Bad result when train with small batch size

[tranquan687]

I trained many times with batch size = 4 on RTX2080 Ti and got the best result below, I don't know the reason, does anyone meet the same issue?

Car [email protected], 0.70, 0.70:
bbox AP:84.6559, 76.2390, 69.5820
bev  AP:30.3075, 22.8254, 20.0140
3d   AP:22.3193, 17.3183, 15.5445
aos  AP:83.87, 74.73, 67.76
Car [email protected], 0.70, 0.70:
bbox AP:87.6583, 77.4353, 70.4441
bev  AP:26.2874, 18.2709, 14.9717
3d   AP:17.4402, 12.1115, 9.9476
aos  AP:86.76, 75.79, 68.42
Car [email protected], 0.50, 0.50:
bbox AP:84.6559, 76.2390, 69.5820
bev  AP:60.3748, 45.3538, 39.6286
3d   AP:55.1636, 40.5460, 36.5379
aos  AP:83.87, 74.73, 67.76
Car [email protected], 0.50, 0.50:
bbox AP:87.6583, 77.4353, 70.4441
bev  AP:60.5232, 43.0466, 36.9430
3d   AP:54.3911, 38.4424, 33.1688
aos  AP:86.76, 75.79, 68.42

[Ivan-Tang-3D]

Sorry for the delay, the decrease of results brought by the small batch size is natural.It is better for the A100 or 3090.