def _callback(pi):
Y_hat = TeaSpoon.exe(pred, [pi, X])
out = {}
out['MSE'] = np.mean(Y_hat==y)
out['Loss'] = TeaSpoon.exe(loss, [pi, X,y])
opt = {'freq':10}
return out, opt
def callback(pi):
Y_hat = TS.exe(pred_T, [pi, X])
out = {}
out['ACC'] = np.mean(Y_hat==y)
out['Loss'] = TS.exe(loss_T, [pi, X,y])
out['pi'] = 3.14
opt = {'freq':1}
return out, opt
def init(X,y,r=0):
'''
'''
N, F = X.shape
C = len(np.unique(y))
w = np.random.rand(F, C)
b = np.ones(C)
para = {}
para['w'] = TeaSpoon.parameter(w, const=False)
para['b'] = TeaSpoon.parameter(b, const=False)
para['r'] = TeaSpoon.parameter(r, const=True)
return para
def test_default(self):
p1, res = TeaSpoon.optimize(
fun=loss,
p0=p0,
jac=grad,
callback='default',
args=(X, y),
method='BFGS',
options = {'maxiter': 20, 'disp': 0},
)
loss_0 = TeaSpoon.exe(loss, [p0, X, y])
loss_1 = TeaSpoon.exe(loss, [p1, X, y])
assert loss_0 > loss_1
def test_none(self):
p1, res = TeaSpoon.optimize(
fun=loss,
p0=p0,
jac=grad,
args=(X, y),
method='BFGS',
options = {'maxiter': 20, 'disp': 0},
)
def test_default(self):
p1, res = TeaSpoon.optimize(
fun=loss,
p0=p0,
jac=grad,
callback='default',
args=(X, y),
method='BFGS',
options = {'maxiter': 20, 'disp': 0},
)
def test_user_counter(self):
global glob_counter
glob_counter = 0
def _callback(pi):
Y_hat = TeaSpoon.exe(pred, [pi, X])
out = {}
out['MSE'] = np.mean(Y_hat==y)
out['Loss'] = TeaSpoon.exe(loss, [pi, X,y])
opt = {'freq':10}
return out, opt
p1, res = TeaSpoon.optimize(
fun=loss,
p0=p0,
jac=grad,
callback=_callback,
args=(X, y),
method='BFGS',
options = {'maxiter': 20, 'disp': 0},
)
import unittest
import TeaSpoon
import numpy as np
from examples import log_reg
X,y = log_reg.rand_dset()
p0 = log_reg.init(X,y)
pred = TeaSpoon.compile(log_reg.pred, [p0, X], jac=False)
loss, grad = TeaSpoon.compile(log_reg.loss, [p0, X, y], jac=True)
class testcase:
def test_default(self):
p1, res = TeaSpoon.optimize(
fun=loss,
p0=p0,
jac=grad,
callback='default',
args=(X, y),
method='BFGS',
options = {'maxiter': 20, 'disp': 0},
)
loss_0 = TeaSpoon.exe(loss, [p0, X, y])
loss_1 = TeaSpoon.exe(loss, [p1, X, y])
assert loss_0 > loss_1
######################################################
# generate random dataset:
N = 600 # samples
F = 5 # features
C = 4 # labels
y = np.random.randint(0, C, N)
X = np.random.rand(N, F) * np.array([y+1]).T
######################################################
# initial model
w = np.random.uniform(-1,1,(F, C))
b = np.zeros(C)
p0 = {}
p0['w'] = TS.parameter(w, const=False)
p0['b'] = TS.parameter(b, const=False)
# define loss function
def loss(pi, X, y):
'''
compute loss (average negative conditional log likelihood)
'''
w = pi['w'].value
b = pi['b'].value
P = TT.nnet.softmax( T.dot(X, w) + b )
idx = TT.arange(y.shape[0]).astype('int64')
idy = y.astype('int64')
return -TT.mean(TT.log(P[idx, idy]))
