def central_diff(self, f, epsilon, theta):
print epsilon
x = theta.get_matrix()
n = x.shape[0]
g = np.zeros(x.shape)
if x.ndim == 2:
for i in range(x.shape[0]):
for j in range(x.shape[1]):
upper = x.copy()
upper[i, j] += epsilon
lower = x.copy()
lower[i, j] -= epsilon
g[i,
j] = ((f(DataFrame.from_matrix(upper)) -
f(DataFrame.from_matrix(lower))) / (2 * epsilon))
elif x.ndim == 1:
for i in range(x.shape[0]):
upper = x.copy()
upper[i] += epsilon
lower = x.copy()
lower[i] -= epsilon
g[i] = ((f(DataFrame.from_matrix(upper)) -
f(DataFrame.from_matrix(lower))) / (2 * epsilon))
else:
raise ValueError
return g
def test_softmax_reg_loss(self):
df = DataFrame()
epsilon = 1e-4
y_path = ("y/", "y/")
theta_path = ("theta/", "theta/")
X_path = ("X/", "X/")
k = 10
n, m = 5, 8
df[X_path] = DataFrame.from_matrix(nprand.rand(n, m))
df[theta_path] = DataFrame.from_matrix(nprand.rand(k, m))
y = np.zeros((n, k), dtype=bool)
for i in range(n):
j = nprand.randint(k)
y[i, j] = True
df[y_path] = DataFrame.from_matrix(y)
reg = 0.0001
softmax = lambda theta_df: SoftmaxRegression(theta_df, df[X_path], df[
y_path], reg).f()
g_central = self.central_diff(softmax, epsilon, df[theta_path])
g1 = SoftmaxRegression(df[theta_path], df[X_path], df[y_path], reg).g()
# print g_central
assert (np.allclose(g_central, g1))
def central_diff(self,f,epsilon,theta):
print epsilon
x = theta.get_matrix()
n = x.shape[0]
g = np.zeros(x.shape)
if x.ndim == 2:
for i in range(x.shape[0]):
for j in range(x.shape[1]):
upper = x.copy()
upper[i,j] += epsilon
lower = x.copy()
lower[i,j] -= epsilon
g[i,j] = ((f(DataFrame.from_matrix(upper))
-f(DataFrame.from_matrix(lower)))
/(2*epsilon))
elif x.ndim == 1:
for i in range(x.shape[0]):
upper = x.copy()
upper[i] += epsilon
lower = x.copy()
lower[i] -= epsilon
g[i] = ((f(DataFrame.from_matrix(upper))
-f(DataFrame.from_matrix(lower)))
/(2*epsilon))
else:
raise ValueError
return g
def test_softmax_reg_loss(self):
df = DataFrame()
epsilon = 1e-4
y_path = ("y/","y/")
theta_path = ("theta/","theta/")
X_path = ("X/","X/")
k = 10
n,m = 5,8
df[X_path] = DataFrame.from_matrix(nprand.rand(n,m))
df[theta_path] = DataFrame.from_matrix(nprand.rand(k,m))
y = np.zeros((n,k),dtype=bool)
for i in range(n):
j = nprand.randint(k)
y[i,j] = True
df[y_path] = DataFrame.from_matrix(y)
reg = 0.0001
softmax = lambda theta_df: SoftmaxRegression(theta_df, df[X_path],
df[y_path], reg).f()
g_central = self.central_diff(softmax,epsilon,df[theta_path])
g1 = SoftmaxRegression(df[theta_path], df[X_path], df[y_path], reg).g()
# print g_central
assert(np.allclose(g_central,g1))
# Test batch by checking average gradient
# g2 = np.zeros((k,m))
# for i in range(n):
# g2 += Softmax.g(df[theta_path], df[X_path], df[y_path], reg)
# g2 /= n
# assert(np.allclose(g_central,g2))
def test_permutation(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
permute_path1 = ("row2/", "col1/")
M1 = nprand.rand(3, 5)
df[M1_path] = DataFrame.from_matrix(M1)
df[permute_path1] = Permute(df[M1_path])
p_df = df["auto/row1/", "auto/permutation/"]
p = p_df.get_matrix().ravel()
assert (df[permute_path1].get_matrix() == M1[p, :]).all()
def test_dot(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
M2_path = ("row2/", "col2/")
dot_path1 = ("row1/", "col2/")
M1 = nprand.rand(3, 5)
M2 = nprand.rand(5, 8)
df[M1_path] = DataFrame.from_matrix(M1)
df[M2_path].set_matrix(M2)
df[dot_path1] = Dot(df[M1_path], df[M2_path])
assert (df[dot_path1].get_matrix() == M1.dot(M2)).all()
def func(self, target_df, X_df):
X = X_df.get_matrix()
P = np.random.permutation(X.shape[0])
row_labels = X_df._row_index.keys()
col_labels = X_df._col_index.keys()
(row_query, col_query) = X_df.pwd()
X_df._top_df[_auto_dir + row_query, _auto_dir + "permutation/"] = DataFrame.from_matrix(P[:, None])
# print "finished permute"
return DataFrame.from_matrix(X[P, :], row_labels, col_labels)
def func(self, target_df, X_df):
X = X_df.get_matrix()
P = np.random.permutation(X.shape[0])
row_labels = X_df._row_index.keys()
col_labels = X_df._col_index.keys()
(row_query, col_query) = X_df.pwd()
X_df._top_df[_auto_dir+row_query,_auto_dir+"permutation/"] = \
DataFrame.from_matrix(P[:,None])
# print "finished permute"
return DataFrame.from_matrix(X[P, :], row_labels, col_labels)
def test_linear(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
M2_path = ("row2/", "col2/")
linear_path1 = ("row1/", "col2/")
M1 = nprand.rand(3, 5)
M2 = nprand.rand(3, 5)
df[M1_path] = DataFrame.from_matrix(M1)
df[M2_path].set_matrix(M2)
a = 2
b = -3
df[linear_path1] = Linear(a, df[M1_path], b, df[M2_path])
assert (df[linear_path1].get_matrix() == a * M1 + b * M2).all()
def test_gd(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
M2_path = ("row2/", "col2/")
batch1_path = ("row1/", "col1/batch1/")
batch2_path = ("row1/", "col1/batch2/")
x0_path = ("x0/", "y0/")
M1 = nprand.rand(3, 5)
df[batch1_path].set_matrix(M1)
M2 = np.zeros((3, 5))
df[M2_path] = GD(SquareTest, M2, df[M1_path], step_size=1)
sleep(1)
# df[M2_path].stop()
assert np.allclose(df[M2_path].get_matrix(), df[M1_path].get_matrix())
# Assert that the input structure has been replicated
assert (
df["row2/",
"col2/batch1/"].get_matrix() == df[M2_path].get_matrix()).all()
# Now attempt to extend the parameter matrix
M3 = nprand.rand(3, 4)
df[batch2_path].set_matrix(M3)
sleep(1)
assert df[M2_path].shape == df[M1_path].shape
assert df[M2_path].shape == (3, 9)
assert np.allclose(df[M2_path].get_matrix(), df[M1_path].get_matrix())
df[M2_path].stop()
def test_zero_mean(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
M2_path = ("row2/", "col2/")
M1 = nprand.rand(3, 5)
M1_zm = M1 - np.mean(M1, axis=0)
df[M1_path].set_matrix(M1)
df[M2_path] = ZeroMean(df[M1_path])
assert (df[M2_path].get_matrix() == M1_zm).all()
def func(self,target_df,X_df, num_bases=50):
X = X_df.get_matrix()
X_m = np.mean(X,axis=0) # mean
X_zm = X - X_m # X with 0 mean
u,s,v_T = la.svd(X_zm)
row_labels = [str(i) for i in range(X.shape[1])]
col_labels = [str(i) for i in range(num_bases)]
return DataFrame.from_matrix(np.real(v_T.T[:,:num_bases]),row_labels,col_labels)
def func(self, target_df, X_df, num_bases=50):
X = X_df.get_matrix()
X_m = np.mean(X, axis=0) # mean
X_zm = X - X_m # X with 0 mean
u, s, v_T = la.svd(X_zm)
row_labels = [str(i) for i in range(X.shape[1])]
col_labels = [str(i) for i in range(num_bases)]
return DataFrame.from_matrix(np.real(v_T.T[:, :num_bases]), row_labels,
col_labels)
def test_permutation(self):
df = DataFrame()
M1_path = ("row1/","col1/")
permute_path1 = ("row2/","col1/")
M1 = nprand.rand(3,5)
df[M1_path] = DataFrame.from_matrix(M1)
df[permute_path1] = Permute(df[M1_path])
p_df = df["auto/row1/","auto/permutation/"]
p = p_df.get_matrix().ravel()
assert (df[permute_path1].get_matrix()==M1[p,:]).all()
def test_dot(self):
df = DataFrame()
M1_path = ("row1/","col1/")
M2_path = ("row2/","col2/")
dot_path1 = ("row1/","col2/")
M1 = nprand.rand(3,5)
M2 = nprand.rand(5,8)
df[M1_path] = DataFrame.from_matrix(M1)
df[M2_path].set_matrix(M2)
df[dot_path1] = Dot(df[M1_path],df[M2_path])
assert (df[dot_path1].get_matrix()==M1.dot(M2)).all()
def test_sgd(self):
# Also test sgd
close = np.array([[-44.25076083, 38.62854577],
[-38.41473092, 36.29945225],
[-31.43300105, 30.79620632],
[-21.27706071, 24.08638079],
[-14.00259076, 6.54438641],
[ 11.52354442, -6.07783327],
[ 48.69374796, -38.64696136],
[ 95.49682071, -84.38906967]])
df = DataFrame()
path = "row/","col/"
df["xrow/","xcol/"]= DataFrame.from_matrix(np.arange(16).reshape(8,2))
df["yrow/","ycol/"] = DataFrame.from_matrix(np.arange(8).reshape(8,1))
X_df = df["xrow/","xcol/"]
y_df = df["yrow/","ycol/"]
df[path] = SGD(SquareTest,close,y_df,batch_size=8,step_size=0.5)
sleep(1)
df[path].stop()
assert np.allclose(df[path].get_matrix(), y_df.get_matrix())
def test_tuple_to_query(self):
df = DataFrame()
# Test conversion of hashable elements to their actual queries
string = "randomstring"
slice_hash, slice_actual = (slice, (2, 4, 1)), slice(2, 4, 1)
list_hash, list_actual = (list, (1, 2, 3, 4, 5, 6)), [1, 2, 3, 4, 5, 6]
assert df._tuple_element_to_query(string) == string
assert df._tuple_element_to_query(slice_hash) == slice_actual
assert df._tuple_element_to_query(list_hash) == list_actual
assert df._query_to_tuple_element(string) == string
assert df._query_to_tuple_element(slice_actual) == slice_hash
assert df._query_to_tuple_element(list_actual) == list_hash
def test_linear(self):
df = DataFrame()
M1_path = ("row1/","col1/")
M2_path = ("row2/","col2/")
linear_path1 = ("row1/","col2/")
M1 = nprand.rand(3,5)
M2 = nprand.rand(3,5)
df[M1_path] = DataFrame.from_matrix(M1)
df[M2_path].set_matrix(M2)
a = 2
b = -3
df[linear_path1] = Linear(a,df[M1_path],b,df[M2_path])
assert (df[linear_path1].get_matrix()==a*M1+b*M2).all()
def func(self,target_df,a,X_df,b,Y_df,row_labels=None,col_labels=None):
"""Fetch matrices from dataframes, and return the resulting linear
combination in a dataframe"""
x = X_df.get_matrix()
y = Y_df.get_matrix()
if row_labels==None:
row_labels = X_df._row_index.keys()
if col_labels==None:
col_labels = X_df._col_index.keys()
if (x.shape != y.shape):
raise ValueError
return DataFrame.from_matrix(a*x+b*y,row_labels,col_labels)
def test_sgd(self):
# Also test sgd
close = np.array([[-44.25076083, 38.62854577],
[-38.41473092, 36.29945225],
[-31.43300105, 30.79620632],
[-21.27706071, 24.08638079],
[-14.00259076, 6.54438641],
[11.52354442, -6.07783327],
[48.69374796, -38.64696136],
[95.49682071, -84.38906967]])
df = DataFrame()
path = "row/", "col/"
df["xrow/",
"xcol/"] = DataFrame.from_matrix(np.arange(16).reshape(8, 2))
df["yrow/",
"ycol/"] = DataFrame.from_matrix(np.arange(8).reshape(8, 1))
X_df = df["xrow/", "xcol/"]
y_df = df["yrow/", "ycol/"]
df[path] = SGD(SquareTest, close, y_df, batch_size=8, step_size=0.5)
sleep(1)
df[path].stop()
assert np.allclose(df[path].get_matrix(), y_df.get_matrix())
def test_PCA_basis(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
M2_path = ("row2/", "col2/")
n = 10
m = 5
d = 3
M1 = nprand.rand(n, m)
M1 = M1 - np.mean(M1, axis=0)
# print M1
df[M1_path].set_matrix(M1)
df[M2_path] = PCABasis(df[M1_path], d)
u, s, v_T = numpy.linalg.svd(M1, full_matrices=False)
s[d + 1:] = 0
v = v_T.T[:, :d]
M1_reconstructed = u.dot(np.diag(s).dot(v_T))
# print M1
# print M1_reconstructed
M1_reconstructed2 = M1.dot(v).dot(v.T)
# print M1_reconstructed2
# print M1.dot(v.dot(v.T))
covmat = (1. / (n - 1)) * M1.T.dot(M1)
evs, evmat = scipy.linalg.eig(covmat)
p = np.argsort(evs)[::-1]
evmat_sorted = evmat[:, p][:, :d]
M1_reconstructed3 = M1.dot(evmat_sorted).dot(evmat_sorted.T)
basis = df[M2_path].get_matrix()
for i in range(evmat_sorted.shape[1]):
assert np.isclose(basis[:,i], evmat_sorted[:,i]).all() or \
np.isclose(basis[:,i],-evmat_sorted[:,i]).all()
M3_path = ("row3/", "col3/")
M3 = nprand.rand(2 * n, m)
M3 = M3 - np.mean(M1, axis=0)
df[M3_path].set_matrix(M3)
pca_path = ("pca/", "pca/")
df[pca_path] = PCA(df[M1_path], df[M3_path], d)
pca = df[pca_path].get_matrix()
proj = M3.dot(evmat_sorted)
for i in range(pca.shape[1]):
assert np.isclose(pca[:,i], proj[:,i]).all() or \
np.isclose(pca[:,i],-proj[:,i]).all()
def test_tuple_to_query(self):
df = DataFrame()
# Test conversion of hashable elements to their actual queries
string = "randomstring"
slice_hash, slice_actual = (slice,(2,4,1)), slice(2,4,1)
list_hash, list_actual = (list,(1,2,3,4,5,6)), [1,2,3,4,5,6]
assert df._tuple_element_to_query(string) == string
assert df._tuple_element_to_query(slice_hash) == slice_actual
assert df._tuple_element_to_query(list_hash) == list_actual
assert df._query_to_tuple_element(string) == string
assert df._query_to_tuple_element(slice_actual) == slice_hash
assert df._query_to_tuple_element(list_actual) == list_hash
def test_one_hot_encoding(self):
df = DataFrame()
M1_path = ("row1/", "col1/")
M2_path = ("row2/", "col2/")
n = 10
m = 5
M1 = np.vstack(
[nprand.randint(0, m, (n, 1)),
np.arange(m).reshape(m, 1)])
M2 = np.zeros((n + m, m))
for i in range(n + m):
M2[i, M1[i]] = 1
df[M1_path].set_matrix(M1)
df[M2_path] = OneHotEncoding(df[M1_path])
assert (df[M2_path].get_matrix() == M2).all()
def func(self,
target_df,
a,
X_df,
b,
Y_df,
row_labels=None,
col_labels=None):
"""Fetch matrices from dataframes, and return the resulting linear
combination in a dataframe"""
x = X_df.get_matrix()
y = Y_df.get_matrix()
if row_labels == None:
row_labels = X_df._row_index.keys()
if col_labels == None:
col_labels = X_df._col_index.keys()
if (x.shape != y.shape):
raise ValueError
return DataFrame.from_matrix(a * x + b * y, row_labels, col_labels)
def func(self, target_df, X_df):
return DataFrame.from_matrix(X_df.get_matrix())
def test_simple_query(self):
df = DataFrame.from_matrix(np.arange(6).reshape(2,3))
assert df._is_simple_query()
assert df["row/","col/"]._is_simple_query()
assert df["row/","col/"][:,:]._is_simple_query()
assert df["row/","col/"][0:1,2:3]._is_simple_query()
def func(self, target_df, X_df, Y_df):
x = X_df.get_matrix()
y = Y_df.get_matrix()
row_labels = X_df._row_index.keys()
col_labels = Y_df._col_index.keys()
return DataFrame.from_matrix(x.dot(y), row_labels, col_labels)
def func(self, target_df, X_df):
X = X_df.get_matrix()
X_m = np.mean(X, axis=0)
X_zm = X - X_m # X with zero mean
return DataFrame.from_matrix(X_zm)
def test_simple_query(self):
df = DataFrame.from_matrix(np.arange(6).reshape(2, 3))
assert df._is_simple_query()
assert df["row/", "col/"]._is_simple_query()
assert df["row/", "col/"][:, :]._is_simple_query()
assert df["row/", "col/"][0:1, 2:3]._is_simple_query()
def test_setitem(self):
df = DataFrame()
rows = "row/"
cols = "col/"
M = np.arange(6).reshape(2, 3)
df.__setitem__((slice(None, None, None), slice(None, None, None)), M)
assert (df.get_matrix() == M).all()
df = DataFrame()
df.__setitem__((slice(None, None, None), slice(None, None, None)),
M,
rows=["a", "b"],
cols=["c", "d", "e"])
assert (df.get_matrix() == M).all()
df = DataFrame()
df["x/", "y/"].__setitem__(
(slice(None, None, None), slice(None, None, None)),
M,
rows=["a", "b"],
cols=["c", "d", "e"])
assert (df.get_matrix() == M).all()
assert (df["x/", "y/"].get_matrix() == M).all()
df["x/", "y/"].__setitem__(
(slice(None, None, None), slice(None, None, None)), 2)
assert (df.get_matrix() == 2).all()
def func(self,target_df,X_df,Y_df):
x = X_df.get_matrix()
y = Y_df.get_matrix()
row_labels = X_df._row_index.keys()
col_labels = Y_df._col_index.keys()
return DataFrame.from_matrix(x.dot(y),row_labels,col_labels)
def func(self,target_df,X_df):
X = X_df.get_matrix()
X_m = np.mean(X, axis=0)
X_zm = X - X_m # X with zero mean
return DataFrame.from_matrix(X_zm)
def test_setitem(self):
df = DataFrame()
rows = "row/"
cols = "col/"
M = np.arange(6).reshape(2,3)
df.__setitem__((slice(None,None,None),slice(None,None,None)), M)
assert (df.get_matrix()==M).all()
df = DataFrame()
df.__setitem__((slice(None,None,None),slice(None,None,None)), M,
rows=["a","b"],cols=["c","d","e"])
assert (df.get_matrix()==M).all()
df = DataFrame()
df["x/","y/"].__setitem__((slice(None,None,None),slice(None,None,None)), M,
rows=["a","b"],cols=["c","d","e"])
assert (df.get_matrix() == M).all()
assert (df["x/","y/"].get_matrix() == M).all()
df["x/","y/"].__setitem__((slice(None,None,None),slice(None,None,None)),2)
assert (df.get_matrix() == 2).all()
def func(self,target_df, X_df):
return DataFrame.from_matrix(X_df.get_matrix())