深度学习 1660s AMP 加速问题

[faller]

使用 apex amp 后内存占用下降,但时间是原来的 5 倍, 前 5 耗时如下,切换 O1 O2 会导致其发生变化, 应该?换卡?!

ncalls	tottime	percall	cumtime	percall	filename:lineno(function)
82	342.244	4.174	342.244	4.174	{method 'item' of 'torch._C._TensorBase' objects}
91411	65.138	0.001	65.138	0.001	{method 'decode' of 'ImagingDecoder' objects}
1280	8.153	0.006	8.153	0.006	{imread}
80	1.464	0.018	1.464	0.018	{method 'run_backward' of 'torch._C._EngineBase' objects}
1280	0.654	0.001	0.654	0.001	{built-in method PIL._imaging.new}