def __init__ ( self,
input,
input_shape,
id,
num_classes = 10,
rng = None,
input_params = None,
borrow = True,
# batch_norm = False,
activation = 'softmax',
verbose = 2
):
super(classifier_layer,self).__init__(id = id, type = 'classifier', verbose = verbose)
if rng is None:
rng = numpy.random
if verbose >= 3:
print("... Creating classifier layer")
# initialize with 0 the weights W as a matrix of shape (n_in, n_out)
self.input = input
# To copy weights previously created or some wierd initializations
if input_params is not None:
self.w = input_params[0]
self.b = input_params[1]
else:
self.w = theano.shared(value=numpy.asarray(0.01 * rng.standard_normal(
size=(input_shape[1], num_classes)),
dtype=theano.config.floatX), name='weights', borrow = borrow)
self.b = theano.shared(value=numpy.zeros((num_classes,),
dtype=theano.config.floatX), name='bias', borrow = borrow)
self.fit = T.dot(input, self.w) + self.b
self.p_y_given_x, softmax_shp = _activate(x = self.fit,
activation = activation,
input_size = num_classes,
verbose = verbose,
dimension = 2)
# compute prediction as class whose probability is maximal in symbolic form
self.inference = self.p_y_given_x # this is where batchnorm test should go.
self.predictions = T.argmax(self.inference, axis=1)
# parameters of the model
self.L1 = abs(self.w).sum()
self.L2 = (self.w ** 2).sum()
self.params = [self.w, self.b]
self.active_params = [self.w, self.b]
self.probabilities = T.log(self.p_y_given_x)
self.output = self.p_y_given_x
self.output_shape = (input_shape[0], num_classes)
self.num_neurons = num_classes
self.activation = activation
self.dropout_rate = 0
self.batch_norm = False
if verbose >= 3:
print("... Classifier layer is created with output shape " + str(self.output_shape))
def test10_abstract_layer_activate_maxout_tuple(self,mock_maxout):
mock_maxout.return_value = (self.input_ndarray,self.input_shape)
self.out,self.out_shape = _activate(self.input_ndarray,("maxout","type",self.input_shape),self.input_shape,self.verbose,**{'dimension':(10,10)})
self.assertTrue(numpy.allclose(self.out,self.input_ndarray))
def test9_abstract_layer_activate_squared(self,mock_squared):
mock_squared.return_value = self.input_ndarray
self.out,self.out_shape = _activate(self.input_ndarray,"squared",self.input_shape,self.verbose)
self.assertTrue(numpy.allclose(self.out,self.input_ndarray))
def test8_abstract_layer_activate_softmax(self,mock_softmax):
mock_softmax.return_value = self.input_ndarray
self.out,self.out_shape = _activate(self.input_ndarray,"softmax",self.input_shape,self.verbose)
self.assertTrue(numpy.allclose(self.out,self.input_ndarray))
def test7_abstract_layer_activate_tanh(self,mock_tanh):
mock_tanh.return_value = self.input_ndarray
self.out,self.out_shape = _activate(self.input_ndarray,"tanh",self.input_shape,self.verbose)
self.assertTrue(numpy.allclose(self.out,self.input_ndarray))
def test5_abstract_layer_activate_abs(self,mock_abs):
mock_abs.return_value = self.input_ndarray
self.out,self.out_shape = _activate(self.input_ndarray,"abs",self.input_shape,self.verbose)
self.assertTrue(numpy.allclose(self.out,self.input_ndarray))
def test4_abstract_layer_activate_relu(self,mock_relu):
mock_relu.return_value = self.input_ndarray
self.out,self.out_shape = _activate(self.input_ndarray,"relu",self.input_shape,self.verbose)
self.assertTrue(numpy.allclose(self.out,self.input_ndarray))