TF2.0CPU and TF2.7GPU giving completely different results

I was running a Matterport MRCNN model on TF2.0 CPU on a custom data set. I ran 3 epochs with 100 steps each starting with COCO default weights. On the test data ,model was able to do object identification with bounding boxes and segmentation masks using the weights file from first epoch itself.

However , when I run the same model on TF2.7 GPU, model does not recognize anything. Not able to figure out if something needs to be done differently on TF2.7GPU