com/Tessellate-Imaging/Monk_Object_Detection/blob/master/4_efficientdet/lib/train.py

modify it to load your own weights, freeze base layers, and use it as transfer learning.

Else you may check up https://github.com/Tessellate-Imaging/Monk_Object_Detection/tree/master/example_notebooks/1_gluoncv_finetune pipeline, it's pure transfer learning.

[abhi-kumar]

Feature has been added.

[abhi-kumar]

The following features have been added.

a) Support for 8 new models

• model_name="efficientnet-b0"
• model_name="efficientnet-b1"
• model_name="efficientnet-b2"
• model_name="efficientnet-b3"
• model_name="efficientnet-b4"
• model_name="efficientnet-b5"
• model_name="efficientnet-b6"
• model_name="efficientnet-b7"
• model_name="efficientnet-b8"

#Usage
gtff.Model(model_name="efficientnet-b0");

b) Load a previously trained model and start a new training
gtf.Model(model_name="efficientnet-b0", load_pretrained_model_from="path to model.pth");