def loadHierarchy(self):
self.cs = pyogmaneo.ComputeSystem()
self.h = pyogmaneo.Hierarchy("pupper.ohr")
def __init__(self,
env,
layerSizes=2 * [Int3(4, 4, 16)],
layerRadius=4,
hiddenSize=Int3(8, 8, 16),
imageRadius=9,
imageScale=1.0,
obsResolution=32,
actionResolution=16,
rewardScale=1.0,
terminalReward=0.0,
infSensitivity=1.0,
nThreads=4,
loadName=None):
self.env = env
pyogmaneo.ComputeSystem.setNumThreads(nThreads)
self.cs = pyogmaneo.ComputeSystem()
self.imEnc = None
self.imEncIndex = -1
self.inputSizes = []
self.inputLows = []
self.inputHighs = []
self.inputTypes = []
self.imageSizes = []
self.imgsPrev = []
self.actionIndices = []
self.rewardScale = rewardScale
self.terminalReward = terminalReward
self.infSensitivity = infSensitivity
if type(self.env.observation_space) is gym.spaces.Discrete:
self.inputSizes.append(Int3(1, 1, self.env.observation_space.n))
self.inputTypes.append(pyogmaneo.inputTypeNone)
self.inputLows.append([0.0])
self.inputHighs.append([0.0])
elif type(self.env.observation_space) is gym.spaces.Box:
if len(self.env.observation_space.shape) == 1 or len(
self.env.observation_space.shape) == 0:
squareSize = int(
np.ceil(np.sqrt(len(self.env.observation_space.low))))
squareTotal = squareSize * squareSize
self.inputSizes.append(
Int3(squareSize, squareSize, obsResolution))
self.inputTypes.append(pyogmaneo.inputTypeNone)
lows = list(self.env.observation_space.low)
highs = list(self.env.observation_space.high)
# Detect large numbers/inf
for i in range(len(lows)):
if abs(lows[i]) > 100000 or abs(highs[i]) > 100000:
# Indicate inf by making low greater than high
lows[i] = 1.0
highs[i] = -1.0
self.inputLows.append(lows)
self.inputHighs.append(highs)
elif len(self.env.observation_space.shape) == 2:
scaledSize = (int(self.env.observation_space.shape[0] *
imageScale),
int(self.env.observation_space.shape[1] *
imageScale), 1)
self.imageSizes.append(scaledSize)
elif len(self.env.observation_space.shape) == 3:
scaledSize = (int(self.env.observation_space.shape[0] *
imageScale),
int(self.env.observation_space.shape[1] *
imageScale), 3)
self.imageSizes.append(scaledSize)
else:
raise Exception("Unsupported Box input: Dimensions too high " +
str(self.env.observation_space.shape))
else:
raise Exception("Unsupported input type " +
str(type(self.env.observation_space)))
if len(self.imageSizes) > 0:
vlds = []
for i in range(len(self.imageSizes)):
vld = pyogmaneo.ImVisibleLayerDesc(
Int3(self.imageSizes[i][0], self.imageSizes[i][1],
self.imageSizes[i][2]), imageRadius)
vlds.append(vld)
self.imgsPrev.append(np.zeros(self.imageSizes[i]))
self.imEnc = pyogmaneo.ImageEncoder(self.cs, hiddenSize, vlds)
self.imEncIndex = len(self.inputSizes)
self.inputSizes.append(hiddenSize)
self.inputTypes.append(pyogmaneo.inputTypeNone)
self.inputLows.append([0.0])
self.inputHighs.append([1.0])
# Actions
if type(self.env.action_space) is gym.spaces.Discrete:
self.actionIndices.append(len(self.inputSizes))
self.inputSizes.append(Int3(1, 1, self.env.action_space.n))
self.inputTypes.append(pyogmaneo.inputTypeAction)
self.inputLows.append([0.0])
self.inputHighs.append([0.0])
elif type(self.env.action_space) is gym.spaces.Box:
if len(self.env.action_space.shape) < 3:
if len(self.env.action_space.shape) == 2:
self.actionIndices.append(len(self.inputSizes))
self.inputSizes.append(
Int3(self.env.action_space.shape[0],
self.env.action_space.shape[1], actionResolution))
self.inputTypes.append(pyogmaneo.inputTypeAction)
lows = list(self.env.action_space.low)
highs = list(self.env.action_space.high)
self.inputLows.append(lows)
self.inputHighs.append(highs)
else:
squareSize = int(
np.ceil(np.sqrt(len(self.env.action_space.low))))
squareTotal = squareSize * squareSize
self.actionIndices.append(len(self.inputSizes))
self.inputSizes.append(
Int3(squareSize, squareSize, actionResolution))
self.inputTypes.append(pyogmaneo.inputTypeAction)
lows = list(self.env.action_space.low)
highs = list(self.env.action_space.high)
self.inputLows.append(lows)
self.inputHighs.append(highs)
else:
raise Exception(
"Unsupported Box action: Dimensions too high " +
str(self.env.action_space.shape))
else:
raise Exception("Unsupported action type " +
str(type(self.env.action_space)))
lds = []
for i in range(len(layerSizes)):
ld = pyogmaneo.LayerDesc()
ld.hiddenSize = layerSizes[i]
ld.ffRadius = layerRadius
ld.pRadius = layerRadius
ld.aRadius = layerRadius
lds.append(ld)
if loadName is None:
self.h = pyogmaneo.Hierarchy(self.cs, self.inputSizes,
self.inputTypes, lds)
else:
self.h = pyogmaneo.Hierarchy(loadName)
self.actions = []
for i in range(len(self.actionIndices)):
index = self.actionIndices[i]
size = len(self.inputLows[index])
startAct = []
for j in range(size):
startAct.append(np.random.randint(0, self.inputSizes[index].z))
self.actions.append(startAct)
#
# This copy of EOgmaNeo is licensed to you under the terms described
# in the PYEOGMANEO_LICENSE.md file included in this distribution.
# ----------------------------------------------------------------------------
# -*- coding: utf-8 -*-
import numpy as np
import pyogmaneo
import matplotlib.pyplot as plt
# Set the number of threads
pyogmaneo.ComputeSystem.setNumThreads(4)
# Create the compute system
cs = pyogmaneo.ComputeSystem()
# This defines the resolution of the input encoding - we are using a simple single column that represents a bounded scalar through a one-hot encoding. This value is the number of "bins"
inputColumnSize = 64
# The bounds of the scalar we are encoding (low, high)
bounds = (-1.0, 1.0)
# Define layer descriptors: Parameters of each layer upon creation
lds = []
for i in range(5): # Layers with exponential memory
ld = pyogmaneo.LayerDesc()
# Set the hidden (encoder) layer size: width x height x columnSize
ld.hiddenSize = pyogmaneo.Int3(4, 4, 16)