I am trying to migrate my project from TF1 to TF2. However, after certain round of training, I noticed that the training conducted in TF2 was not capable of capturing certain object characteristic (which was doing well in TF1).
I am using Mask RCNN Inception Resnet V2 pre-trained model for both scenarios, the model for TF2 was obtained from models/research/object_detection/g3doc/tf2_detection_zoo.md at master · tensorflow/models · GitHub and models/research/object_detection/g3doc/tf1_detection_zoo.md at master · tensorflow/models · GitHub for TF1.
The image set used for training and evaluation were same for both scenario and parameter in the pipeline.config is minimal (only in the feature extractor section, exception occurred if I changed it to what it is in TF1).
How should I do to make the Mask R-CNN Inception Resnet V2 to perform as good as in TF1? Or is it possible to convert the model from TF1 model zoo so that it is compatible with TF2?
I run the inference on same dataset in both TF1 and TF2, and TF1 got significant lesser overkill in overall while TF2 was too easily affected by background noise.