can explain why do this (h + 1) // 2, (w + 1) // 2

[cometta]

Video-Pre-Training/lib/impala_cnn.py

Line 127 in 077ba2b

return (self.outchan, (h + 1) // 2, (w + 1) // 2)

Can explain why this line we do this +1 for height and weight before divide?

Video-Pre-Training/lib/impala_cnn.py

Line 168 in 077ba2b

init_scale=math.sqrt(len(chans)),

any paper i can read on why set init scaling using the number of channel ?

[Miffyli]

Hey. Unfortunately I (and other authors of VPT probably) do not have good answers to give off the bat, as these codes have probably lived in OpenAI's codebase for quite a while the original reasoning for them lost in time. Usually these might be things that "implemented found to work well" or saw it somewhere else.