Softmax Regressionの精度が低すぎる
Macbook air, Jupyter-Notebook, Python2.7
上記のリンクを参考にしソフトマックス回帰の実装を試みました。accurate: 0.104384133612と精度が約10パーセントしかありません。cost functionもすぐに極大点にたどり着きます、(局所解とは思えない、、)。どこがおかしいのでしょうか？
初歩的な内容かもしれませんが、よろしくお願いします。
def softmax(z):
e = np.exp(z - np.max(z))
#return e/np.sum(e)
return np.maximum(1e-5, e/np.sum(e))
n_class = 10
# Weight
W = np.random.randn(X_train.shape[1], n_class)
# Z = XW (+ b)
def transfunc(X, W):
return np.dot(X, W)
Z = transfunc(X_train, W)
def initialize_label(X):
Z = transfunc(X, W)
return np.argmax(np.apply_along_axis(softmax, 1, Z), axis = 1)
# Assign each data to a random label
labels = initialize_label(X_train); labels.shape
def loss_function(X, theta):
num = X.shape[0]
J = 0
z = np.dot(X, theta)
for i in xrange(num):
for k in xrange(n_class):
if labels[i] == k:
J = J - np.log(softmax(z[i])[k])
else:
J = J - 0
return J
def gradient_k(X, theta, k):
num = X.shape[0]
z = np.dot(X, theta)
grad = np.zeros_like(theta[:,k])
for i in xrange(num):
if labels[i] == k:
grad = grad - X[i]*(1 - softmax(z[i])[k])
else:
grad = grad + X[i]*softmax(z[i])[k]
return grad
loss_cost = [0]
def update(X, theta, eta = 0.0001, max_iter = 100):
num = X.shape[0]
for i in xrange(max_iter):
tmp_theta = theta
for k in xrange(n_class):
tmp_theta[:,k] = tmp_theta[:,k] + eta*gradient_k(X, theta, k)
theta = tmp_theta
loss_cost.append(loss_function(X, theta))
if abs(loss_function(X, theta)-loss_cost[i]) < 1e-5:
print "convege, {}:iter".format(i)
break
return theta