def main(args):
if args.X_path[:-3] == 'csv':
X = np.loadtxt(args.X_path, delimiter=',')
# elif args.X_path[:-3] == 'xls':
else:
X = pandas.read_excel(args.X_path)
X = np.array(X)
W_est = notears.notears_linear_l1(X, lambda1=args.lambda1, loss_type=args.loss_type)
assert utils.is_dag(W_est)
np.savetxt(args.W_path, W_est, delimiter=',')
plot_matrix.plot_matrix(args)
if h_new > 0.25 * h:
rho *= 10
else:
break
w_est, h = w_new, h_new
alpha += rho * h
if h <= h_tol or rho >= rho_max:
break
W_est = _adj(w_est)
W_est[np.abs(W_est) < w_threshold] = 0
return W_est
if __name__ == '__main__':
import utils as ut
ut.set_random_seed(1)
n, d, s0, graph_type, sem_type = 100, 20, 20, 'ER', 'gauss'
B_true = ut.simulate_dag(d, s0, graph_type)
W_true = ut.simulate_parameter(B_true)
np.savetxt('W_true.csv', W_true, delimiter=',')
X = ut.simulate_linear_sem(W_true, n, sem_type)
np.savetxt('X.csv', X, delimiter=',')
W_est = notears_linear_l1(X, lambda1=0.1, loss_type='l2')
assert ut.is_dag(W_est)
np.savetxt('W_est.csv', W_est, delimiter=',')
acc = ut.count_accuracy(B_true, W_est != 0)
print(acc)
def main(args):
starttime = datetime.datetime.now()
torch.set_default_dtype(torch.double)
np.set_printoptions(precision=3)
# load data
X_envs, X_test = utils.load_data(dataset_name=args.dataset_name,
preprocess=args.preprocess,
num_envs=args.num_envs)
# set target variable Y
y_index = args.y_index
vertex_label = utils.load_vertex_label(dataset_name=args.dataset_name)
# adjust dataset X according to y_index
X_swap = []
for X_env in X_envs:
swap(X_env, 0, y_index)
X_swap.append(X_env)
X = np.stack(X_swap)
# swap X_test
swap(X_test, 0, y_index)
# we do not need swap label because when drawing, we keep the original order.
# Build model
n_envs, n, d = X.shape
Y_pred_hidden = [args.Y_hidden] * (args.Y_hidden_layer - 1)
notears_hidden = [args.hidden] * args.notears_hidden_layer
model = isl.isl_module(n_envs=n_envs,
Y_dims=[d, args.hidden] + Y_pred_hidden + [1],
dims=[d] + notears_hidden + [1],
bias=True)
if args.continue_path:
pretrained_path = args.continue_path
model.load_state_dict(torch.load(pretrained_path))
model.to(device)
# conduct Invariant Structure Learning (ISL)
y_pred_loss, w_est_origin_envs = isl.notears_nonlinear(
model,
X,
y_index=0,
lambda1=args.lambda1,
lambda2=args.lambda2,
lambda1_Y=args.lambda1_Y,
lambda2_Y_fc1=args.lambda2_Y_fc1,
lambda2_Y_fc2=args.lambda2_Y_fc2,
w_threshold=0,
beta=args.beta) # y_index always be 0, w-thresh always be zero
# tune W for the best shd score
g, w_true = utils.load_w_true(args.dataset_name)
# save results
os.makedirs(args.Output_path, exist_ok=True)
for i in range(len(w_est_origin_envs)): # for each environment
swap_row_column(w_est_origin_envs[i], 0,
y_index) # recover to original order
if args.Only_Y_DAG: # for boston housing, we only care Y related DAG
w_est_origin_envs[i][:, 1:] = np.zeros(w_true[:, 1:].shape)
# rank W
wthresh_w_shd_dict = utils.rank_W(w_est_origin_envs[i],
10,
90,
w_true,
w_threshold_low=0,
w_threshold_high=10,
step_size=0.02)
# select W
w_est = None
w_threshold = None
for threshold, w_element in wthresh_w_shd_dict.items():
if utils.is_dag(w_element[0]):
w_est = w_element[0]
w_threshold = threshold
break
exp = (args.Output_path + (
'pretrain-IRM_BH_norm={}_wthreshold={}_beta={}_y_index={}_lam1' +
'={}_lam2fc1={}_lam2fc2_laterY{}NO{}_Yhid={}').format(
args.normalization, w_threshold, args.beta, args.y_index,
args.lambda1_Y, args.lambda2_Y_fc1, args.lambda2_Y_fc2,
args.Y_hidden_layer, args.notears_hidden_layer, args.Y_hidden))
utils.save_dag(w_est, f'{exp}_{i}', vertex_label=vertex_label)
np.savetxt(f'{exp}_West_env_{i}.csv', w_est, fmt='%.2f', delimiter=',')
np.savetxt(f'{exp}_West_origin_env_{i}.csv',
w_est_origin_envs[i],
fmt='%.2f',
delimiter=',')
# save accuracy
acc = metrics.count_accuracy(w_true, w_est != 0)
# acc = {}
y = model.test(X_test)
mse = squared_loss(y[:, 0], X_test[:, 0])
acc['mse'] = mse
print(acc)
with open(f'{exp}_metrics.json', 'w+') as f:
json.dump(acc, f)
# save and load model weights
model_save_path = exp + '_save_model.pt'
torch.save(model.state_dict(), model_save_path)
endtime = datetime.datetime.now()
print('total time is ', (endtime - starttime).seconds)
def main(args):
torch.set_default_dtype(torch.double)
np.set_printoptions(precision=3)
# load data
X = utils.load_data(dataset_name=args.dataset_name, num_envs=args.num_envs)
# load vertex labels
vertex_label = utils.load_vertex_label(dataset_name=args.dataset_name)
# the initial potential causal parents for each variable
if args.dataset_name == 'Insurance':
Init_Relation = {
'PropCost': ['ThisCarCost', 'OtherCarCost'], # 0
'GoodStudent': ['Age', 'SocioEcon'], # 1
'Age': ['GoodStudent', 'RiskAversion'], # 2
'SocioEcon': ['RiskAversion', 'MakeModel', 'HomeBase'], # 3
'RiskAversion':
['SocioEcon', 'DrivQuality', 'DrivingSkill', 'HomeBase'], # 4
'VehicleYear':
['SocioEcon', 'MakeModel', 'Antilock', 'CarValue', 'Airbag'], # 5
'ThisCarDam': ['RuggedAuto', 'Accident', 'ThisCarCost'], # 6
'RuggedAuto':
['VehicleYear', 'MakeModel', 'Antilock', 'Cushioning'], # 7
'Accident': ['ThisCarDam', 'RuggedAuto'], # 8
'MakeModel':
['SocioEcon', 'VehicleYear', 'RuggedAuto', 'CarValue'], # 9
'DrivQuality': ['RiskAversion', 'DrivingSkill'], # 10
'Mileage': ['MakeModel', 'CarValue'], # 11
'Antilock': ['VehicleYear', 'MakeModel'], # 12
'DrivingSkill': ['Age', 'DrivQuality'], # 13
'SeniorTrain': ['Age', 'RiskAversion'], # 14
'ThisCarCost': ['RuggedAuto', 'CarValue'], # 15
'Theft':
['ThisCarDam', 'MakeModel', 'CarValue', 'HomeBase',
'AntiTheft'], # 16
'CarValue': ['VehicleYear', 'MakeModel'], # 17
'HomeBase': ['SocioEcon', 'RiskAversion'], # 18
'AntiTheft': ['SocioEcon', 'RiskAversion'], # 19
'OtherCarCost': ['RuggedAuto', 'Accident'], # 20
'OtherCar': ['SocioEcon'], # 21
'MedCost': ['Age', 'Accident', 'Cushioning'], # 22
'Cushioning': ['RuggedAuto', 'Airbag'], # 23
'Airbag': ['VehicleYear'], # 24
'ILiCost': ['Accident'], # 25
'DrivHist': ['RiskAversion', 'DrivingSkill'], # 26
}
elif args.dataset_name == 'Sachs':
Init_Relation = {
'Raf': ['Mek'],
'Mek': ['Raf'],
'Plcg': ['PIP2'],
'PIP2': ['Plcg', 'PIP3'],
'PIP3': ['Plcg', 'PIP2'],
'Erk': ['Akt', 'Mek', 'PKA'],
'Akt': ['PKA', 'Erk'],
'PKA': ['Akt'],
'PKC': ['P38'],
'P38': ['PKA', 'PKC'],
'Jnk': ['PKC'],
}
Init_DAG = np.zeros((len(vertex_label), len(vertex_label)))
for x in vertex_label:
x_index = vertex_label.index(x)
for y in Init_Relation[x]: # for each causal parent of x
y_index = vertex_label.index(y)
Init_DAG[y_index, x_index] = 1
n, d = X.shape
model = isl.notearsmlp_self_supervised(dims=[d, args.hidden, 1],
bias=True,
Init_DAG=Init_DAG)
model.to(device)
# To use a different feature as label, change y_index to column index of the
# feature.
w_est_origin = isl.notears_nonlinear(model,
X,
lambda1=args.lambda1,
lambda2=args.lambda2,
w_threshold=0) # keep origin w_est
# tune W for the best shd score
g, w_true = utils.load_w_true(args.dataset_name)
wthresh_w_shd_dict = utils.rank_W(w_est_origin,
10,
90,
w_true,
w_threshold_low=0,
w_threshold_high=10,
step_size=0.01)
w_est = None
w_threshold = None
for threshold, w_element in wthresh_w_shd_dict.items():
if utils.is_dag(w_element[0]):
w_est = w_element[0]
w_threshold = threshold
break
exp = (args.Output_path +
'notear_mlp_sachs_norm={}wthreshold={}lambda1={}lambda2={}'.format(
args.normalization, w_threshold, args.lambda1, args.lambda2))
os.makedirs(args.Output_path, exist_ok=True)
np.savetxt(exp + '_w_est_origin.csv', w_est_origin, delimiter=',')
np.savetxt(exp + '_w_est.csv', w_est, delimiter=',')
utils.save_dag(w_est, exp + '_w_est', vertex_label=vertex_label)
acc = metrics.count_accuracy(w_true, w_est != 0)
print(acc)
y = model(torch.from_numpy(X))
y = y.cpu().detach().numpy()
mse = squared_loss(y[:, 0], X[:, 0])
print('mse:', mse)
acc['mse'] = mse
with open(f'{exp}_metrics.json', 'w+') as f:
json.dump(acc, f)
def train(X, b_true, vertex_label, model_type, args):
"""Trains the model."""
n_envs, n, d = X.shape
os.makedirs(args.Output_path, exist_ok=True)
if model_type == 'notear-mlp':
X = np.vstack(X)
model = nonlinear.NotearsMLP(dims=[d, args.hidden, 1], bias=True)
w_est_origin = nonlinear.notears_nonlinear(model,
X,
lambda1=args.lambda1,
lambda2=args.lambda2,
w_threshold=0)
wthresh_w_shd_dict = utils.rank_W(w_est_origin,
2,
10,
b_true,
w_threshold_low=0,
w_threshold_high=5,
step_size=0.02)
w_est = None
w_threshold = None
for threshold, w_element in wthresh_w_shd_dict.items():
if utils.is_dag(w_element[0]):
w_est = w_element[0]
w_threshold = threshold
break
exp = f'notear_mlp_syn_{args.synthetic_source}_W-thresh{round(w_threshold, 3)}'
# save the learned W matrix
np.savetxt(args.Output_path + f'{exp}_West.csv',
w_est,
fmt='%.3f',
delimiter=',')
np.savetxt(args.Output_path + f'{exp}_WOrigin.csv',
w_est_origin,
fmt='%.3f',
delimiter=',')
# save the learned DAG
utils.save_dag(w_est,
args.Output_path + f'{exp}_DAG',
vertex_label=vertex_label)
# count the accuracy of Y
b_true_Y = np.zeros_like(b_true)
b_true_Y[:, 0] = b_true[:, 0]
w_est_Y = np.zeros_like(w_est)
w_est_Y[:, 0] = w_est[:, 0]
acc = metrices.count_accuracy(b_true_Y, w_est_Y != 0)
metrics[f'{exp}_train_acc'] = acc
y = model(torch.from_numpy(X))
y = y.cpu().detach().numpy()
mse = mean_squared_loss(y.shape[0], y[:, 0], X[:, 0])
print('mse:', mse)
metrics[f'{model_type}_train_MSE'] = mse
elif model_type == 'notear-castle':
X = np.vstack(X)
model = nonlinear_castle.NotearsMLP(dims=[d, args.hidden, 1],
bias=True)
# To use a different feature as label, change y_index to column
# index of the feature.
w_est_origin = nonlinear_castle.notears_nonlinear(
model,
X,
lambda1=args.lambda1,
lambda2=args.lambda2,
w_threshold=args.w_threshold)
wthresh_w_shd_dict = utils.rank_W(w_est_origin,
2,
10,
b_true,
w_threshold_low=0,
w_threshold_high=5,
step_size=0.02)
w_est = None
w_threshold = None
for threshold, w_element in wthresh_w_shd_dict.items():
if utils.is_dag(w_element[0]):
w_est = w_element[0]
w_threshold = threshold
break
exp = f'notear_castle_syn_{args.synthetic_source}_W-thresh-{round(w_threshold, 3)}'
# save the learned W matrix
np.savetxt(args.Output_path + f'{exp}_West.csv',
w_est,
fmt='%.2f',
delimiter=',')
np.savetxt(args.Output_path + f'{exp}_WOrigin.csv',
w_est_origin,
fmt='%.3f',
delimiter=',')
# save the learned DAG
utils.save_dag(w_est,
args.Output_path + f'{exp}_DAG',
vertex_label=vertex_label)
# estimate the accuracy of Y
b_true_Y = np.zeros_like(b_true)
b_true_Y[:, 0] = b_true[:, 0]
w_est_Y = np.zeros_like(w_est)
w_est_Y[:, 0] = w_est[:, 0]
acc = metrices.count_accuracy(b_true_Y, w_est_Y != 0)
metrics[f'{exp}_train_acc'] = acc
y = model(torch.from_numpy(X))
y = y.cpu().detach().numpy()
mse = mean_squared_loss(y.shape[0], y[:, 0], X[:, 0])
print('mse:', mse)
metrics[f'{model_type}_train_MSE'] = mse
elif model_type == 'ISL':
model = isl.isl_module(n_envs=n_envs,
Y_dims=[d, args.hidden, 1],
dims=[d, args.hidden, 1],
bias=True)
model.to(device)
# To use a different feature as label, change y_index to column index of
# the feature.
_, w_est_origin_envs = isl.notears_nonlinear(
model,
X,
y_index=0,
lambda1=args.lambda1,
lambda2=args.lambda2,
lambda1_Y=args.lambda1_Y,
lambda2_Y_fc1=args.lambda2_Y_fc1,
lambda2_Y_fc2=args.lambda2_Y_fc2,
beta=args.beta,
w_threshold=0)
for i in range(len(w_est_origin_envs)):
wthresh_w_shd_dict = utils.rank_W(w_est_origin_envs[i],
2,
30,
b_true,
w_threshold_low=0,
w_threshold_high=5,
step_size=0.02)
w_est = None
w_threshold = None
for threshold, w_element in wthresh_w_shd_dict.items():
if utils.is_dag(w_element[0]):
w_est = w_element[0]
w_threshold = threshold
break
exp = f'notear_ISL_syn_{args.synthetic_source}_W-thresh{round(w_threshold, 3)}'
# save the leared W matrix
np.savetxt(args.Output_path + f'{exp}_West_{i}.csv',
w_est,
fmt='%.3f',
delimiter=',')
np.savetxt(args.Output_path + f'{exp}_WOrigin_env_{i}.csv',
w_est_origin_envs[i],
fmt='%.3f',
delimiter=',')
# save the learned DAG
utils.save_dag(w_est,
args.Output_path + f'{exp}_DAG_{i}',
vertex_label=vertex_label)
# count the accuracy of Y
b_true_Y = np.zeros_like(b_true)
b_true_Y[:, 0] = b_true[:, 0]
w_est_Y = np.zeros_like(w_est)
w_est_Y[:, 0] = w_est[:, 0]
acc = metrices.count_accuracy(b_true_Y, w_est_Y != 0)
print(acc)
metrics[f'{exp}_train_env_{i}_acc'] = acc
y = model.test(X)
mse = mean_squared_loss(y.shape[0] * y.shape[1], y,
X[:, :, 0][:, :, np.newaxis])
print('mse:', mse)
metrics[f'{exp}_train_env_{i}_mse'] = mse
else:
ValueError('Invalid model type')
return model, w_est
utils.set_random_seed(5)
n = 1000 # samples
n_i = 8 # data points (X) per sample
d = 8 # DAG vertices
s0 = 8 # DAG edges
e = 20 # epigenetic markers
k = 4 # archetypes
graph_type, sem_type = 'ER', 'gauss'
# Create network and epigenetic archetypes
W_k = np.ones((k, d, d))
e_k = np.ones((k, e))
# Ensure any combination of networks is DAG
while not utils.is_dag(np.sum(W_k, axis=0)):
for j in range(k):
B_true = utils.simulate_dag(d, s0, graph_type)
W_true = utils.simulate_parameter(B_true)
W_k[j] = W_true
e_k[j] = utils.simulate_epigenome(e)
np.savez('outputs/archetypes.npz', W_k=W_k, e_k=e_k)
# Generate n samples from k archetypes
subtypes = np.zeros((n, k))
W_n = np.zeros((n, d, d))
e_n = np.zeros((n, e))
X_n = np.zeros((n, n_i, d))
for i in range(n):
weights = np.random.uniform(0, 1, k)
weights /= np.sum(weights)
lag_range = range(1, 12 + 1)
all_variable_names = list(df.columns)
plotname = "TemporalModel__" + file
plotname_REDUCED = "TemporalModel*REDUCED*__" + file
#lambda1 used in example provided by authors is 0.1
#default w_threshold=0.3
for lambda1 in [0.1]:
for w_threshold in [0.3]:
W_est = notears.notears_linear_l1(X.copy(),
lambda1=lambda1,
loss_type='l2',
w_threshold=w_threshold)
if not utils.is_dag(W_est):
plotname_add = '!!!NODAG!!!'
else:
plotname_add = ''
np.savetxt('AdjMatrix_' + plotname + '_lambda1=' + str(lambda1) +
'_Wthresh=' + str(w_threshold) + plotname_add + '.csv',
W_est,
delimiter=',')
plot_graph.visualize_temporal(
np.around(W_est, 2),
feature_names=feature_names,
full_feature_names=full_feature_names,
lag_range=lag_range,
all_variable_names=all_variable_names,
def count_accuracy(B_true, B_est):
"""Computes various accuracy metrics for B_est.
true positive = predicted association exists in condition in correct
direction
reverse = predicted association exists in condition in opposite direction
false positive = predicted association does not exist in condition
Args:
B_true (np.ndarray): [d, d] ground truth graph, {0, 1}
B_est (np.ndarray): [d, d] estimate, {0, 1, -1}, -1 is undirected edge in
CPDAG.
Returns:
fdr: (reverse + false positive) / prediction positive
tpr: (true positive) / condition positive
fpr: (reverse + false positive) / condition negative
shd: undirected extra + undirected missing + reverse
nnz: prediction positive
"""
if (B_est == -1).any(): # cpdag
if not ((B_est == 0) | (B_est == 1) | (B_est == -1)).all():
raise ValueError('B_est should take value in {0,1,-1}')
if ((B_est == -1) & (B_est.T == -1)).any():
raise ValueError('undirected edge should only appear once')
else: # dag
if not ((B_est == 0) | (B_est == 1)).all():
raise ValueError('B_est should take value in {0,1}')
if not utils.is_dag(B_est):
raise ValueError('B_est should be a DAG')
d = B_true.shape[0]
# linear index of nonzeros
pred_und = np.flatnonzero(B_est == -1)
pred = np.flatnonzero(B_est == 1)
cond = np.flatnonzero(B_true)
cond_reversed = np.flatnonzero(B_true.T)
cond_skeleton = np.concatenate([cond, cond_reversed])
# true pos
true_pos = np.intersect1d(pred, cond, assume_unique=True)
# treat undirected edge favorably
true_pos_und = np.intersect1d(pred_und, cond_skeleton, assume_unique=True)
true_pos = np.concatenate([true_pos, true_pos_und])
# false pos
false_pos = np.setdiff1d(pred, cond_skeleton, assume_unique=True)
false_pos_und = np.setdiff1d(pred_und, cond_skeleton, assume_unique=True)
false_pos = np.concatenate([false_pos, false_pos_und])
# reverse
extra = np.setdiff1d(pred, cond, assume_unique=True)
reverse = np.intersect1d(extra, cond_reversed, assume_unique=True)
# compute ratio
pred_size = len(pred) + len(pred_und)
cond_neg_size = 0.5 * d * (d - 1) - len(cond)
fdr = float(len(reverse) + len(false_pos)) / max(pred_size, 1)
tpr = float(len(true_pos)) / max(len(cond), 1)
fpr = float(len(reverse) + len(false_pos)) / max(cond_neg_size, 1)
# structural hamming distance
pred_lower = np.flatnonzero(np.tril(B_est + B_est.T))
cond_lower = np.flatnonzero(np.tril(B_true + B_true.T))
extra_lower = np.setdiff1d(pred_lower, cond_lower, assume_unique=True)
missing_lower = np.setdiff1d(cond_lower, pred_lower, assume_unique=True)
shd = len(extra_lower) + len(missing_lower) + len(reverse)
return {'fdr': fdr, 'tpr': tpr, 'fpr': fpr, 'shd': shd, 'nnz': pred_size}
rho *= 10
else:
break
w_est, h = w_new, h_new
alpha += rho * h
if h <= h_tol or rho >= rho_max:
break
W_est = _adj(w_est)
W_est[np.abs(W_est) < w_threshold] = 0
return W_est
if __name__ == '__main__':
import utils
utils.set_random_seed(1)
n, d, s0, graph_type, sem_type = 100, 20, 20, 'ER', 'gauss'
B_true = utils.simulate_dag(d, s0, graph_type)
W_true = utils.simulate_parameter(B_true)
np.savetxt('W_true.csv', W_true, delimiter=',')
X = utils.simulate_linear_sem(W_true, n, sem_type)
np.savetxt('X.csv', X, delimiter=',')
W_est = notears_linear(X, lambda1=0.1, loss_type='l2')
assert utils.is_dag(W_est)
np.savetxt('W_est.csv', W_est, delimiter=',')
acc = utils.count_accuracy(B_true, W_est != 0)
print(acc)
