hiwonjoon/README.md

## README.md

      
    Raw
  

              README.md
            
          
    <<< Depreciated >>>

Conda Install

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod +x Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
<answer appropriately>
<recommend to setup conda on /scratch/cluster/<username>. Since miniconda can become really big!>
conda install -y 'python==3.6.*' # Downgrade python version


Install Cuda / Cudnn

<use conda if possible>
<if not, just download cuda and cudnn and extract it. then, add cuda directory to a path/ld_library_path>


Install basic required libraries using conda

conda install pkg-config
conda install patchelf
conda install -c menpo osmesa


Install basic required librarires using pip

pip insatll ipython
pip install pkgconfig
pip install tensorflow-gpu==1.13.1
pip install gym


Install mujoco-py

Check whether osmesa library installed correctly by pkg-config --cflags osmesa. The return should be -I<your conda include directory>
< Download mujoco200_linux.zip and extract on ~/.mujoco/ also, add mjkey.txt' >
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<path to your mujoco dir>
wget https://launchpad.net/ubuntu/+archive/primary/+files/libgcrypt11_1.5.3-2ubuntu4.2_amd64.deb
dpkg -x libgcrypt11_1.5.3-2ubuntu4.2_amd64.deb
mv lib/x86_64-linux-gnu/* <your miniconda dir>/lib
git clone git@github.com:hiwonjoon/mujoco-py.git # For a detailed change I made, please see the commit
cd mujoco-py
pip install .
cd ..
rm -rf libgcrypt11_1.5.3-2ubuntu4.2_amd64.deb mujoco-py

Check the installation of mujoco-py by running a test script
import mujoco_py
import os
mj_path, _ = mujoco_py.utils.discover_mujoco()
xml_path = os.path.join(mj_path, 'model', 'humanoid.xml')
model = mujoco_py.load_model_from_path(xml_path)
sim = mujoco_py.MjSim(model)

print(sim.data.qpos)
# [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]

sim.step()
print(sim.data.qpos)


Test Gym Headless Rendering

import gym
env = gym.make('Reacher-v2')
env.reset()
env.render('rgb_array').shape
# should be [500,500,3]


## dm_control_gpu_headless_rendering_test.py
from dm_control import mujoco
# Load a model from an MJCF XML string.
xml_string = """
<mujoco>
  <worldbody>
    <light name="top" pos="0 0 1.5"/>
    <geom name="floor" type="plane" size="1 1 .1"/>
    <body name="box" pos="0 0 .3">
      <joint name="up_down" type="slide" axis="0 0 1"/>
      <geom name="box" type="box" size=".2 .2 .2" rgba="1 0 0 1"/>
      <geom name="sphere" pos=".2 .2 .2" size=".1" rgba="0 1 0 1"/>
    </body>
  </worldbody>
</mujoco>
"""
physics = mujoco.Physics.from_xml_string(xml_string)
# Render the default camera view as a numpy array of pixels.

while True:
    pixels = physics.render()

## mujoco_py_render_test.py
import mujoco_py

import os
mj_path, _ = mujoco_py.utils.discover_mujoco()
xml_path = os.path.join(mj_path, 'model', 'humanoid.xml')
model = mujoco_py.load_model_from_path(xml_path)
sim = mujoco_py.MjSim(model)

viewer = mujoco_py.MjRenderContextOffscreen(sim, -1)

while True:
    viewer.render(300, 300, camera_id=-1)
    data = viewer.read_pixels(300, 300, depth=False)
    print(data.shape)

## mujoco_py_render_test_with_dm_control.py
from dm_control import mujoco
# Load a model from an MJCF XML string.
xml_string = """
<mujoco>
  <worldbody>
    <light name="top" pos="0 0 1.5"/>
    <geom name="floor" type="plane" size="1 1 .1"/>
    <body name="box" pos="0 0 .3">
      <joint name="up_down" type="slide" axis="0 0 1"/>
      <geom name="box" type="box" size=".2 .2 .2" rgba="1 0 0 1"/>
      <geom name="sphere" pos=".2 .2 .2" size=".1" rgba="0 1 0 1"/>
    </body>
  </worldbody>
</mujoco>
"""
physics = mujoco.Physics.from_xml_string(xml_string)
# Render the default camera view as a numpy array of pixels.

pixels = physics.render()

import mujoco_py

import os
mj_path, _ = mujoco_py.utils.discover_mujoco()
xml_path = os.path.join(mj_path, 'model', 'humanoid.xml')
model = mujoco_py.load_model_from_path(xml_path)
sim = mujoco_py.MjSim(model)

viewer = mujoco_py.MjRenderContextOffscreen(sim, -1)

while True:
    viewer.render(300, 300, camera_id=-1)
    data = viewer.read_pixels(300, 300, depth=False)
    print(data.shape)


## note.md

      
    Raw
  

              note.md
            
          
    Conda setup

conda create --no-default-packages -n tf2
conda activate tf2
Install Python

conda install -y 'python==3.7.*'
Install Cuda & Cudnn

conda install cudatoolkit=10.1
conda install cudnn
Install tensorflow

pip install tensorflow-gpu
Tensorflow GPU Test

export CUDA_VISIBLE_DEVICES=1
python tf-test.py
Install dm_control

Check GLEW is installed (required for EGL rendering)

dpkg -s libglew2.0
ldconfig -p | grep GLEW
pip install dm_control
The current version of dm_control has a PyOpenGL library compatibility issue. see the note below

vi /envs/tf2/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_ext.py
Test dm_control GPU rendering (watch nvidia-smi; the process should be seen on the process list,

and gpu usage should  go up)

python dm_control_gpu_headless_rendering_test.py
Install mujoco-py

pip3 install -U 'mujoco-py<2.1,>=2.0'
surprisingly enough, there is no hickup. OPENAI start to manage their code I guess.

Test mujoco_py rendering... (watch nvidia-smi)

python mujoco_py_render_test.py
Nope. they won't render using GPU.

But, when you load proper library through dm_control, then... (watch nvidia-smi)

python mujoco_py_render_test_with_gpu.py
it will be rendered with GPU!

############# Side Note #################
dm_control OPENGL version compatibility issue:

change code here
https://github.com/deepmind/dm_control/blob/30069ac11b60ee71acbd9159547d0bc334d63281/dm_control/_render/pyopengl/egl_ext.py#L42
ctypes.POINTER(EGLDeviceEXT), ctypes.POINTER(EGL.EGLint)) --> ctypes.POINTER(EGL.EGLDeviceEXT), ctypes.POINTER(EGL.EGLint))

  
## tf-test.py
from __future__ import absolute_import, division, print_function, unicode_literals

import tensorflow as tf

from tensorflow.keras.layers import Dense, Flatten, Conv2D
from tensorflow.keras import Model

mnist = tf.keras.datasets.mnist

(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

# 채널 차원을 추가합니다.
x_train = x_train[..., tf.newaxis]
x_test = x_test[..., tf.newaxis]

train_ds = tf.data.Dataset.from_tensor_slices(
    (x_train, y_train)).shuffle(10000).batch(32)
test_ds = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(32)

class MyModel(Model):
  def __init__(self):
    super(MyModel, self).__init__()
    self.conv1 = Conv2D(32, 3, activation='relu')
    self.flatten = Flatten()
    self.d1 = Dense(128, activation='relu')
    self.d2 = Dense(10, activation='softmax')

  def call(self, x):
    x = self.conv1(x)
    x = self.flatten(x)
    x = self.d1(x)
    return self.d2(x)

model = MyModel()

loss_object = tf.keras.losses.SparseCategoricalCrossentropy()

optimizer = tf.keras.optimizers.Adam()

train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')

test_loss = tf.keras.metrics.Mean(name='test_loss')
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')

@tf.function
def train_step(images, labels):
  with tf.GradientTape() as tape:
    predictions = model(images)
    loss = loss_object(labels, predictions)
  gradients = tape.gradient(loss, model.trainable_variables)
  optimizer.apply_gradients(zip(gradients, model.trainable_variables))

  train_loss(loss)
  train_accuracy(labels, predictions)

@tf.function
def test_step(images, labels):
  predictions = model(images)
  t_loss = loss_object(labels, predictions)

  test_loss(t_loss)
  test_accuracy(labels, predictions)

EPOCHS = 5

for epoch in range(EPOCHS):
  for images, labels in train_ds:
    train_step(images, labels)

  for test_images, test_labels in test_ds:
    test_step(test_images, test_labels)

  template = 'epoch: {}, loss: {}, acc: {},  test_loss: {},  testacc: {}'
  print (template.format(epoch+1,
                         train_loss.result(),
                         train_accuracy.result()*100,
                         test_loss.result(),
                         test_accuracy.result()*100))
	from dm_control import mujoco
	# Load a model from an MJCF XML string.
	xml_string = """
	<mujoco>
	<worldbody>
	<light name="top" pos="0 0 1.5"/>
	<geom name="floor" type="plane" size="1 1 .1"/>
	<body name="box" pos="0 0 .3">
	<joint name="up_down" type="slide" axis="0 0 1"/>
	<geom name="box" type="box" size=".2 .2 .2" rgba="1 0 0 1"/>
	<geom name="sphere" pos=".2 .2 .2" size=".1" rgba="0 1 0 1"/>
	</body>
	</worldbody>
	</mujoco>
	"""
	physics = mujoco.Physics.from_xml_string(xml_string)
	# Render the default camera view as a numpy array of pixels.

	while True:
	pixels = physics.render()
	import mujoco_py

	import os
	mj_path, _ = mujoco_py.utils.discover_mujoco()
	xml_path = os.path.join(mj_path, 'model', 'humanoid.xml')
	model = mujoco_py.load_model_from_path(xml_path)
	sim = mujoco_py.MjSim(model)

	viewer = mujoco_py.MjRenderContextOffscreen(sim, -1)

	while True:
	viewer.render(300, 300, camera_id=-1)
	data = viewer.read_pixels(300, 300, depth=False)
	print(data.shape)
	from __future__ import absolute_import, division, print_function, unicode_literals

	import tensorflow as tf

	from tensorflow.keras.layers import Dense, Flatten, Conv2D
	from tensorflow.keras import Model

	mnist = tf.keras.datasets.mnist

	(x_train, y_train), (x_test, y_test) = mnist.load_data()
	x_train, x_test = x_train / 255.0, x_test / 255.0

	# 채널 차원을 추가합니다.
	x_train = x_train[..., tf.newaxis]
	x_test = x_test[..., tf.newaxis]

	train_ds = tf.data.Dataset.from_tensor_slices(
	(x_train, y_train)).shuffle(10000).batch(32)
	test_ds = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(32)

	class MyModel(Model):
	def __init__(self):
	super(MyModel, self).__init__()
	self.conv1 = Conv2D(32, 3, activation='relu')
	self.flatten = Flatten()
	self.d1 = Dense(128, activation='relu')
	self.d2 = Dense(10, activation='softmax')

	def call(self, x):
	x = self.conv1(x)
	x = self.flatten(x)
	x = self.d1(x)
	return self.d2(x)

	model = MyModel()

	loss_object = tf.keras.losses.SparseCategoricalCrossentropy()

	optimizer = tf.keras.optimizers.Adam()

	train_loss = tf.keras.metrics.Mean(name='train_loss')
	train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')

	test_loss = tf.keras.metrics.Mean(name='test_loss')
	test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')

	@tf.function
	def train_step(images, labels):
	with tf.GradientTape() as tape:
	predictions = model(images)
	loss = loss_object(labels, predictions)
	gradients = tape.gradient(loss, model.trainable_variables)
	optimizer.apply_gradients(zip(gradients, model.trainable_variables))

	train_loss(loss)
	train_accuracy(labels, predictions)

	@tf.function
	def test_step(images, labels):
	predictions = model(images)
	t_loss = loss_object(labels, predictions)

	test_loss(t_loss)
	test_accuracy(labels, predictions)

	EPOCHS = 5

	for epoch in range(EPOCHS):
	for images, labels in train_ds:
	train_step(images, labels)

	for test_images, test_labels in test_ds:
	test_step(test_images, test_labels)

	template = 'epoch: {}, loss: {}, acc: {}, test_loss: {}, testacc: {}'
	print (template.format(epoch+1,
	train_loss.result(),
	train_accuracy.result()*100,
	test_loss.result(),
	test_accuracy.result()*100))