kashif

from __future__ import print_function

from keras.datasets import cifar10
from keras.layers import merge, Input
from keras.layers.convolutional import Convolution2D, ZeroPadding2D, AveragePooling2D
from keras.layers.core import Dense, Activation, Flatten
from keras.layers.normalization import BatchNormalization
from keras.models import Model
from keras.preprocessing.image import ImageDataGenerator
from keras.utils import np_utils

kashif

from __future__ import print_function

from keras.datasets import cifar10
from keras.layers import merge, Input
from keras.layers.convolutional import Convolution2D, ZeroPadding2D, AveragePooling2D
from keras.layers.core import Dense, Activation, Flatten, Dropout
from keras.layers.normalization import BatchNormalization
from keras.models import Model
from keras.preprocessing.image import ImageDataGenerator
from keras.utils import np_utils

kashif

Cross Entropy Method

How do we solve for the policy optimization problem which is to maximize the total reward given some parametrized policy?

Discounted future reward

To begin with, for an episode the total reward is the sum of all the rewards. If our environment is stochastic, we can never be sure if we will get the same rewards the next time we perform the same actions. Thus the more we go into the future the more the total future reward may diverge. So for that reason it is common to use the discounted future reward where the parameter discount is called the discount factor and is between 0 and 1.

A good strategy for an agent would be to always choose an action that maximizes the (discounted) future reward. In other words we want to maximize the expected reward per episode.

kashif

name: "autocolorize"
input: "data"
input_dim: 1
input_dim: 1
input_dim: 514
input_dim: 514
layer {
  name: "data"
  type: "Input"
  top: "data"

kashif

import numpy as np
import gym
from gym.spaces import Discrete, Box
from gym.wrappers import Monitor
from keras.models import Sequential
from keras.layers import Dense, Activation, Flatten


# ================================================================
# Policies

kashif

'''Trains a simple convnet on the Zalando MNIST dataset.

Gets to 81.03% test accuracy after 30 epochs
(there is still a lot of margin for parameter tuning).
3 seconds per epoch on a GeForce GTX 980 GPU with CuDNN 5.
'''

from __future__ import print_function
import numpy as np
from mnist import MNIST

kashif

class AMSgrad(Optimizer):
    """AMSGrad optimizer.

    Default parameters follow those provided in the Adam paper.

    # Arguments
        lr: float >= 0. Learning rate.
        beta_1: float, 0 < beta < 1. Generally close to 1.
        beta_2: float, 0 < beta < 1. Generally close to 1.
        epsilon: float >= 0. Fuzz factor.

kashif

class AccSGD(Optimizer):
    """AccSGD optimizer.

    Arguments:
        lr (float): learning rate
        kappa (float, optional): ratio of long to short step (default: 1000)
        xi (float, optional): statistical advantage parameter (default: 10)
        smallConst (float, optional): any value <=1 (default: 0.7)

    # References

kashif

def input_fn(file_pattern, labels,
             image_size=(224,224),
             shuffle=False,
             batch_size=64, 
             num_epochs=None, 
             buffer_size=4096,
             prefetch_buffer_size=None):

    table = tf.contrib.lookup.index_table_from_tensor(mapping=tf.constant(labels))
    num_classes = len(labels)

kashif

import torch
import torch.nn as nn


class EvoNorm2d(nn.Module):
    __constants__ = ['num_features', 'eps', 'nonlinearity']
    
    def __init__(self, num_features, eps=1e-5, nonlinearity=True):
        super(EvoNorm2d, self).__init__()