チュートリアルから作った多層パーセプトロン

my_mlp_tutorial.py

# -*- coding: utf-8 -*-
"""
お手製のニューラルネットワーク
トレーニング編
"""
import sys
import numpy as np
from collections import defaultdict

def create_features(line, ids):
    phi = [0] * len(ids)
    words = line.rstrip().split()
    for word in words:
        phi[ids["UNI:"+word]] += 1
    return phi

def word2ids(fi):
    ids = defaultdict(lambda: len(ids))
    for line in fi:
        label,title = line.rstrip().split("\t")

        words = title.split()
        for word in words:
            ids["UNI:"+word]
    return ids

def forward_nn(net,phi0):
    phi_list = [phi0,"*","*"] #各層の値が入る
    for i, layer in enumerate(net):
        w = layer["w"]
        b = layer["b"]
        phi_list[i+1] = np.tanh(np.dot(w,phi_list[i])+b)
    # print phi_list
    return phi_list

def backward_nn(net,phi, true_label):
    J = len(net)
    delta = [0,0,np.array([true_label-phi[J][0]])]
    delta_prime = [0] * (J+1)

    for i in range(J)[::-1]:
        delta_prime[i+1] = delta[i+1] * (1 - phi[i+1]**2)
        w = net[i]["w"]
        b = net[i]["b"]
        delta[i] = np.dot(delta_prime[i+1],w)
    return delta_prime

def update_weights(net,phi,delta,_lambda):
    for i in range(len(net)):
        w = net[i]["w"]
        b = net[i]["b"]
        w += _lambda * np.outer(delta[i+1], phi[i])
        b += _lambda * delta[i+1]

if __name__ == "__main__":
    with open("./titles-en-train.labeled", "r") as fi:
        feat_lab = []
#データ全体をなめて，素性のインデックスを出す
#素性のインデックス（全体の長さ）が必須であるため．
        ids = word2ids(fi)
        fi.seek(0)
#ここからfeature,labelのセットのリストを作る
        for line in fi:
            label, title = line.rstrip().split("\t")
            feat_lab.append((create_features(title, ids), float(label)))

#ネットワークをランダムな値で初期化
#[0層,1層,...]
#-0.5から0.5の間で初期化した
    net = [{"w":np.random.rand(2,27190)-0.5, "b":np.random.rand(2)-0.5},\
     {"w":np.random.rand(1,2)-0.5, "b":np.random.rand(1)-0.5}]
    print net

#学習を行う
    for phi,label in feat_lab:
        phi_list = forward_nn(net, phi)
        delta_prime = backward_nn(net,phi_list,label)
        update_weights(net,phi,delta_prime, 0.1)

    print net
    print ids