Problem when the training is triggered from a pre trained model

There is a problem when the training is triggered using a pre-trained model

The test case bob.learn.tensorflow.test.test_cnn_pretrained_model:test_cnn_pretrained covers this workflow, but for some reason the training get stuck (in the session.run) when you have a network trained with several thousands iterations (I know it is weird, the iteration number is just a counter). This is annoying since our biggest queue allows us to run for 4 days only.