_feature_extractor from faster-rcnn_resnet is not a trackable object. Impossible to get Weights restored.

[jjuanvision]

I'm trying to restore the weights via checkpoint. The issue that I come across is that the _feature_extractor is not a trackable object.

tf.train.Checkpoint(_feature_extractor=detection_model._feature_extractor). This just happens with the faster rcnn meta architecture as opposite with the ssd one.

After that, I tried to restore the weights just for the prediction heads (box head and class ones). I'm able to do that but then I have another issue with initializing the variables by passing dummy images.

The environment is the same as the google colab Object Detection API Colab TF2.

Thanks in advance!

[jjuanvision]

https://gist.github.com/jjuanvision/31c5445f9a2f9eac8abc9a1a9b0b14ee

Here, I leave enclosed the gist where you can reproduce the error following every step from the notebook.

Now, I just restore the weights for the first stage box predictor and left aside initialized the feature extractor so that I can go on. I manage to fine tuned but then when I try to predict it throws an error related to " ValueError: Dimensions 0 and 1 are not compatible".