num_states = 210 (Atari environment) indicates you need 210 frames to propagate, seems infeasible?

[QasimWani]

https://github.com/uvipen/Super-mario-bros-A3C-pytorch/blob/004d331c5f16022dfcf5d62d10574a7737db0bca/train.py#L49

In the first convolution layer in model.py, you specify the input channels as num_inputs. After digging through your repo, you're calling create_train_env() in src/env.py For Atari environments, breakout/pong/space-force, the observation space is (210, 160, 3). This means that you're picking up 210 frames before training the actual agent. Doesn't this result in huge training error initially, since the number of channels is significantly larger than most A3C papers indicate, maybe like 2 or 3? What's your justification towards such high of a channel number?