RpnL1loss is very high when I use 8gpus and change lr to 0.01

[anorthman]

dataset:coco2017

I added it to bbox_transform.py, but it can't solve the high rpnl1loss. dw = np.minimum(dw, np.log(1000. / 16.)) dh = np.minimum(dh, np.log(1000. / 16.))

If I change lr to 0.001, it will works. But lr is always set 0.02 for batchsize16 or 0.01 for batchsize8 in many papers. Why did I meet it when I changed lr to 0.01? It looks decreasing.