How to do transfer learning in only last layer in Tensorflow?

[AI-LastWish]

I followed this tutorial https://github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10

I know that by default, TF API initialize from a checkpoint but allow retraining for all layers.

But I just want to retrain only the last layer, because I want to use pre-trained model in COCO to train in Open Images dataset.

Are there anyone can tell me how to do that?

System information

What is the top-level directory of the model you are using:(I don't really understand the meaning of "top-level directory of the model")
I just followed https://github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10

I used faster_rcnn_nas_lowproposals_coco downloaded from
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

OS Platform and Distribution:
Ubuntu 16.04
TensorFlow installed from:
GPU version of Tensorflow
TensorFlow version:
v1.12
Bazel version:
NA: don't know what is Bazel
CUDA/cuDNN version:
NA: don't know, but it's not the related to the problem
GPU model and memory:
Titan V
Exact command to reproduce:
NA

[rootkitchao]

You can set from_detection_checkpoint: false.
I think train a model in OID by init from a coco pre-trained model checkpoint may save some time,but also reduce the accuracy.

[AI-LastWish]

I'm using faster_rcnn_nas_coco.config, so you mean set from_detection_checkpoint: false in this file?
Could explain the meaning of that setting: What happen if I set from_detection_checkpoint: true and what will happen from_detection_checkpoint: false?

Moreover, I retrain only the last layer because I retrain faster_rcnn_nas_lowproposals_coco with input size: 1200x1200, which will led to 'OOM out of memory' error because the input resolution is so high, and it requires more RAM than RAM in my TITAN V(11 GB RAM)

So if I set from_detection_checkpoint: false, will the model requires less GPU memory than the original from_detection_checkpoint: true?

[rootkitchao]

Sorry, I copied this paragraph directly from my config, you should set it to true.
I don't think it will save much memory, and the memory overhead of the model itself still exists. If you want to use high resolution, you can only replace models with fewer parameters.

[AI-LastWish]

But if I train with only the last layer, will it reduce the accuracy because of the model is optimized for COCO, not for Open Images. So if I retrain only the last layer, will I get bad result with Open Images?

[rootkitchao]

But if I train with only the last layer, will it reduce the accuracy because of the model is optimized for COCO, not for Open Images. So if I retrain only the last layer, will I get bad result with Open Images?
Yes, of course. The MSCOCO data set is relatively small, so it is initialized from the ILSVRC checkpoint during training.But the OID v4 data set is relatively large, only need to train from scratch.So this will definitely reduce the accuracy slightly.

[ymodak]

Closing this issue since its addressed by rootkitchao's comment above. Feel free to reopen if have any further questions. Thanks!