Python3.6でAttributeError: 'FigureCanvasMac' object has no attribute 'renderer'と出る

Question

前提・実現したいこと

前回の質問同様，オープンソースのニューラルネットワークのコード（Variational Integrator Networks）を用いて，この論文で行われているideal single pendulumの学習を再現したいと思っています．
環境構築をなんとか済ませ，実際にコードを回すために

python run_exp.py experiments VIN_VV pendulum pixels 1 60 1

を実行したところ，以下のようにFigureCanvasMacでAttributeErrorがでてしまいました．

AttributeError: 'FigureCanvasMac' object has no attribute 'renderer'
  In call to configurable 'create_dataset' (<function create_dataset at 0x13f1a6ae8>)
  In call to configurable 'main' (<function main at 0x13f1a6d90>)

発生している問題・エラーメッセージ

(VIN0503-Python3.6) ユーザー名@ユーザー名s-MacBook-Pro Variational_Integrator_Networks % python run_exp.py experiments VIN_VV pendulum pixels 1 60 1
 2021-05-04 01:24:09.657712: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2021-05-04 01:24:09.694326: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7ffec8f05b90 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2021-05-04 01:24:09.694348: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/numpy/core/_asarray.py:136: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray
  return array(a, dtype, copy=False, order=order, subok=True)
Traceback (most recent call last):
  File "run_exp.py", line 241, in <module>
    main(root_dir, model_name, system_name, observations, num_train_traj, num_train_steps, seed)
  File "/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/gin/config.py", line 1078, in gin_wrapper
    utils.augment_exception_message_and_reraise(e, err_str)
  File "/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/gin/utils.py", line 49, in augment_exception_message_and_reraise
    six.raise_from(proxy.with_traceback(exception.__traceback__), None)
  File "<string>", line 3, in raise_from
  File "/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/gin/config.py", line 1055, in gin_wrapper
    return fn(*new_args, **new_kwargs)
  File "run_exp.py", line 200, in main
    build_utils.create_dataset(system, model.horizon, num_train_traj, num_train_steps)
  File "/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/gin/config.py", line 1078, in gin_wrapper
    utils.augment_exception_message_and_reraise(e, err_str)
  File "/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/gin/utils.py", line 49, in augment_exception_message_and_reraise
    six.raise_from(proxy.with_traceback(exception.__traceback__), None)
  File "<string>", line 3, in raise_from
  File "/Users/ユーザー名/VIN0503-Python3.6/lib/python3.6/site-packages/gin/config.py", line 1055, in gin_wrapper
    return fn(*new_args, **new_kwargs)
  File "/Users/ユーザー名/VIN0503-Python3.6/Variational_Integrator_Networks/build_utils.py", line 119, in create_dataset
    obs, states = system.run(num_traj, num_steps, step_size)
  File "/Users/ユーザー名/VIN0503-Python3.6/Variational_Integrator_Networks/systems.py", line 73, in run
    return self._state_to_observations(qqd), qqd
  File "/Users/ユーザー名/VIN0503-Python3.6/Variational_Integrator_Networks/systems.py", line 33, in _state_to_observations
    return self._state_to_pixels(qqdot)
  File "/Users/ユーザー名/VIN0503-Python3.6/Variational_Integrator_Networks/systems.py", line 115, in _state_to_pixels
    y_t = np.array(f.canvas.renderer.buffer_rgba())[:, :, :1]
AttributeError: 'FigureCanvasMac' object has no attribute 'renderer'
  In call to configurable 'create_dataset' (<function create_dataset at 0x13f1a6ae8>)
  In call to configurable 'main' (<function main at 0x13f1a6d90>)

該当のソースコード

systems.py のファイル内容は以下の通りになります．文字数の関係から一部のみの抜粋となります．

import numpy as np
from scipy.integrate import ode

import plot_utils
import matplotlib.pyplot as plt


class System(object):

    def __init__(self, observations, seed, **kwargs):

        self.observations = observations
        self.RNG = np.random.RandomState(seed)
        self.system_param = kwargs
        self._init_system(**kwargs)

    def _init_system(self, **kwargs):
        raise NotImplementedError()

    def _random_init_state(self):
        raise NotImplementedError()

    def _ODE(self, t, y):
        raise NotImplementedError()

    def _state_to_observations(self, qqdot):
        if self.observations == "observed":
            return qqdot
        elif self.observations == "noisy":
            noise = self.RNG.randn(*qqdot.shape) * self.system_param["noise_std"]
            return qqdot + noise
        elif self.observations == "pixels":
            return self._state_to_pixels(qqdot)

    def _state_to_pixels(self, qqdot):
        raise NotImplementedError()

    def compute_energy(self, qqdot):
        raise NotImplementedError()

    def integrate_ODE(self, num_steps, step_size, y0=None, rtol=1e-12):

        T = num_steps * step_size
        t = np.linspace(0.0, T, num_steps)

        if y0 is None:
            y0 = self._random_init_state()

        solver = ode(self._ODE).set_integrator('dop853', rtol=rtol)
        sol = np.empty((len(t), 2))
        sol[0] = y0
        solver.set_initial_value(y0)
        k = 1
        while solver.successful() and solver.t < T:
            solver.integrate(t[k])
            sol[k] = solver.y
            k += 1
        
        return sol

    def run(self, num_traj, num_steps, step_size, y0=None):

        qqd = []
        for n in range(num_traj):
            if y0 is not None:
                y0_n = y0[n]
            else:
                y0_n = None
            qqd_n = self.integrate_ODE(num_steps, step_size, y0=y0_n)
            qqd.append(qqd_n[None])
        qqd = np.vstack(qqd)

        return self._state_to_observations(qqd), qqd

class Pendulum(System):

    def _init_system(self, mass=1.0, length=1.0, g=9.82, friction=0., **kwargs):

        def pendulum_ODE(t, y):
            q, qdot = y
            dydt = [qdot, -friction*qdot - (g/length)*np.sin(q)]
            return dydt

        self._ODE = pendulum_ODE

    def _random_init_state(self):

        sign = [-1., 1.]
        q0 = np.pi - self.RNG.uniform(low=0.1, high=1./2.*np.pi)
        q0 = self.RNG.choice(sign)*q0
        qdot0 = self.RNG.uniform(low=-1., high=1.)
        y0 = np.hstack([q0, qdot0])
        return y0

    def _state_to_pixels(self, qqdot):

        length = self.system_param["length"]
        q, qdot = np.split(qqdot, 2, axis=-1)
        x = length * np.sin(q)
        y = -length * np.cos(q)

        f, ax = plt.subplots(figsize=(1, 1), dpi=28)
        f_lim = length+0.2

        ys = []
        for traj in range(q.shape[0]):
            y_traj = []
            for t in range(q.shape[1]):
                plt.cla()
                ax.set_xlim(-f_lim, f_lim)
                ax.set_ylim(-f_lim, f_lim)
                ax.plot([0, x[traj, t]],[0, y[traj, t]], linewidth=12, color="black")
                ax.axis('off')
                f.canvas.draw()
                y_t = np.array(f.canvas.renderer.buffer_rgba())[:, :, :1]
                y_t[y_t > 0] = 255
                y_t[y_t == 0] = 1.
                y_t[y_t == 255] = 0.
                y_traj.append(np.float32(y_t[None]))
            y_traj = np.vstack(y_traj)
            ys.append(y_traj[None])
        ys = np.vstack(ys)
        ys = ys.reshape(*ys.shape[:2], 784)

        return ys

    def compute_energy(self, qqdot):

        mass = self.system_param["mass"]
        length = self.system_param["length"]
        g = self.system_param["g"]

        q, qdot = np.split(qqdot, 2, axis=-1)
        K = 0.5 * mass * (length**2) * qdot**2 
        U = -mass * g * length * np.cos(q)
        E = K + U

        return E

試したこと

元々Python3.7.5でやっていたのですが，

NotImplementedError: Cannot convert a symbolic Tensor (sequential_2_1/lstm/strided_slice:0) to a numpy array.
  In call to configurable 'create_model' (<function create_model at 0x134c40a70>)
  In call to configurable 'main' (<function main at 0x1366c3dd0>)

というエラーが出たため調べたところ，tensorflowのgithubにて同様のエラーに関するイシューを発見し，そこでPython3.6でやるといいと書いてあったため，Pythonのバージョンを3.7.5から3.6.0に下げました．

また，AttributeErrorの方のエラー文でも検索してみたところ，こちらも同様のエラーに関するイシューを発見しました．こちらで言及されている通り，

import matplotlib; matplotlib.use('agg')

を，systems.pyの1行目に追記した上で再度実行しましたが，同様のエラーが出てしまいました．

補足情報（FW/ツールのバージョンなど）

0S： MacOS Catalina ver10.15.7
仮想環境名：VIN0503-Python3.6

pip==21.0.1
Pytho==3.6.0
jax==0.1.60
numpy==1.19.5
matplotlib==3.2.0

Answer 1

おはようございます。問題提起確認しました。

Pythonファイルの上の方に以下を追加する。を試していただけますか？

import matplotlib
matplotlib.use('TKAgg')

・Matplotlib fig.canvas.renderer works on Win, fails on Mac

それでもダメなら、matplotlib==3.2.0のバージョンなので多分違うのですが、

np.array(f.canvas.renderer.buffer_rgba()) → np.array(f.canvas.renderer._renderer)

へ変更いただけますか？

・グラフをndarrayに変換してPillow（PIL）で表示する

Comments

[hamster-m]

前回に引き続きすばやいご回答ありがとうございます．

import matplotlib
matplotlib.use('TKAgg')

上記の2行をsystems.pyの1行目と2行目に追加したところ無事解決できました！ありがとうございました！

[退会済みユーザー]

お、よかった！！
引き続き応援してます〜。

[hamster-m]

ありがとうございます！！