Mac M1 conda env

[edwhu]

Hello,

I wanted to get this running on my M1 laptop for local development. The installation process deviated quite a bit from the README.

In particular, I had to setup tensorflow and jax differently than the README workflow. It went something like this:

conda create -n dreamerv3 python=3.8
conda activate dreamerv3

conda install -c apple tensorflow-deps
conda install jax -c conda-forge
pip install tensorflow-macos tensorflow_probability

pip install -r requirements.txt # comment out tensorflow-cpu in requirements.txt first.

I verified this environment by running example.py and by repeating the whole process again (conda installation and running example.py).

To make things easier for future users, I exported the conda env file. Then they don't need to manually install the conda dependencies and edit the requirements.txt.

Perhaps this will be useful for others?

[giteverything]

really nice work! When I run the example.py, some error occurs:

RuntimeError: Unknown backend: 'gpu' requested, but no platforms that are instances of gpu are present. Platforms
present are: interpreter,cpu

any idea how to fix it?

more details here:
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮
│ /Users/arthur/src/dreamerv3/example.py:53 in │
│ │
│ 50 │
│ 51 │
│ 52 if name == 'main': │
│ ❱ 53 main() │
│ 54 │
│ │
│ /Users/arthur/src/dreamerv3/example.py:42 in main │
│ │
│ 39 env = dreamerv3.wrap_env(env, config) │
│ 40 env = embodied.BatchEnv([env], parallel=False) │
│ 41 │
│ ❱ 42 agent = dreamerv3.Agent(env.obs_space, env.act_space, step, config) │
│ 43 replay = embodied.replay.Uniform( │
│ 44 │ config.batch_length, config.replay_size, logdir / 'replay') │
│ 45 args = embodied.Config( │
│ │
│ /Users/arthur/src/dreamerv3/dreamerv3/jaxagent.py:20 in init │
│ │
│ 17 │ configs = agent_cls.configs │
│ 18 │ inner = agent_cls │
│ 19 │ def init(self, *args, **kwargs): │
│ ❱ 20 │ super().init(agent_cls, *args, **kwargs) │
│ 21 return Agent │
│ 22 │
│ 23 │
│ │
│ /Users/arthur/src/dreamerv3/dreamerv3/jaxagent.py:35 in init │
│ │
│ 32 │ self.agent = agent_cls(obs_space, act_space, step, config, name='agent') │
│ 33 │ self.rng = np.random.default_rng(config.seed) │
│ 34 │ │
│ ❱ 35 │ available = jax.devices(self.config.platform) │
│ 36 │ self.policy_devices = [available[i] for i in self.config.policy_devices] │
│ 37 │ self.train_devices = [available[i] for i in self.config.train_devices] │
│ 38 │ self.single_device = (self.policy_devices == self.train_devices) and ( │
│ │
│ /Users/arthur/miniconda3/envs/dreamerv3/lib/python3.8/site-packages/jax/_src/lib/xla_bridge.py:4 │
│ 83 in devices │
│ │
│ 480 Returns: │
│ 481 │ List of Device subclasses. │
│ 482 """ │
│ ❱ 483 return get_backend(backend).devices() │
│ 484 │
│ 485 │
│ 486 def default_backend() -> str: │
│ │
│ /Users/arthur/miniconda3/envs/dreamerv3/lib/python3.8/site-packages/jax/_src/lib/xla_bridge.py:4 │
│ 25 in get_backend │
│ │
│ 422 │
│ 423 @lru_cache(maxsize=None) # don't use util.memoize because there is no X64 dependence. │
│ 424 def get_backend(platform=None): │
│ ❱ 425 return _get_backend_uncached(platform) │
│ 426 │
│ 427 │
│ 428 def get_device_backend(device=None): │
│ │
│ /Users/arthur/miniconda3/envs/dreamerv3/lib/python3.8/site-packages/jax/_src/lib/xla_bridge.py:4 │
│ 11 in _get_backend_uncached │
│ │
│ 408 │
│ 409 bs = backends() │
│ 410 if platform is not None: │
│ ❱ 411 │ platform = canonicalize_platform(platform) │
│ 412 │ backend = bs.get(platform, None) │
│ 413 │ if backend is None: │
│ 414 │ if platform in _backends_errors: │
│ │
│ /Users/arthur/miniconda3/envs/dreamerv3/lib/python3.8/site-packages/jax/_src/lib/xla_bridge.py:2 │
│ 94 in canonicalize_platform │
│ │
│ 291 for p in platforms: │
│ 292 │ if p in b.keys(): │
│ 293 │ return p │
│ ❱ 294 raise RuntimeError(f"Unknown backend: '{platform}' requested, but no " │
│ 295 │ │ │ │ │ f"platforms that are instances of {platform} are present. " │
│ 296 │ │ │ │ │ "Platforms present are: " + ",".join(b.keys())) │
│ 297 │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
RuntimeError: Unknown backend: 'gpu' requested, but no platforms that are instances of gpu are present. Platforms
present are: interpreter,cpu

[edwhu]

Hi, if you are on M1 mac, there is no GPU as the error message says. Just specify cpu instead of gpu in the config.