def weight_decay():
with open('./experiment_configs/10_20_config_search.yaml',
'r+') as f_search:
config = yaml.load(f_search)
config['use_growl'] = True
config['growl_params'] = [[1e-20, 1e-20], [3e-1, 3e-3]]
config['use_group_lasso'] = False
config['use_wd'] = True
config['log_dir'] = log_dir
lr = 1e-3
config['learning_rate'] = lr
# for baseline
config['num_epochs'] = 300
config['retraining'] = True
wd = [0.01] * 5
for i, wd_i in enumerate(wd):
config['wd'] = wd[i]
config['res_dir'] = config[
'log_dir'] + '/weight_decay_{}_p_{}_idx_{}'.format(
wd_i, config['preference'], i)
config['plot_dir'] = config['res_dir'] + '/plot_res'
config['summary_dir'] = config['res_dir'] + '/summary_res'
train_mnist(config)
def grlasso_l2():
with open('./experiment_configs/10_20_config_search.yaml',
'r+') as f_search:
config = yaml.load(f_search)
config['use_growl'] = False
config['use_group_lasso'] = True
config['use_wd'] = True
lr = 1e-3
config['learning_rate'] = lr
config['log_dir'] = log_dir
lambda1 = [80] * 5
lambda_wd = [0.001]
preference = [0.6]
for p in preference:
config['preference'] = p
for wd in lambda_wd:
for idx1, lambda1_i in enumerate(lambda1):
config['growl_params'] = [[lambda1_i, 0]] * NUM_LAYER
config['wd'] = wd
config['res_dir'] = config[
'log_dir'] + '/grEN_l1_{}_l_wd_{}_p_{}_idx_{}'.format(
lambda1_i, wd, config['preference'], idx1)
config['plot_dir'] = config['res_dir'] + '/plot_res'
config['summary_dir'] = config['res_dir'] + '/summary_res'
train_mnist(config)
def PLD():
with open('./experiment_configs/10_20_config_search.yaml',
'r+') as f_search:
config = yaml.load(f_search)
config['use_growl'] = True
config['use_group_lasso'] = False
config['use_wd'] = False
config['reg_params_type'] = 'PLD'
PLD_transition = 0.5
preference = 0.6
lr = 1e-3
config['learning_rate'] = lr
config['log_dir'] = log_dir
lambda1 = 87
lambda1 = [12] * 5
lambda2 = [0.215]
lambda_wd = [0]
for idx1, lambda1_i in enumerate(lambda1):
for lambda2_i in lambda2:
for wd in lambda_wd:
config['preference'] = preference
config['wd'] = wd
config['growl_params'] = [[lambda1_i, lambda2_i]
] * NUM_LAYER
config['PLD_transition'] = PLD_transition
config['res_dir'] = config[
'log_dir'] + '/PLDwd_lr_{}_l1_{}_l2_{}_PLDt_{}_p_{}_wd_{}_idx_{}'.format(
lr, lambda1_i, lambda2_i, PLD_transition,
config['preference'], wd, idx1)
config['plot_dir'] = config['res_dir'] + '/plot_res'
config['summary_dir'] = config['res_dir'] + '/summary_res'
train_mnist(config)
def group_lasso():
with open('./experiment_configs/10_20_config_search.yaml',
'r+') as f_search:
config = yaml.load(f_search)
config['use_growl'] = False
config['use_group_lasso'] = True
config['use_wd'] = False
lr = 1e-3
config['learning_rate'] = lr
config['log_dir'] = log_dir
lambda1 = [60] * 5
for idx1, lambda1_i in enumerate(lambda1):
config['growl_params'] = [[lambda1_i, 0]] * NUM_LAYER
config['res_dir'] = config[
'log_dir'] + '/group_lasso_lr_{}_l1_p_{}_{}_idx1_{}'.format(
lr, lambda1_i, config['preference'], idx1)
config['plot_dir'] = config['res_dir'] + '/plot_res'
config['summary_dir'] = config['res_dir'] + '/summary_res'
train_mnist(config)
def baseline():
with open('./experiment_configs/10_20_config_search.yaml',
'r+') as f_search:
config = yaml.load(f_search)
config['use_growl'] = False
config['growl_params'] = [[1e-20, 1e-20], [3e-1, 3e-3]]
config['use_group_lasso'] = False
config['use_wd'] = False
config['log_dir'] = log_dir
lr = 1e-3
config['learning_rate'] = lr
# for baseline
config['num_epochs'] = 400
config['retraining'] = False
for i in range(5):
config['res_dir'] = config['log_dir'] + '/baseline_idx_{}'.format(
i)
config['plot_dir'] = config['res_dir'] + '/plot_res'
config['summary_dir'] = config['res_dir'] + '/summary_res'
train_mnist(config)
def train_and_test():
# load intact line drawings
intact_dataset = KFoldDataset('line_drawings')
X_intact, Y_intact, class_indices = intact_dataset.get_data(
IMG_SIZE, 'grayscale')
# load dR weighted grayscale drawings
# weighted_dataset = KFoldDataset('toronto_dR_weighted')
# X_weighted,Y_weighted,class_indices = weighted_dataset.get_data(IMG_SIZE, 'grayscale')
symmetric_dataset = KFoldDataset('dR_symmetric')
X_sym, Y_sym, class_indices = symmetric_dataset.get_data(
IMG_SIZE, 'grayscale')
asymmetric_dataset = KFoldDataset('dR_asymmetric')
X_asym, Y_asym, class_indices = asymmetric_dataset.get_data(
IMG_SIZE, 'grayscale')
# init 5-fold cross validation
kfold = StratifiedKFold(n_splits=5, shuffle=True)
# combine datasets into single 2-channel set
X = np.ndarray(shape=(X_intact.shape[0:3] + (2, )))
# put each dataset in respective channels
X[:, :, :, 0] = X_intact.squeeze()
# X[:,:,:,1] = X_weighted.squeeze()
# X[:,:,:,1] = X_sym.squeeze()
X[:, :, :, 1] = X_asym.squeeze()
# convert ground-truth one-hot encodings to class labels
labels = [np.argmax(y) for y in Y_intact]
two_channel_scores = [None] * 5
nb_classes = 6
i = 0
for train_idx, test_idx in kfold.split(X, labels):
print('fold ' + str(i + 1) + ' of 5')
input_shape = X.shape[1:]
model = compile_mnist(input_shape, nb_classes)
model = train_mnist(model, X[train_idx], Y_intact[train_idx],
X[test_idx], Y_intact[test_idx])
score = model.evaluate(X[test_idx], Y_intact[test_idx])
two_channel_scores[i] = score[1]
K.clear_session()
i += 1
print(two_channel_scores)
print(np.mean(two_channel_scores))
def train_and_test():
# load rgb images
rgb_dataset = KFoldDataset('toronto_rgb')
X_rgb,Y_rgb,class_indices = rgb_dataset.get_data(IMG_SIZE, 'rgb')
# load intact line drawings
intact_dataset = KFoldDataset('toronto_line_drawings')
X_intact,Y_intact,class_indices = intact_dataset.get_data(IMG_SIZE, COLOR_MODE)
# load symmetric splits
sym_dataset = KFoldDataset('toronto_dR_symmetric')
X_sym,Y_sym,class_indices = sym_dataset.get_data(IMG_SIZE, COLOR_MODE)
# load asymmetric splits
asym_dataset = KFoldDataset('toronto_dR_asymmetric')
X_asym,Y_asym,class_indices = asym_dataset.get_data(IMG_SIZE, COLOR_MODE)
# init 5-fold cross validation
kfold = StratifiedKFold(n_splits=5, shuffle=True)
# combine datasets into single 3-channel set
# new array to hold combined images
X = np.ndarray(shape=(X_intact.shape[0:3] + (3,)))
# put each dataset in respective channels
X[:,:,:,0] = X_intact.squeeze()
X[:,:,:,1] = X_sym.squeeze()
X[:,:,:,2] = X_asym.squeeze()
# convert ground-truth one-hot encodings to class labels
labels = [np.argmax(y) for y in Y_intact]
# train for each fold
rgb_scores = [None] * 5
intact_scores = [None] * 5
three_channel_scores = [None] * 5
nb_classes = 6
i=0
for train_idx, test_idx in kfold.split(X, labels):
print('fold ' + str(i+1) + ' of 5')
# train on rgb images ==================
input_shape = X_rgb.shape[1:]
model = compile_mnist(input_shape, nb_classes)
model = train_mnist(model,
X_rgb[train_idx], Y_rgb[train_idx],
X_rgb[test_idx], Y_rgb[test_idx]
)
score = model.evaluate(X_rgb[test_idx], Y_rgb[test_idx])
rgb_scores[i] = score[1]
# Train on just intact images =====
input_shape = X_intact.shape[1:]
model = compile_mnist(input_shape, nb_classes)
model = train_mnist(model,
X_intact[train_idx], Y_intact[train_idx],
X_intact[test_idx], Y_intact[test_idx]
)
score = model.evaluate(X_intact[test_idx], Y_intact[test_idx])
intact_scores[i] = score[1]
# Train on 3-channel combo images ======
# load model
input_shape = X.shape[1:]
model = compile_mnist(input_shape, nb_classes)
# train model
model = train_mnist(model,
X[train_idx], Y_intact[train_idx],
X[test_idx], Y_intact[test_idx]
)
# get final evaluation
score = model.evaluate(X[test_idx], Y_intact[test_idx])
# append only accuracy
three_channel_scores[i] = score[1]
# clear tensorflow memory
K.clear_session()
i+=1
# print final score
print(rgb_scores)
print(np.mean(rgb_scores))
print(intact_scores)
print(np.mean(intact_scores))
print(three_channel_scores)
print(np.mean(three_channel_scores))
