def evaluate(Y,
Yhat,
S2=None,
mY=None,
sY=None,
metrics=['Rho', 'RMSE', 'SMSE', 'EXPV', 'MSLL']):
feature_num = Y.shape[1]
# find and remove bad variables from the response variables
nz = np.where(
np.bitwise_and(np.isfinite(Y).any(axis=0),
np.var(Y, axis=0) != 0))[0]
MSE = np.mean((Y - Yhat)**2, axis=0)
results = dict()
if 'RMSE' in metrics:
RMSE = np.sqrt(MSE)
results['RMSE'] = RMSE
if 'Rho' in metrics:
Rho = np.zeros(feature_num)
pRho = np.ones(feature_num)
Rho[nz], pRho[nz] = compute_pearsonr(Y[:, nz], Yhat[:, nz])
results['Rho'] = Rho
results['pRho'] = pRho
if 'SMSE' in metrics:
SMSE = np.zeros_like(MSE)
SMSE[nz] = MSE[nz] / np.var(Y[:, nz], axis=0)
results['SMSE'] = SMSE
if 'EXPV' in metrics:
EXPV = np.zeros(feature_num)
EXPV[nz] = explained_var(Y[:, nz], Yhat[:, nz])
results['EXPV'] = EXPV
if 'MSLL' in metrics:
if ((S2 is not None) and (mY is not None) and (sY is not None)):
MSLL = np.zeros(feature_num)
MSLL[nz] = compute_MSLL(Y[:, nz], Yhat[:, nz], S2[:, nz],
mY[nz].reshape(-1, 1).T,
(sY[nz]**2).reshape(-1, 1).T)
results['MSLL'] = MSLL
return results
def estimate(args):
torch.set_default_dtype(torch.float32)
args.type = 'MT'
print('Loading the input Data ...')
responses = fileio.load_nifti(args.respfile,
vol=True).transpose([3, 0, 1, 2])
response_shape = responses.shape
with open(args.covfile, 'rb') as handle:
covariates = pickle.load(handle)['covariates']
with open(args.testcovfile, 'rb') as handle:
test_covariates = pickle.load(handle)['test_covariates']
if args.mask is not None:
mask = fileio.load_nifti(args.mask, vol=True)
mask = fileio.create_mask(mask, mask=None)
else:
mask = fileio.create_mask(responses[0, :, :, :], mask=None)
if args.testrespfile is not None:
test_responses = fileio.load_nifti(args.testrespfile,
vol=True).transpose([3, 0, 1, 2])
test_responses_shape = test_responses.shape
print('Normalizing the input Data ...')
covariates_scaler = StandardScaler()
covariates = covariates_scaler.fit_transform(covariates)
test_covariates = covariates_scaler.transform(test_covariates)
response_scaler = MinMaxScaler()
responses = unravel_2D(response_scaler.fit_transform(ravel_2D(responses)),
response_shape)
if args.testrespfile is not None:
test_responses = unravel_2D(
response_scaler.transform(ravel_2D(test_responses)),
test_responses_shape)
test_responses = np.expand_dims(test_responses, axis=1)
factor = args.m
x_context = np.zeros([covariates.shape[0], factor, covariates.shape[1]],
dtype=np.float32)
y_context = np.zeros([
responses.shape[0], factor, responses.shape[1], responses.shape[2],
responses.shape[3]
],
dtype=np.float32)
x_all = np.zeros([covariates.shape[0], factor, covariates.shape[1]],
dtype=np.float32)
x_context_test = np.zeros(
[test_covariates.shape[0], factor, test_covariates.shape[1]],
dtype=np.float32)
y_context_test = np.zeros([
test_covariates.shape[0], factor, responses.shape[1],
responses.shape[2], responses.shape[3]
],
dtype=np.float32)
print('Estimating the fixed-effects ...')
for i in range(factor):
x_context[:, i, :] = covariates[:, :]
x_context_test[:, i, :] = test_covariates[:, :]
idx = np.random.randint(0, covariates.shape[0], covariates.shape[0])
if args.estimator == 'ST':
for j in range(responses.shape[1]):
for k in range(responses.shape[2]):
for l in range(responses.shape[3]):
reg = LinearRegression()
reg.fit(x_context[idx, i, :], responses[idx, j, k, l])
y_context[:, i, j, k, l] = reg.predict(x_context[:,
i, :])
y_context_test[:, i, j, k,
l] = reg.predict(x_context_test[:,
i, :])
elif args.estimator == 'MT':
reg = MultiTaskLasso(alpha=0.1)
reg.fit(
x_context[idx, i, :],
np.reshape(responses[idx, :, :, :],
[covariates.shape[0],
np.prod(responses.shape[1:])]))
y_context[:, i, :, :, :] = np.reshape(
reg.predict(x_context[:, i, :]), [
x_context.shape[0], responses.shape[1], responses.shape[2],
responses.shape[3]
])
y_context_test[:, i, :, :, :] = np.reshape(
reg.predict(x_context_test[:, i, :]), [
x_context_test.shape[0], responses.shape[1],
responses.shape[2], responses.shape[3]
])
print('Fixed-effect %d of %d is computed!' % (i + 1, factor))
x_all = x_context
responses = np.expand_dims(responses, axis=1).repeat(factor, axis=1)
################################## TRAINING #################################
encoder = Encoder(x_context, y_context, args).to(args.device)
args.cnn_feature_num = encoder.cnn_feature_num
decoder = Decoder(x_context, y_context, args).to(args.device)
model = NP(encoder, decoder, args).to(args.device)
print('Estimating the Random-effect ...')
k = 1
epochs = [
int(args.epochs / 4),
int(args.epochs / 2),
int(args.epochs / 5),
int(args.epochs - args.epochs / 4 - args.epochs / 2 - args.epochs / 5)
]
mini_batch_num = args.batchnum
batch_size = int(x_context.shape[0] / mini_batch_num)
model.train()
for e in range(len(epochs)):
optimizer = optim.Adam(model.parameters(), lr=10**(-e - 2))
for j in range(epochs[e]):
train_loss = 0
rand_idx = np.random.permutation(x_context.shape[0])
for i in range(mini_batch_num):
optimizer.zero_grad()
idx = rand_idx[i * batch_size:(i + 1) * batch_size]
y_hat, z_all, z_context, dummy = model(
torch.tensor(x_context[idx, :, :], device=args.device),
torch.tensor(y_context[idx, :, :, :, :],
device=args.device),
torch.tensor(x_all[idx, :, :], device=args.device),
torch.tensor(responses[idx, :, :, :, :],
device=args.device))
loss = np_loss(
y_hat,
torch.tensor(responses[idx, :, :, :, :],
device=args.device), z_all, z_context)
loss.backward()
train_loss += loss.item()
optimizer.step()
print('Epoch: %d, Loss:%f, Average Loss:%f' %
(k, train_loss, train_loss / responses.shape[0]))
k += 1
################################## Evaluation #################################
print('Predicting on Test Data ...')
model.eval()
model.apply(apply_dropout_test)
with torch.no_grad():
y_hat, z_all, z_context, y_sigma = model(
torch.tensor(x_context_test, device=args.device),
torch.tensor(y_context_test, device=args.device),
n=15)
if args.testrespfile is not None:
test_loss = np_loss(y_hat[0:test_responses_shape[0], :],
torch.tensor(test_responses, device=args.device),
z_all, z_context).item()
print('Average Test Loss:%f' % (test_loss / test_responses_shape[0]))
RMSE = np.sqrt(
np.mean((test_responses -
y_hat[0:test_responses_shape[0], :].cpu().numpy())**2,
axis=0)).squeeze() * mask
SMSE = RMSE**2 / np.var(test_responses, axis=0).squeeze()
Rho, pRho = compute_pearsonr(
test_responses.squeeze(),
y_hat[0:test_responses_shape[0], :].cpu().numpy().squeeze())
EXPV = explained_var(
test_responses.squeeze(),
y_hat[0:test_responses_shape[0], :].cpu().numpy().squeeze()) * mask
MSLL = compute_MSLL(
test_responses.squeeze(),
y_hat[0:test_responses_shape[0], :].cpu().numpy().squeeze(),
y_sigma[0:test_responses_shape[0], :].cpu().numpy().squeeze()**2,
train_mean=test_responses.mean(0),
train_var=test_responses.var(0)).squeeze() * mask
NPMs = (test_responses -
y_hat[0:test_responses_shape[0], :].cpu().numpy()) / (
y_sigma[0:test_responses_shape[0], :].cpu().numpy())
NPMs = NPMs.squeeze()
NPMs = NPMs * mask
NPMs = np.nan_to_num(NPMs)
temp = NPMs.reshape(
[NPMs.shape[0], NPMs.shape[1] * NPMs.shape[2] * NPMs.shape[3]])
EVD_params = extreme_value_prob_fit(temp, 0.01)
abnormal_probs = extreme_value_prob(EVD_params, temp, 0.01)
############################## SAVING RESULTS #################################
print('Saving Results to: %s' % (args.outdir))
exfile = args.respfile
y_hat = y_hat.squeeze().cpu().numpy()
y_hat = response_scaler.inverse_transform(ravel_2D(y_hat))
y_hat = y_hat[:, mask.flatten()]
fileio.save(y_hat.T,
args.outdir + '/yhat.nii.gz',
example=exfile,
mask=mask)
ys2 = y_sigma.squeeze().cpu().numpy()
ys2 = ravel_2D(ys2) * (response_scaler.data_max_ -
response_scaler.data_min_)
ys2 = ys2**2
ys2 = ys2[:, mask.flatten()]
fileio.save(ys2.T, args.outdir + '/ys2.nii.gz', example=exfile, mask=mask)
if args.testrespfile is not None:
NPMs = ravel_2D(NPMs)[:, mask.flatten()]
fileio.save(NPMs.T,
args.outdir + '/Z.nii.gz',
example=exfile,
mask=mask)
fileio.save(Rho.flatten()[mask.flatten()],
args.outdir + '/Rho.nii.gz',
example=exfile,
mask=mask)
fileio.save(pRho.flatten()[mask.flatten()],
args.outdir + '/pRho.nii.gz',
example=exfile,
mask=mask)
fileio.save(RMSE.flatten()[mask.flatten()],
args.outdir + '/rmse.nii.gz',
example=exfile,
mask=mask)
fileio.save(SMSE.flatten()[mask.flatten()],
args.outdir + '/smse.nii.gz',
example=exfile,
mask=mask)
fileio.save(EXPV.flatten()[mask.flatten()],
args.outdir + '/expv.nii.gz',
example=exfile,
mask=mask)
fileio.save(MSLL.flatten()[mask.flatten()],
args.outdir + '/msll.nii.gz',
example=exfile,
mask=mask)
with open(args.outdir + 'model.pkl', 'wb') as handle:
pickle.dump(
{
'model': model,
'covariates_scaler': covariates_scaler,
'response_scaler': response_scaler,
'EVD_params': EVD_params,
'abnormal_probs': abnormal_probs
},
handle,
protocol=pickle.HIGHEST_PROTOCOL)
###############################################################################
print('DONE!')
def estimate(respfile,
covfile,
maskfile=None,
cvfolds=None,
testcov=None,
testresp=None,
alg='gpr',
configparam=None,
saveoutput=True,
outputsuffix=None,
standardize=True):
""" Estimate a normative model
This will estimate a model in one of two settings according to the
particular parameters specified (see below):
* under k-fold cross-validation
required settings 1) respfile 2) covfile 3) cvfolds>2
* estimating a training dataset then applying to a second test dataset
required sessting 1) respfile 2) covfile 3) testcov 4) testresp
* estimating on a training dataset ouput of forward maps mean and se
required sessting 1) respfile 2) covfile 3) testcov
The models are estimated on the basis of data stored on disk in ascii or
neuroimaging data formats (nifti or cifti). Ascii data should be in
tab or space delimited format with the number of subjects in rows and the
number of variables in columns. Neuroimaging data will be reshaped
into the appropriate format
Basic usage::
estimate(respfile, covfile, [extra_arguments])
where the variables are defined below. Note that either the cfolds
parameter or (testcov, testresp) should be specified, but not both.
:param respfile: response variables for the normative model
:param covfile: covariates used to predict the response variable
:param maskfile: mask used to apply to the data (nifti only)
:param cvfolds: Number of cross-validation folds
:param testcov: Test covariates
:param testresp: Test responses
:param alg: Algorithm for normative model
:param configparam: Parameters controlling the estimation algorithm
:param saveoutput: Save the output to disk? Otherwise returned as arrays
:param outputsuffix: Text string to add to the output filenames
All outputs are written to disk in the same format as the input. These are:
:outputs: * yhat - predictive mean
* ys2 - predictive variance
* Hyp - hyperparameters
* Z - deviance scores
* Rho - Pearson correlation between true and predicted responses
* pRho - parametric p-value for this correlation
* rmse - root mean squared error between true/predicted responses
* smse - standardised mean squared error
The outputsuffix may be useful to estimate multiple normative models in the
same directory (e.g. for custom cross-validation schemes)
"""
# load data
print("Processing data in " + respfile)
X = fileio.load(covfile)
Y, maskvol = load_response_vars(respfile, maskfile)
if len(Y.shape) == 1:
Y = Y[:, np.newaxis]
if len(X.shape) == 1:
X = X[:, np.newaxis]
Nmod = Y.shape[1]
if testcov is not None:
# we have a separate test dataset
Xte = fileio.load(testcov)
testids = range(X.shape[0], X.shape[0] + Xte.shape[0])
if len(Xte.shape) == 1:
Xte = Xte[:, np.newaxis]
if testresp is not None:
Yte, testmask = load_response_vars(testresp, maskfile)
if len(Yte.shape) == 1:
Yte = Yte[:, np.newaxis]
else:
sub_te = Xte.shape[0]
Yte = np.zeros([sub_te, Nmod])
# Initialise normative model
nm = norm_init(X, alg=alg, configparam=configparam)
# treat as a single train-test split
splits = CustomCV((range(0, X.shape[0]), ), (testids, ))
Y = np.concatenate((Y, Yte), axis=0)
X = np.concatenate((X, Xte), axis=0)
# force the number of cross-validation folds to 1
if cvfolds is not None and cvfolds != 1:
print("Ignoring cross-valdation specification (test data given)")
cvfolds = 1
else:
# we are running under cross-validation
splits = KFold(n_splits=cvfolds)
testids = range(0, X.shape[0])
# Initialise normative model
nm = norm_init(X, alg=alg, configparam=configparam)
# find and remove bad variables from the response variables
# note: the covariates are assumed to have already been checked
nz = np.where(
np.bitwise_and(np.isfinite(Y).any(axis=0),
np.var(Y, axis=0) != 0))[0]
# run cross-validation loop
Yhat = np.zeros_like(Y)
S2 = np.zeros_like(Y)
Hyp = np.zeros((Nmod, nm.n_params, cvfolds))
Z = np.zeros_like(Y)
nlZ = np.zeros((Nmod, cvfolds))
for idx in enumerate(splits.split(X)):
fold = idx[0]
tr = idx[1][0]
te = idx[1][1]
# standardize responses and covariates, ignoring invalid entries
iy, jy = np.ix_(tr, nz)
mY = np.mean(Y[iy, jy], axis=0)
sY = np.std(Y[iy, jy], axis=0)
if standardize:
Yz = np.zeros_like(Y)
Yz[:, nz] = (Y[:, nz] - mY) / sY
mX = np.mean(X[tr, :], axis=0)
sX = np.std(X[tr, :], axis=0)
Xz = (X - mX) / sX
else:
Yz = Y
Xz = X
# estimate the models for all subjects
for i in range(0, len(nz)): # range(0, Nmod):
print("Estimating model ", i + 1, "of", len(nz))
try:
nm = norm_init(Xz[tr, :],
Yz[tr, nz[i]],
alg=alg,
configparam=configparam)
Hyp[nz[i], :, fold] = nm.estimate(Xz[tr, :], Yz[tr, nz[i]])
if (alg == 'hbr'):
if nm.configs['new_site'] == True:
nm.estimate_on_new_sites(
Xz[te, :], Y[te, nz[i]]
) # The test/train division is done internally
yhat, s2 = nm.predict_on_new_sites(Xz[te, :])
else:
yhat, s2 = nm.predict(Xz[te, :], Xz[tr, :],
Yz[tr, nz[i]], Hyp[nz[i], :,
fold])
else:
yhat, s2 = nm.predict(Xz[te, :], Xz[tr, :], Yz[tr, nz[i]],
Hyp[nz[i], :, fold])
if standardize:
Yhat[te, nz[i]] = yhat * sY[i] + mY[i]
S2[te, nz[i]] = s2 * sY[i]**2
else:
Yhat[te, nz[i]] = yhat
S2[te, nz[i]] = s2
nlZ[nz[i], fold] = nm.neg_log_lik
if testcov is None:
Z[te, nz[i]] = (Y[te, nz[i]] - Yhat[te, nz[i]]) / \
np.sqrt(S2[te, nz[i]])
else:
if testresp is not None:
Z[te, nz[i]] = (Y[te, nz[i]] - Yhat[te, nz[i]]) / \
np.sqrt(S2[te, nz[i]])
except Exception as e:
exc_type, exc_obj, exc_tb = sys.exc_info()
fname = os.path.split(exc_tb.tb_frame.f_code.co_filename)[1]
print("Model ", i + 1, "of", len(nz),
"FAILED!..skipping and writing NaN to outputs")
print("Exception:")
print(e)
print(exc_type, fname, exc_tb.tb_lineno)
Hyp[nz[i], :, fold] = float('nan')
Yhat[te, nz[i]] = float('nan')
S2[te, nz[i]] = float('nan')
nlZ[nz[i], fold] = float('nan')
if testcov is None:
Z[te, nz[i]] = float('nan')
else:
if testresp is not None:
Z[te, nz[i]] = float('nan')
# compute performance metrics
if testcov is None:
MSE = np.mean((Y[testids, :] - Yhat[testids, :])**2, axis=0)
RMSE = np.sqrt(MSE)
# for the remaining variables, we need to ignore zero variances
SMSE = np.zeros_like(MSE)
Rho = np.zeros(Nmod)
pRho = np.ones(Nmod)
EXPV = np.zeros(Nmod)
MSLL = np.zeros(Nmod)
iy, jy = np.ix_(testids, nz) # ids for tested samples nonzero values
SMSE[nz] = MSE[nz] / np.var(Y[iy, jy], axis=0)
Rho[nz], pRho[nz] = compute_pearsonr(Y[iy, jy], Yhat[iy, jy])
EXPV[nz] = explained_var(Y[iy, jy], Yhat[iy, jy])
MSLL[nz] = compute_MSLL(Y[iy, jy], Yhat[iy, jy], S2[iy, jy],
mY.reshape(-1, 1).T, (sY**2).reshape(-1, 1).T)
else:
if testresp is not None:
MSE = np.mean((Y[testids, :] - Yhat[testids, :])**2, axis=0)
RMSE = np.sqrt(MSE)
# for the remaining variables, we need to ignore zero variances
SMSE = np.zeros_like(MSE)
Rho = np.zeros(Nmod)
pRho = np.ones(Nmod)
EXPV = np.zeros(Nmod)
MSLL = np.zeros(Nmod)
iy, jy = np.ix_(testids, nz) # ids tested samples nonzero values
SMSE[nz] = MSE[nz] / np.var(Y[iy, jy], axis=0)
Rho[nz], pRho[nz] = compute_pearsonr(Y[iy, jy], Yhat[iy, jy])
EXPV[nz] = explained_var(Y[iy, jy], Yhat[iy, jy])
MSLL[nz] = compute_MSLL(Y[iy, jy], Yhat[iy, jy], S2[iy, jy],
mY.reshape(-1, 1).T,
(sY**2).reshape(-1, 1).T)
# Set writing options
if saveoutput:
print("Writing output ...")
if fileio.file_type(respfile) == 'cifti' or \
fileio.file_type(respfile) == 'nifti':
exfile = respfile
else:
exfile = None
if outputsuffix is not None:
ext = str(outputsuffix) + fileio.file_extension(respfile)
else:
ext = fileio.file_extension(respfile)
# Write output
if testcov is None:
fileio.save(Yhat[testids, :],
'yhat' + ext,
example=exfile,
mask=maskvol)
fileio.save(S2[testids, :],
'ys2' + ext,
example=exfile,
mask=maskvol)
fileio.save(Z[testids, :], 'Z' + ext, example=exfile, mask=maskvol)
fileio.save(Rho, 'Rho' + ext, example=exfile, mask=maskvol)
fileio.save(pRho, 'pRho' + ext, example=exfile, mask=maskvol)
fileio.save(RMSE, 'rmse' + ext, example=exfile, mask=maskvol)
fileio.save(SMSE, 'smse' + ext, example=exfile, mask=maskvol)
fileio.save(EXPV, 'expv' + ext, example=exfile, mask=maskvol)
fileio.save(MSLL, 'msll' + ext, example=exfile, mask=maskvol)
if cvfolds is None:
fileio.save(Hyp[:, :, 0],
'Hyp' + ext,
example=exfile,
mask=maskvol)
else:
for idx in enumerate(splits.split(X)):
fold = idx[0]
fileio.save(Hyp[:, :, fold].T,
'Hyp_' + str(fold + 1) + ext,
example=exfile,
mask=maskvol)
else:
if testresp is None:
fileio.save(Yhat[testids, :],
'yhat' + ext,
example=exfile,
mask=maskvol)
fileio.save(S2[testids, :],
'ys2' + ext,
example=exfile,
mask=maskvol)
fileio.save(Hyp[:, :, 0],
'Hyp' + ext,
example=exfile,
mask=maskvol)
else:
fileio.save(Yhat[testids, :],
'yhat' + ext,
example=exfile,
mask=maskvol)
fileio.save(S2[testids, :],
'ys2' + ext,
example=exfile,
mask=maskvol)
fileio.save(Z[testids, :],
'Z' + ext,
example=exfile,
mask=maskvol)
fileio.save(Rho, 'Rho' + ext, example=exfile, mask=maskvol)
fileio.save(pRho, 'pRho' + ext, example=exfile, mask=maskvol)
fileio.save(RMSE, 'rmse' + ext, example=exfile, mask=maskvol)
fileio.save(SMSE, 'smse' + ext, example=exfile, mask=maskvol)
fileio.save(EXPV, 'expv' + ext, example=exfile, mask=maskvol)
fileio.save(MSLL, 'msll' + ext, example=exfile, mask=maskvol)
if cvfolds is None:
fileio.save(Hyp[:, :, 0].T,
'Hyp' + ext,
example=exfile,
mask=maskvol)
else:
for idx in enumerate(splits.split(X)):
fold = idx[0]
fileio.save(Hyp[:, :, fold].T,
'Hyp_' + str(fold + 1) + ext,
example=exfile,
mask=maskvol)
else:
if testcov is None:
output = (Yhat[testids, :], S2[testids, :], Hyp, Z[testids, :],
Rho, pRho, RMSE, SMSE, EXPV, MSLL)
else:
if testresp is None:
output = (Yhat[testids, :], S2[testids, :], Hyp)
else:
output = (Yhat[testids, :], S2[testids, :], Hyp, Z[testids, :],
Rho, pRho, RMSE, SMSE, EXPV, MSLL)
return output