在COCO数据集上使用sparse_inst_r50vd_dcn_giam_aug.yaml重头训练，精度并没有达到37.9，低了十几个点

[116022017144]

未使用fp16，batch_size=32,BASE_LR: 0.00005 STEPS: (210000, 250000) MAX_ITER: 270000 WEIGHT_DECAY: 0.05,其余参数配置未改，训练了212559iter，AP只有12.6,这是怎么回事呢？fp16对最终精度结果有很大的影响么？
[02/16 06:24:14] d2.evaluation.fast_eval_api INFO: Evaluate annotation type segm
[02/16 06:24:34] d2.evaluation.fast_eval_api INFO: COCOeval_opt.evaluate() finished in 19.81 seconds.
[02/16 06:24:34] d2.evaluation.fast_eval_api INFO: Accumulating evaluation results...
[02/16 06:24:36] d2.evaluation.fast_eval_api INFO: COCOeval_opt.accumulate() finished in 2.17 seconds.
[02/16 06:24:36] d2.evaluation.coco_evaluation INFO: Evaluation results for segm:

AP	AP50	AP75	APs	APm	APl
12.604	22.949	12.255	3.839	12.047	19.988

[wondervictor]

Hi @116022017144, FP16 is not crucial to the final performance. Could you provide the log file with me?

[lsm140]

未使用fp16，batch_size=32

未使用fp16，batch_size都能有32。泰裤辣！不使用fp16，我只能bs=4，开了就bs=8。即使选一块3090，不开fp16也就bs=8