for p in [0] + [pp for pp in np.logspace(-4, 0, 30)]:
filepath = os.path.join(os.getcwd(), 'results', 'mnist_5layer_sweep',
'trial_p%.06f' % (p))
os.mkdir(filepath)
log_file = open(os.path.join(filepath, 'log.txt'), 'a')
sys.stdout = log_file
trial['topology']['bypass p'] = float(p)
per_layer_rates = []
correction_matrices = []
training_errors = []
test_errors = []
states = []
n_epochs = trial['number of epochs']
Network = eqp.Network(trial['topology'], trial['hyperparameters'],
trial['configuration'], trial['dataset'])
initial_W = Network.W.clone().cpu().squeeze().numpy()
initial_W_mask = Network.W_mask.clone().cpu().squeeze().numpy()
with open(os.path.join(filepath, 'init.pickle'), 'wb') as F:
pickle.dump(
{
'trial settings': trial,
'initial weight': initial_W,
'initial mask': initial_W_mask
}, F)
Network.calculate_test_error()
training_errors.append(Network.test_error)
test_errors.append(Network.test_error)
mean_dW = torch.zeros(Network.W.shape).to(trial['configuration']['device'])
assert False"""
with open(os.path.join(filepath, 'init.pickle'), 'wb') as F:
pickle.dump(
{
'topology': topology,
'hyperparameters': hyperparameters,
'configuration': configuration
}, F)
per_layer_rates = []
training_errors = []
test_errors = []
n_epochs = 10
Network = eqp.Network(topology, hyperparameters, configuration,
datasets.Diabetes)
Network.calculate_test_error()
training_errors.append(Network.test_error)
test_errors.append(Network.test_error)
for epoch_idx in range(n_epochs):
print('Starting epoch %d.' % (epoch_idx + 1))
t0 = time.time()
Network.train_epoch()
Network.calculate_test_error()
training_errors.append(Network.training_error)
test_errors.append(Network.test_error)
per_layer_rates.append([])
for i in range(len(Network.interlayer_connections)):
per_layer_rates[-1].append(
np.mean([p[i] for p in Network.per_layer_rates]))
with open(os.path.join(filepath, 'e%d.pickle' % (epoch_idx)), 'wb') as F: