Apple Silicon M1 Cannot convert MPS Tensor to float64

[traderpedroso]

how could i convert to float32? on pure torch was simple, just converting tensor to dtype=torch.float32

full error

TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

[araffin]

Duplicate of #914 and also related more to PyTorch than SB3...

[traderpedroso]

Duplicate of #914 and also related more to PyTorch than SB3...

Its a different topic! MPS framework only support float32 add dtype=th.float32 on utils.py

def obs_as_tensor(
    obs: Union[np.ndarray, Dict[Union[str, int], np.ndarray]], device: th.device
) -> Union[th.Tensor, TensorDict]:
    """
    Moves the observation to the given device.

    :param obs:
    :param device: PyTorch device
    :return: PyTorch tensor of the observation on a desired device.
    """
    if isinstance(obs, np.ndarray):
        return th.as_tensor(obs, dtype=th.float32).to(device)
    elif isinstance(obs, dict):
        return {key: th.as_tensor(_obs).to(device) for (key, _obs) in obs.items()}
    else:
        raise Exception(f"Unrecognized type of observation {type(obs)}")

[araffin]

i assume you are using a custom env? (please fill the issue template completely)
a solution would be a gym wrapper that cast observations to float64.

[traderpedroso]

i assume you are using a custom env? (please fill the issue template completely)
a solution would be a gym wrapper that cast observations to float64.

yes using gym_anytrading

import gym
import gym_anytrading
from stable_baselines3.ppo.policies import MlpPolicy
from stable_baselines3 import A2C, PPO
from stable_baselines3.common.evaluation import evaluate_policy
from stable_baselines3.common.env_util import make_vec_env
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
import quantstats as qs


df = pd.read_csv("tickers.csv")

f = df.copy()[['open', 'high', 'low', 'close', 'volume']]
df = df.reset_index()
df.columns = ['Date', 'Open', 'High', 'Low', 'Close', 'Volume']
df = df.set_index('Date')


df.isnull().sum().sum() # there are no nans
df.fillna(method="ffill", inplace=True)
df = df.loc[~df.index.duplicated(keep = 'first')]   
df = df.dropna()
df = df.fillna(method="ffill")
df = df.dropna()
df = df.astype('float32')


window_size = 50
start_index = window_size
end_index = len(df)


def create_env(df, start_index, end_index, window_size):
    env = gym.make('forex-v0', df=df,  frame_bound=(start_index,
                   end_index), window_size=window_size)
    env.trade_fee = 0.01
    return env


def env_maker():
    return create_env(df, start_index, end_index, window_size)

env = make_vec_env(lambda: env_maker(), n_envs=1000) # 1000
model = PPO(MlpPolicy, env, device="mps", verbose=1) # PPO(MlpPolicy, env, verbose=1)

# Random evolute before training
mean_reward, std_reward = evaluate_policy(model, env, n_eval_episodes=10)
print(f"mean_reward:{mean_reward:.2f} +/- {std_reward:.2f}")

OUTPUT

TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

[araffin]

could you confirm that using a wrapper that cast observations to float32 (also for obs space) solve this issue ?

[traderpedroso]

could you confirm that using a wrapper that cast observations to float32 (also for obs space) solve this issue ?

Yes, I did it!
found out float64 won work with metal API as I said here #914, and thank you for your warning I want to use it only for portfolio manager calculate position size base volatility

[charliemagee]

I found a workaround on M1 Max. Switched to 3D animation mode. Used Zoe "Depth Algorithm." Unchecked "Use depth warping."

[TimEyerish]

I just had similar happen. I shortened prompts and switched models, it's working again.