def main():
D = 4
L = 2
# randomly generate parameter values for testing
np.random.seed(28031987)
W1 = np.random.randn(D,2*D)
b1 = np.random.randn(D)
W2 = np.random.randn(2*D,D)
b2 = np.random.randn(2*D)
Wlabel = np.random.randn(L,D)
# flattened arrays
W1flat = W1.flatten()
W2flat = W2.flatten()
Wlabelflat = Wlabel.flatten()
allflat = np.append(W1flat,b1)
allflat = np.append(allflat,W2flat)
allflat = np.append(allflat,b2)
allflat = np.append(allflat,Wlabelflat)
Np = len(allflat)
# hyperparameters
eps = 0.00000001
lambda_reg = 0.0
alpha = 0.2
# get the data
neg_list,pos_list = format_data()
neg_list = neg_list[:1]
pos_list = pos_list[:1]
vocab = np.random.randn(268810,D)
# do k_fold
mean,SEM = k_fold(neg_list,pos_list,10,L,alpha,lambda_reg,vocab,normalized=False)
exit(0)
# calculate finite difference approximation to gradient
numgrad = np.zeros(Np)
for i in range(Np):
print 'P ' + str(i)
allflat[i] += eps
fxpe = full_j(allflat,D,L,lambda_reg,alpha,neg_list,pos_list,
vocab,normalized=False)
allflat[i] -= 2.0*eps
fxme = full_j(allflat,D,L,lambda_reg,alpha,neg_list,pos_list,
vocab,normalized=False)
allflat[i] += eps
numgrad[i] = (fxpe - fxme)/(2.0*eps)
# now use backprop
bpgrad = compute_grad(allflat,D,L,lambda_reg,alpha,neg_list,pos_list,
vocab,normalized=False)
print numgrad
print bpgrad
print numgrad-bpgrad
def top_matches(prefs, person, n=5):
""" Rank movie critics matches for `person` from `perfs`.
:param prefs: dict - critics as defined in data.py
:param person: str - movie critic's name to compare others to
:param n: int - number of movie critics to rank
:return: list[tuple(float, str)] - ranking of (score, critic's name)
"""
svd = dt.train_svd(prefs)
result = svd.similar(person, n)
return dt.format_data(result)
def get_recommendations(prefs, person):
""" Recommend movies to `person` matching others' ratings with his.
:param prefs: dict - critics as defined in data.py
:param person: str - movie critic's name to advise
:return: list[tuple(float, str)] - ranking of (score, film's title)
Note that in this function, only movie which user has not rated
will be recommended
"""
n = 3
svd = dt.train_svd(prefs)
result = svd.recommend(person, n, only_unknowns=True, is_row=True)
return dt.format_data(result)