v 04 env.step returns only zeros for visual observations about animalai-olympics HOT 4 CLOSED

On Thu, Jun 6, 2019 at 11:51 AM Benjamin Beyret ***@***.***> wrote: Hi Phil, The speed at which you will be able to run the environment and train is very dependent on your hardware. Also, you need to differentiate two things that can take more or less time: - Taking a single step in the environment - Training of your algorithm If I understand your question correctly, your point "1." is regarding training (which library to use to train the agent), and point "2." is related to taking steps in the environment. Regarding training, you are completely free to implement whichever technique you want using any library you prefer (OpenCV in your case). However, this should not be linked to the performance of the environment and we cannot really help on that side as this part of the challenge is what is required from the participants. On the environment side, we have optimised the executable as much as possible to prevent the call to env.step() to be a bottleneck. To test for this, you can try to run a few thousand steps in the environment alone (without any training) and see how long it takes, put this code in your example folder and run it: from animalai.envs import UnityEnvironment from animalai.envs.arena_config import ArenaConfig import numpy as np import time env = UnityEnvironment( n_arenas=1, file_name='../env/AnimalAI', worker_id=1, seed=0, docker_training=False, play=False ) env.reset(arenas_configurations=ArenaConfig('configs/justFood.yaml')) N = 10000 before = time.time() for _ in range(N): env.step(np.random.randint(0, 3, size=2)) after = time.time() print('Steps per second: {0}'.format(N / (after - before))) env.close() on Linux with an i7 and GTX730 I get to 320 steps per second. Finally, if you wish to re-implement your own communicator and training interface you can do so, but keep in mind that at the time of submission we will expect an agent that implements agent.py (*the syntax of which will most likely change by the launch date*). If you still want to follow this route we will not provide help in any other language than python. I can only encourage you to read and debug the code in animalai/envs and familiarise yourself with protobuf which is the protocol used to communicate. Hope this helps, Benjamin — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#14?email_source=notifications&email_token=AAOSQUVZOVMJ3627C3D3SS3PZDM3RA5CNFSM4HQJJH42YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODXCKZWI#issuecomment-499428569>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAOSQUUXZLPGTT764UHBFRLPZDM3RANCNFSM4HQJJH4Q> .