def trte_split(X, Y, tr_frac):
"""Split the data in X/Y into training and testing portions."""
if gp.is_garray(X):
X = X.as_numpy_array()
else:
X = np.array(X)
if gp.is_garray(Y):
Y = Y.as_numpy_array()
else:
Y = np.array(Y)
obs_count = X.shape[0]
obs_dim = X.shape[1]
tr_count = round(tr_frac * obs_count)
te_count = obs_count - tr_count
Xtr = np.zeros((tr_count, X.shape[1]))
Ytr = np.zeros((tr_count, Y.shape[1]))
Xte = np.zeros((te_count, X.shape[1]))
Yte = np.zeros((te_count, Y.shape[1]))
idx = npr.permutation(range(obs_count))
# Basic manual iteration
for i in range(obs_count):
if (i < tr_count):
Xtr[i,:] = X[idx[i],:]
Ytr[i,:] = Y[idx[i],:]
else:
Xte[(i - tr_count),:] = X[idx[i],:]
Yte[(i - tr_count),:] = Y[idx[i],:]
return [gp.garray(Xtr), gp.garray(Ytr), gp.garray(Xte), gp.garray(Yte)]
def trte_split(X, Y, tr_frac):
"""Split the data in X/Y into training and testing portions."""
if gp.is_garray(X):
X = X.as_numpy_array()
else:
X = np.array(X)
if gp.is_garray(Y):
Y = Y.as_numpy_array()
else:
Y = np.array(Y)
obs_count = X.shape[0]
obs_dim = X.shape[1]
tr_count = round(tr_frac * obs_count)
te_count = obs_count - tr_count
Xtr = np.zeros((tr_count, X.shape[1]))
Ytr = np.zeros((tr_count, Y.shape[1]))
Xte = np.zeros((te_count, X.shape[1]))
Yte = np.zeros((te_count, Y.shape[1]))
idx = npr.permutation(range(obs_count))
# Basic manual iteration
for i in range(obs_count):
if (i < tr_count):
Xtr[i, :] = X[idx[i], :]
Ytr[i, :] = Y[idx[i], :]
else:
Xte[(i - tr_count), :] = X[idx[i], :]
Yte[(i - tr_count), :] = Y[idx[i], :]
return [gp.garray(Xtr), gp.garray(Ytr), gp.garray(Xte), gp.garray(Yte)]
def vector_weights(self, Wm=gp.garray(())):
"""Return the weights in Wm or self.W, vectorized."""
if (Wm.size == 0):
Wm = self.W
if not gp.is_garray(Wm):
Wm = gp.garray(Wm)
Wv = Wm.reshape((Wm.size, 1))
return Wv
def vector_weights(self, Wm=gp.garray(())):
"""Return the weights in Wm or self.W, vectorized."""
if (Wm.size == 0):
Wm = self.W
if not gp.is_garray(Wm):
Wm = gp.garray(Wm)
Wv = Wm.reshape((Wm.size, 1))
return Wv
def array(x, dtype=None, **kwargs):
if gnp.is_garray(x):
if dtype is gpu_float32:
return x
else:
return np.array(gnp.as_numpy_array(x), dtype=dtype, **kwargs)
else:
if dtype is gpu_float32:
return gnp.as_garray(np.array(x, **kwargs))
else:
return np.array(x, dtype=dtype, **kwargs)
def matrix_weights(self, Wv=gp.garray(())):
"""Return the weights in Wv, or self.W, matrized."""
if (Wv.size == 0):
Wm = self.Wm
else:
if not gp.is_garray(Wv):
Wv = gp.garray(Wv)
if (Wv.size != self.weight_count()):
raise Exception('Wrong-sized Wv.')
Wm = Wv.reshape((self.dim_output, self.dim_input))
return Wm
def matrix_weights(self, Wv=gp.garray(())):
"""Return the weights in Wv, or self.W, matrized."""
if (Wv.size == 0):
Wm = self.Wm
else:
if not gp.is_garray(Wv):
Wv = gp.garray(Wv)
if (Wv.size != self.weight_count()):
raise Exception('Wrong-sized Wv.')
Wm = Wv.reshape((self.dim_output,self.dim_input))
return Wm
def set_weights(self, Ws):
"""Set weights for this network to those in the array Ws.
Each weight array Wi = Ws[i] should be of the proper size for
parameterizing the LNLayer at self.layers[i].
"""
for i in range(self.layer_count):
if not gp.is_garray(Ws[i]):
Ws[i] = gp.garray(Ws[i])
for i in range(self.layer_count):
self.layers[i].set_weights(Ws[i])
return
def set_weights(self, Wm):
"""Set weights in this layer to the given values.
This performs a copy, so modifications of the given Wm, e.g. during
network training, won't affect the values set for self.W.
"""
if ((Wm.shape[0] != self.dim_output) or \
(Wm.shape[1] != self.dim_input)):
raise Exception('Wrong-sized Wm.')
if not gp.is_garray(Wm):
Wm = gp.garray(Wm)
self.W = Wm.copy()
return
def set_weights(self, Wm):
"""Set weights in this layer to the given values.
This performs a copy, so modifications of the given Wm, e.g. during
network training, won't affect the values set for self.W.
"""
if ((Wm.shape[0] != self.dim_output) or \
(Wm.shape[1] != self.dim_input)):
raise Exception('Wrong-sized Wm.')
if not gp.is_garray(Wm):
Wm = gp.garray(Wm)
self.W = Wm.copy()
return
def class_cats(Yi):
"""Change +1/-1 class indicator matrix to categorical vector."""
if not gp.is_garray(Yi):
Yi = gp.garray(Yi)
Yc = Yi.argmax(axis=1)
return Yc
def class_cats(Yi):
"""Change +1/-1 class indicator matrix to categorical vector."""
if not gp.is_garray(Yi):
Yi = gp.garray(Yi)
Yc = Yi.argmax(axis=1)
return Yc
