Use_amp:true causes Nan about zoedepth HOT 1 OPEN

I encountered NaN problem when I tried to support fp16 inference.
The cause is that some of MiDaS outputs are very large values, and that turns into inf in float16.
A workaround is to clip MiDaS output within float16 range.
nagadomi@384f325
Also, running DPTDepthModel.head with float16 degraded the accuracy, so it was better to use float32 for that one.
nagadomi/MiDaS_iw3@2e398c9
Note that I tried this only for inference.

--use_amp 0

{'a1': 0.954, 'a2': 0.995, 'a3': 0.999, 'abs_rel': 0.075, 'rmse': 0.27, 'log_10': 0.032, 'rmse_log': 0.095, 'silog': 7.034, 'sq_rel': 0.03}
654/654 [02:27<00:00,  4.42it/s]

--use_amp 1

{'a1': 0.954, 'a2': 0.995, 'a3': 0.999, 'abs_rel': 0.075, 'rmse': 0.271, 'log_10': 0.033, 'rmse_log': 0.095, 'silog': 7.039, 'sq_rel': 0.03}
654/654 [01:50<00:00,  5.90it/s]

from zoedepth.