평가함수, 평균제곱오차 & 교차엔트로피오차

pydl_costfuncs.py

import sys, os
import numpy as np

# 정답표, One hot label
t = np.array([
	[0,0,1,0,0,0,0,0,0,0], # 2
	[0,0,0,1,0,0,0,0,0,0], # 3
	[0,0,0,0,0,1,0,0,0,0], # 5
	[1,0,0,0,0,0,0,0,0,0], # 0
	[0,1,0,0,0,0,0,0,0,0]  # 1
])

# 정답일 확률
y = np.array([
	[.0, .0, .8, .0, .0, .0, .0, .0, .0, .0],
	[.0, .0, .0, .5, .0, .0, .0, .0, .0, .0],
	[.0, .0, .0, .0, .0, .3, .0, .0, .0, .0],
	[.7, .0, .0, .0, .0, .0, .0, .0, .0, .0],
	[.0, .2, .0, .0, .0, .0, .0, .0, .0, .0]
])

# 평균 제곱 오차mean squared error, MSE
def meanSquaredError(y, t):
	return 0.5 * np.sum( (y-t)**2 )

# 교차 엔트로피 오차cross entropy error, CEE
def crossEntropyError(y, t):
	return -np.sum( t * np.log(y+1e-7) )

def crossEntropyError_total(y, t):
	total_size = y.shape[0]
	if y.ndim == 1: # 항목이 하나인 1차원 배열인 경우, 2차원 배열로 변경.
		y.reshape(1, total_size)
		t.reshape(1, total_size)
	return -np.sum( t * np.log(y+1e-7) ) / total_size

def crossEntropyError_batch(y, t, batch_size):
	total_size = y.shape[0]
	if batch_size > total_size:
		batch_size = total_size

	if y.ndim == 1:
		y.reshape(1, total_size)
		t.reshape(1, total_size)

	batch_y = y[ np.arange(batch_size) ]
	batch_t = t[ np.arange(batch_size) ]

	return -np.sum( batch_t * np.log(batch_y+1e-7) ) / batch_size

print( crossEntropyError(y, t) )
print( crossEntropyError_total(y, t) )
print( crossEntropyError_batch(y, t, 3) )