def loss(y, t):
return nn_plankton.log_loss(y, t) + lambda_reg * reg_term
def loss(y, t):
return nn_plankton.log_loss(y, t) + reg_param * (alpha * L1 +
(1 - alpha) * L2)
def loss(y, t):
return nn_plankton.log_loss(y, t) + lambda_reg * reg_term
def loss(y, t):
return nn_plankton.log_loss(y, t) + reg_param*(alpha * L1 + (1-alpha) * L2)
]
del predictions_list
for i in range(n_models):
print(individual_prediction_errors[i],
os.path.basename(valid_predictions_paths[i]))
print()
# optimizing weights
X = theano.shared(predictions_stack) # source predictions
t = theano.shared(utils.one_hot(t_valid)) # targets
W = T.vector('W')
s = T.nnet.softmax(W).reshape((W.shape[0], 1, 1))
weighted_avg_predictions = T.sum(X * s,
axis=0) # T.tensordot(X, s, [[0], [0]])
error = nn_plankton.log_loss(weighted_avg_predictions, t)
grad = T.grad(error, W)
f = theano.function([W], error)
g = theano.function([W], grad)
w_init = np.zeros(n_models, dtype=theano.config.floatX)
out, loss, _ = scipy.optimize.fmin_l_bfgs_b(f,
w_init,
fprime=g,
pgtol=1e-09,
epsilon=1e-08,
maxfun=10000)
weights = np.exp(out)
weights /= weights.sum()
print "Individual prediction errors"
individual_prediction_errors = [utils.log_loss(p, t_valid) for p in predictions_list]
del predictions_list
for i in xrange(n_models):
print individual_prediction_errors[i], os.path.basename(valid_predictions_paths[i])
print
# optimizing weights
X = theano.shared(predictions_stack) # source predictions
t = theano.shared(utils.one_hot(t_valid)) # targets
W = T.vector('W')
s = T.nnet.softmax(W).reshape((W.shape[0], 1, 1))
weighted_avg_predictions = T.sum(X * s, axis=0) # T.tensordot(X, s, [[0], [0]])
error = nn_plankton.log_loss(weighted_avg_predictions, t)
grad = T.grad(error, W)
f = theano.function([W], error)
g = theano.function([W], grad)
w_init = np.zeros(n_models, dtype=theano.config.floatX)
out, loss, _ = scipy.optimize.fmin_l_bfgs_b(f, w_init, fprime=g, pgtol=1e-09, epsilon=1e-08, maxfun=10000)
weights = np.exp(out)
weights /= weights.sum()
print 'Optimal weights'
for i in xrange(n_models):
print weights[i], os.path.basename(valid_predictions_paths[i])
print
