I'm not sure if I'm doing something wrong, but with fine_tune=False and training from the original pretrained inception network (e.g. retraining the whole network from scratch, not just tuning the detection heads), the network is missing some weights when you import because the detection heads aren't included in the variable scope (or some explanation like that), so I changed line 61 to variables_to_restore=original_inception_vars instead. I'm not sure what the actual intended usage of all this is.
I'm not sure if I'm doing something wrong, but with fine_tune=False and training from the original pretrained inception network (e.g. retraining the whole network from scratch, not just tuning the detection heads), the network is missing some weights when you import because the detection heads aren't included in the variable scope (or some explanation like that), so I changed line 61 to variables_to_restore=original_inception_vars instead. I'm not sure what the actual intended usage of all this is.