iterator = BatchIterator(10, {
'train_set': 'random_see_all',
'test_set': 'continous'
})
dataset.create_placeholders(iterator, device="/cpu:0")
# Create Network
#---------------
# we use a batch_size of 64 and use the dataset.datum shape to
# obtain the shape of 1 observation and create the input shape
my_layer = layers.custom_layer(ops.Dense, ops.BatchNorm, ops.Activation)
dnn = sknet.Network(name='model_base')
dnn.append(
ops.HermiteSplineConv1D(dataset.signals,
J=5,
Q=16,
K=15,
trainable_scales=False,
trainable_knots=False,
complex=False,
trainable_filters=False,
hilbert=False,
init='gabor'))
#dnn.append(ops.BatchNorm(dnn[-1],[0,2]))
dnn.append(tf.abs(dnn[-1]))
dnn.append(tf.expand_dims(dnn[-1], 1))
def get_distance_dense(input,tensor):
def doit(c):
gradient = tf.gradients(tensor[:,c],input)[0]
norm = tf.sqrt(tf.reduce_sum(tf.square(gradient),[1,2,3]))
return tf.abs(tensor[:,c])/norm
distances = tf.map_fn(doit,tf.range(tensor.shape.as_list()[1]),
dtype=tf.float32)
return tf.reduce_min(distances,0)
# Create Network
#---------------
dnn = sknet.Network(name='simple_model')
if DATA_AUGMENTATION:
dnn.append(ops.RandomAxisReverse(dataset.images,axis=[-1]))
dnn.append(ops.RandomCrop(dnn[-1],(28,28),seed=10))
start_op = 2
else:
dnn.append(dataset.images)
start_op = 1
if MODEL=='cnn':
sknet.networks.ConvSmall(dnn,dataset.n_classes)
elif MODEL=='dense':
dnn.append(sknet.ops.Dense(dnn[-1],4096))
dnn.append(sknet.ops.BatchNorm(dnn[-1],0))
dnn.append(sknet.ops.Activation(dnn[-1],0.1))
# number of points for the grid
N = 350
# grid
TIME = np.linspace(-7, 7, N)
X = np.meshgrid(TIME, TIME)
X = np.stack([X[0].flatten(), X[1].flatten()], 1).astype('float32')
# load the dataset
dataset = sknet.Dataset()
dataset['images/train_set'] = X
dataset.create_placeholders(50, {'train_set': "continuous"}, device="/cpu:0")
# Create Network
#---------------
dnn = sknet.Network()
np.random.seed(18 + int(RUN))
# first layer going from 2d to 9 units
W = np.random.randn(2, 4).astype('float32') / np.sqrt(3)
b = np.random.randn(4).astype('float32') / 3
dnn.append(ops.Dense(dataset.images, 4, W=W, b=b))
dnn.append(ops.Activation(dnn[-1], 0.05))
# following layers with same input and output dimension
for i in range(1):
W = np.random.randn(4, 4).astype('float32') / np.sqrt(3)
b = np.random.randn(4).astype('float32') / 3
dnn.append(ops.Dense(dnn[-1], 4, W=W, b=b))
dnn.append(ops.Activation(dnn[-1], 0.05))