X3D for detection MAP mismatch

First of all, thank you very much for this amazing repository.
I am trying to reproduce the results of X3D_M on AVA v2.2 following what the paper says:

> Since our paper focuses on efficiency, by default, we do not increase the spatial resolution of res5 by 2× [15]. Region-of-interest (RoI) features [21] are extracted at the last feature map of res5 by extending a 2D proposal at a frame into a 3D RoI by replicating it along the temporal axis, similar as done in previous work [24, 40, 66], followed by application of frame-wise RoIAlign [27] and temporal global average pooling. The RoI features are then max-pooled and fed to a per-class, sigmoid classifier for prediction.

> The network weights are initialized from the Kinetics models and we use step-wise learning rate decay, that is reduced by 10× when validation error saturates. We train for 14k iterations (68 epochs for ∼211k data), with linear warm-up [23] for the first 1k iterations and use a weight decay of 10−7, as in [15].

However, I am not able to achieve the reported MAP of 23.2.
I am using K400 pretrain for initialisation and my training is stuck at 21.5. I did not expect that huge difference in MAP coming from the pretraining change. So is there anything else not reported on the paper that you use for the detection training? Or do you know which is the expected map when using K400 pretrain?

Thank you again!