def test(self, inp, k=10):
vp0 = sigm(inp)
v0 = 1 * (vp0 > 0.5)
vs = [v0]
hs = []
for k in range(k):
## gibbs
# forward
hp0 = sigm(np.dot(v0, self.w) + self.hbias)
h0 = rng.binomial(1, hp0)
hs.append(h0)
# backward
vp1 = sigm(np.dot(h0, self.w.T) + self.vbias)
v1 = 1 * (vp1 > 0.5)
vs.append(v1)
v0 = v1.copy()
return vs, hs
def step(self, inp):
vp0 = sigm(inp)
v0 = 1 * (vp0 > 0.5)
## gibbs
# forward
hp0 = sigm(np.dot(v0, self.w) + self.hbias)
h0 = rng.binomial(1, hp0)
# backward
vp1 = sigm(np.dot(h0, self.w.T) + self.vbias)
v1 = rng.binomial(1, vp1)
# forward
hp1 = sigm(np.dot(v1, self.w) + self.hbias)
h1 = rng.binomial(1, hp1)
# learn
self.w += self.eta * (np.matmul(v0.T, hp0) - np.matmul(v1.T, hp1))
self.vbias += self.eta * np.mean(v0 - v1, 0)
self.hbias += self.eta * np.mean(hp0 - hp1, 0)
# error
return np.mean((v1 - v0)**2)
def calc_grad(self, w_G, xbias_NG, y_N):
''' Compute gradient of total loss for training logistic regression.
Args
----
w_G : 1D array, size G (G = n_features_including_bias)
Combined vector of weights and bias
xbias_NG : 2D array, size N x G (n_examples x n_features_including_bias)
Input features, with last column of all ones
y_N : 1D array, size N
Binary labels for each example (either 0 or 1)
Returns
-------
grad_wrt_w_G : 1D array, size G
Entry g contains derivative of loss with respect to w_G[g]
'''
#G = w_G.size
N = float(y_N.size)
denom = N * np.log(2)
# TODO calc gradient of L2 penalty term
grad_L2_wrt_w_G = self.alpha * w_G
# TODO calc gradient of log loss term
p = sigm(np.dot(xbias_NG, w_G))
p = np.asarray(p).reshape(-1)
grad_logloss_wrt_w_G = np.dot(y_N - p, xbias_NG)
grad_logloss_wrt_w_G = np.asarray(grad_logloss_wrt_w_G).reshape(-1)
#print(grad_L2_wrt_w_G , grad_logloss_wrt_w_G)
return (grad_L2_wrt_w_G - grad_logloss_wrt_w_G) / denom
def predict_proba(self, x_NF):
''' Produce soft probabilistic predictions for provided input features
Args
----
x_NF : 2D array, size N x F (n_examples x n_features_excluding_bias)
Input features (one row per example).
Returns
-------
yproba_N2 : 2D array, size N x 2
First column gives probability of zero label (negative)
Second column gives probability of one label (positive)
Each entry is a non-negative probability value within (0.0, 1.0)
Each row sums to one
'''
N = x_NF.shape[0]
## TODO write code to do prediction for logistic regression!
# Hint: Be sure to use a numerically stable logistic_sigmoid function
wx = self.w_G[:-1]
b = self.w_G[-1]
# TODO replace the placeholder code below
# Which just predicts 100% probability that class is 0
yproba0_N1 = np.ones((N, 1)) # <-- TODO replace this line
yproba1_N1 = np.zeros((N, 1)) # <-- TODO replace this line
yproba1_N1 = sigm(np.dot(x_NF, wx).T + b)
yproba0_N1 = 1 - yproba1_N1
yproba_N2 = np.column_stack([yproba0_N1, yproba1_N1])
return yproba_N2