lxuechen

from __future__ import division
from numpy import *

class AdaBoost:

    def __init__(self, training_set):
        self.training_set = training_set
        self.N = len(self.training_set)
        self.weights = ones(self.N)/self.N
        self.RULES = []

lxuechen

""" Trains an agent with (stochastic) Policy Gradients on Pong. Uses OpenAI Gym. """
import numpy as np
import cPickle as pickle
import gym

# hyperparameters
H = 200 # number of hidden layer neurons
batch_size = 10 # every how many episodes to do a param update?
learning_rate = 1e-4
gamma = 0.99 # discount factor for reward

lxuechen

import theano
import theano.tensor as T
from theano.tensor.shared_randomstreams import RandomStreams
from theano.sandbox.rng_mrg import MRG_RandomStreams
from lasagne.updates import adam
from lasagne.utils import collect_shared_vars

from sklearn.datasets import fetch_mldata
from sklearn.cross_validation import train_test_split
from sklearn import preprocessing

lxuechen

  def backward_grads(self, y, dy, training=True):
    dy1, dy2 = dy
    y1, y2 = y

    with tf.GradientTape() as gtape:
      gtape.watch(y1)
      gy1 = self.g(y1, training=training)
    grads_combined = gtape.gradient(
        gy1, [y1] + self.g.trainable_variables, output_gradients=dy2)
    dg = grads_combined[1:]

lxuechen

checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer,
        learning_rate=learning_rate, global_step=global_step)

lxuechen

checkpoint.save(file_prefix)
checkpoint.restore(save_path)

lxuechen

tfe = tf.contrib.eager
model.call = tfe.defun(model.call)
model.compute_gradients = tfe.defun(model.compute_gradients)

lxuechen

def apply_gradients(optimizer, gradients, variables, global_step=None):
    optimizer.apply_gradients(
        zip(gradients, variables), global_step=global_step)
apply_gradients = tfe.defun(apply_gradients)

lxuechen

dataset = tf.data.TFRecordDataset(filename)
dataset = dataset.repeat(epochs).map(parser).batch(batch_size)

lxuechen

for image, label in dataset:
  logits = model(image, training=True)
  ...