# ======================================================================
def ffnet(indim, outdim):
outdim = int(outdim)
l_in = lasagne.layers.InputLayer(shape=(None,) + indim)
l_in = lasagne.layers.FlattenLayer(l_in)
l_in = lasagne.layers.DropoutLayer(l_in, p=0.3)
l_hid = lasagne.layers.DenseLayer(l_in, num_units=256,
nonlinearity=lasagne.nonlinearities.rectify)
l_hid = lasagne.layers.DropoutLayer(l_hid, p=0.3)
return lasagne.layers.DenseLayer(l_hid, num_units=outdim,
nonlinearity=lasagne.nonlinearities.softmax)
# ======================================================================
# Load data
# ======================================================================
m = dnntoolkit.model('tmp/tmp.ai')
m.set_model(ffnet, api='lasagne',
indim=ds['X_train'].shape[1:],
outdim=ds['y_train'].shape[1])
net = m.create_model()
y = tensor.matrix(name='y', dtype=theano.config.floatX)
input_var = [l.input_var for l in lasagne.layers.find_layers(net, types=lasagne.layers.InputLayer)]
dropout = lasagne.layers.find_layers(net, types=lasagne.layers.DropoutLayer)
# ====== Create prediction ====== #
y_pred_deter = lasagne.layers.get_output(net, deterministic=True)
f_pred = theano.function(
inputs=input_var,
outputs=y_pred_deter,
allow_input_downcast=True
l_in = lasagne.layers.DropoutLayer(l_in, p=0.3)
l_hid = lasagne.layers.DenseLayer(l_in, num_units=128)
l_hid = lasagne.layers.DropoutLayer(l_hid, p=0.3)
l_out = lasagne.layers.DenseLayer(l_hid, num_units=10,
nonlinearity=lasagne.nonlinearities.softmax)
return l_out
# ===========================================================================
# Test
# ===========================================================================
ds = dnntoolkit.dataset.load_mnist()
print(ds)
m = dnntoolkit.model('/volumes/backup/model/mnist.ai')
m.set_model(test, 'lasagne')
ai = m.create_model()
input_var = [i.input_var
for i in lasagne.layers.find_layers(ai, types=lasagne.layers.InputLayer)]
y = tensor.matrix('y')
y_pred_stoch = lasagne.layers.get_output(ai, deterministic=False)
y_pred_deter = lasagne.layers.get_output(ai, deterministic=True)
cost_monitor = lasagne.objectives.categorical_accuracy(y_pred_deter, y).mean()
cost_train = lasagne.objectives.categorical_crossentropy(y_pred_stoch, y).mean()
regu = lasagne.regularization.L1L2(l2=dnntoolkit.dnn.calc_weights_decay(m.get_nweights()))
cost_regu = regu.apply_network(ai)
cost_train += cost_regu
