def restart_network(self):
if self.traject == 'combine':
print('MNIST Trajectory MINE network')
self.net = networks.statistical_estimator(traject_max_depth = self.traject_max_depth,
traject_num_layers = self.traject_num_layers, traject_stride = self.traject_stride,
traject_kernel = self.traject_kernel, traject_padding = self.traject_padding,
traject_pooling = self.traject_pooling, number_descending_blocks=self.number_descending_blocks,
number_repeating_blocks = self.number_repeating_blocks, repeating_blockd_size = self.repeating_blockd_size)
elif self.traject == 'traject':
print('Traject MINE network')
self.net = networks.conv1d_classifier_(input_dim=[18,1000,0], output_size = 1,
max_depth = self.traject_max_depth,
p_conv = 0, p_fc = 0, BN = False)
else:
print('MNIST MINE network')
if self.net_num == 1:
self.net = networks.statistical_estimator_DCGAN(input_size = 2, output_size = 1)
elif self.net_num == 2:
self.net = networks.statistical_estimator_DCGAN_2(input_size = 1, output_size = 1)
else:
self.net = networks.statistical_estimator_DCGAN_3(input_size = 1, output_size = 1,number_descending_blocks = self.number_descending_blocks,
number_repeating_blocks=self.number_repeating_blocks, repeating_blockd_size = self.repeating_blockd_size)
print('')
print('Restarted Network')
if self.optimizer == 1:
self.mine_net_optim = optim.SGD(self.net.parameters(), lr = self.lr)
print('Optimizer: SGD')
elif self.optimizer == 2:
self.mine_net_optim = optim.Adam(self.net.parameters(), lr = self.lr)
print('Optimizer: Adam')
else:
self.mine_net_optim = optim.RMSprop(self.net.parameters(), lr = self.lr)
print('Optimizer: RMSprop')
#self.scheduler = optim.lr_scheduler.StepLR(self.mine_net_optim, step_size=10*(len(self.dataset)/self.batch_size), gamma=self.gamma)
print('Learning rate = {}'.format(self.lr))
print('Batch size = {}'.format(self.batch_size))
#print('Gamma of lr decay = {}'.format(self.gamma))
print('Using Net {}'.format(self.net_num))
if self.net_num == 3:
print('Number of Descending Blocks is {}'.format(self.number_descending_blocks))
print('Number of times to repeat a block = {}'.format(self.number_repeating_blocks))
print('The fully connected layer to repeat - {}'.format(self.repeating_blockd_size))
pickle.dump(test, f)
else:
print('didnt find train.pickle')
train, test = utils.create_train_test(dr=data_path)
with open('train.pickle', 'wb') as f:
pickle.dump(train, f)
with open('test.pickle', 'wb') as f:
pickle.dump(test, f)
net = networks.conv1d_classifier_(input_dim=[9, 1000, 0],
output_size=10,
p_conv=p_conv,
p_fc=p_fc,
max_depth=max_depth,
num_layers=num_layers,
repeating_block_depth=repeating_block_depth,
repeating_block_size=repeating_block_size,
stride=stride,
kernel=kernel,
padding=0,
pooling=pooling)
print(net)
print('cuda availble?:', torch.cuda.is_available())
#Defining the optimizer
if type(optimizer) != int:
optimizer = int(optimizer)
if optimizer == 1:
optimizer = optim.SGD(net.parameters(), lr=lr)
print('')
print('Optimizer: SGD')
elif optimizer == 2:
def __init__(self, train = True,traject = True, batch = 1000, lr = 3e-3, gamma = 0.001, optimizer=2, net_num = 3,
traject_max_depth = 512, traject_num_layers = 6, traject_stride = [3,1],
traject_kernel = 5, traject_padding = 0,
traject_pooling = [1,2], number_descending_blocks = 3,
number_repeating_blocks=0, repeating_blockd_size=512,
dataset_status = 'same'):
self.net_num = net_num
self.traject = traject
self.dataset_status = dataset_status
print(self.dataset_status)
self.traject_max_depth = traject_max_depth
self.traject_num_layers = traject_num_layers
self.traject_stride = traject_stride
self.traject_kernel = traject_kernel
self.traject_padding = traject_padding
self.traject_pooling = traject_pooling
self.number_descending_blocks = number_descending_blocks
self.number_repeating_blocks = number_repeating_blocks
self.repeating_blockd_size = repeating_blockd_size
#Defining the netwrok:
if self.traject == 'combined':
print('MNIST Trajectory MINE network')
self.net = networks.statistical_estimator(traject_max_depth = self.traject_max_depth,
traject_num_layers = self.traject_num_layers,
traject_stride = self.traject_stride,
traject_kernel = self.traject_kernel,
traject_padding = self.traject_padding,
traject_pooling = self.traject_pooling,
number_descending_blocks=self.number_descending_blocks,
number_repeating_blocks = self.number_repeating_blocks,
repeating_blockd_size = self.repeating_blockd_size)
elif self.traject == 'traject':
print('Traject MINE network')
self.net = networks.conv1d_classifier_(input_dim=[18,1000,0], output_size = 1,
p_conv = 0, p_fc = 0,
max_depth = self.traject_max_depth,
num_layers = self.traject_num_layers,
conv_depth_type = 'decending',
repeating_block_depth = 5,
repeating_block_size = 0,
stride = self.traject_stride,
kernel = self.traject_kernel,
padding = self.traject_padding,
pooling = self.traject_pooling,
BN = False)
else:
print('MNIST MINE network')
if self.net_num == 1:
self.net = networks.statistical_estimator_DCGAN(input_size = 2, output_size = 1)
elif self.net_num == 2:
self.net = networks.statistical_estimator_DCGAN_2(input_size = 1, output_size = 1)
else:
self.net = networks.statistical_estimator_DCGAN_3(input_size = 1, output_size = 1,number_descending_blocks = self.number_descending_blocks,
number_repeating_blocks=self.number_repeating_blocks, repeating_blockd_size = self.repeating_blockd_size)
#self.input1 = input1
#self.input2 = input2
self.lr = lr
self.optimizer = optimizer
if type(optimizer) != int:
optimizer = int(optimizer)
if self.optimizer == 1:
self.mine_net_optim = optim.SGD(self.net.parameters(), lr = self.lr)
print('')
print('Optimizer: SGD')
elif self.optimizer == 2:
self.mine_net_optim = optim.Adam(self.net.parameters(), lr = self.lr)
print('')
print('Optimizer: Adam')
else:
self.mine_net_optim = optim.RMSprop(self.net.parameters(), lr = self.lr)
print('')
print('Optimizer: SGD')
self.train_value = train
#self.dataloader = torch.utils.data.DataLoader(self.dataset, batch_size = batch, shuffle = True)
self.batch_size = batch
#self.scheduler = optim.lr_scheduler.StepLR(self.mine_net_optim, step_size=10*(len(self.dataset)/self.batch_size), gamma=gamma)
self.gamma = gamma
#self.scheduler2 = optim.lr_scheduler.ReduceLROnPlateau(self.mine_net_optim, mode='max', factor=0.5, patience=10, verbose=False, threshold=0.0001, threshold_mode='abs', cooldown=0, min_lr=0, eps=1e-08)
self.results = []
#print all variables of the system:
print('Learning rate = {}'.format(self.lr))
print('Batch size = {}'.format(self.batch_size))
#print('Gamma of lr decay = {}'.format(self.gamma))
print('Using Net {}'.format(self.net_num))
if self.net_num == 3:
print('Number of Descending Blocks is {}'.format(self.number_descending_blocks))
print('Number of times to repeat a block = {}'.format(self.number_repeating_blocks))
print('The fully connected layer to repeat - {}'.format(self.repeating_blockd_size))