def main():
[[X_train, y_train], [X_valid, y_valid],
[X_test, y_test]] = util.load_data(DATA_PATH)
X = X_train.append(X_valid).append(X_test)
y = y_train.append(y_valid).append(y_test)
del X_train
del X_valid
del X_test
del y_train
del y_valid
del y_test
X = make_squared(X, sq_fields)
X = make_interactions(X, interactions)
full = X.copy()
X = X.drop([
'studyArea', 'x', 'y', 'elev_srtm30', 'year',
'varPrecip_growingSeason', 'precip_OctSep:varPrecip_growingSeason'
],
axis=1)
predictors = list(X)
scaler = StandardScaler()
X = scaler.fit_transform(X)
y = y['beetle'].values.reshape(-1)
logistic_clf = LogisticRegression(C=0.001, penalty='l2')
logistic_clf.fit(X, y)
probs = logistic_clf.predict_proba(X)
probs = [p[1] for p in probs]
X_df = pd.DataFrame(data=X, index=full.index, columns=predictors)
X_df['year'] = full['year']
X_df['x'] = full['x']
X_df['y'] = full['y']
X_df['probs'] = probs
X_df['preds'] = X_df['probs'].apply(lambda x: 1
if x >= OPTIMAL_THRESHOLD else 0)
out_data = X_df.loc[X_df.year == 2000, ['x', 'y', 'probs', 'preds']]
out_data = out_data.rename(columns={
'probs': 'probs_2000',
'preds': 'preds_2000'
})
out_data.index = out_data.apply(lambda row: str(row['x']) + str(row['y']),
axis=1)
for year in range(2001, 2015):
year_data = X_df.loc[X_df.year == year, ['x', 'y', 'probs', 'preds']]
year_data.index = year_data.apply(
lambda row: str(row['x']) + str(row['y']), axis=1)
out_data['probs_%s' % year] = year_data['probs']
out_data['preds_%s' % year] = year_data['preds']
out_data.index = range(out_data.shape[0])
print(out_data.head())
out_data.to_csv(HISTORIC_DATA_PATH + 'recent_data_fitted_no_beetle.csv')
def main():
ignore = [
'year', 'studyArea', 'elev_srtm30', 'x', 'y', 'varPrecip_growingSeason'
]
if HISTORIC:
X_test = pd.read_csv(DATA_DIR + 'X_test.csv')
x_min = X_test.x.min()
y_min = X_test.y.min()
x_max = X_test.x.max()
y_max = X_test.y.max()
print('x range: %d - %d' % (x_min, x_max))
print('y range: %d - %d' % (y_min, y_max))
t0 = time.time()
for i in HISTORIC_YEARS:
t_iter = time.time()
path = '%sclean_%d.csv' % (HISTORIC_DATA_DIR, i)
print('Reading data from %s...' % path)
X = pd.read_csv(path)
print('Filtering x, y ranges...')
X = X.loc[((X.x >= x_min)
& (X.x <= x_max)
& (X.y >= y_min)
& (X.y <= y_max)), :]
fields = [col for col in list(X) if col not in ignore]
make_and_save_tensor(X, fields, i)
iter_time = (time.time() - t_iter) / 60
elapsed = (time.time() - t0) / 60
print(
' Iteration time: %.2fminutes\n Elapsed time: %.2fminutes' %
(iter_time, elapsed))
else:
[[X_train, y_train], [X_valid, y_valid],
[X_test, y_test]] = util.load_data(DATA_DIR)
X_train, y_train = util.drop_nans(X_train, y_train,
'varPrecip_growingSeason')
X_valid, y_valid = util.drop_nans(X_valid, y_valid,
'varPrecip_growingSeason')
X_test, y_test = util.drop_nans(X_test, y_test,
'varPrecip_growingSeason')
fields = [col for col in list(X_test) if col not in ignore]
for i in range(2006, 2015):
make_and_save_tensor(X_train, fields, i)
make_and_save_y_matrix(y_train, X_train, i)
for i in range(2003, 2006):
make_and_save_tensor(X_valid, fields, i)
make_and_save_y_matrix(y_valid, X_valid, i)
for i in range(2000, 2003):
make_and_save_tensor(X_test, fields, i)
make_and_save_y_matrix(y_test, X_test, i)
def main():
print('Loading data...')
[[X_train, y_train], [X_valid, y_valid],
[X_test, y_test]] = util.load_data(DATA_DIR)
print('Merging data....')
data = X_train.append(X_valid).append(X_test)
y = y_train.append(y_valid).append(y_test)
data['beetle'] = y
print('Adding the following year beetle data...')
[[X_train, y_train], [X_valid, y_valid],
[X_test, y_test]] = make_new_data_sets(data)
print('X_train: %s\ty_train: %s\n'
'X_valid: %s\ty_valid: %s\n'
'X_test: %s\ty_test: %s' %
(X_train.shape, y_train.shape, X_valid.shape, y_valid.shape,
X_test.shape, y_test.shape))
print('Saving files as "X_train_full.csv" etc....')
save_xy([X_train, y_train], DATA_DIR, 'train_full')
save_xy([X_valid, y_valid], DATA_DIR, 'valid_full')
save_xy([X_test, y_test], DATA_DIR, 'test_full')
def predict_on_image(model, image_path, intensity_correction=0.0):
"""
Use the model to give a prediction on a image.
:param model: A keras model.
:param image_path: The path to a image in geotiff format. Image should be 512x512 and in black and white.
:return: The prediction image as a TrainingImage object.
"""
# Load image
training_image = model_utils.load_data(image_path, image_path)
data_X = model_utils.convert_training_images_to_numpy_arrays([training_image])[0]
data_X += intensity_correction / (2**8 - 1) # Adjust for differing light levels in training and this dataset
data_X = model_utils.fake_colors(data_X)
prediction = model.predict(data_X)
prediction = np.argmax(prediction, axis=-1)
prediction = np.squeeze(prediction)
training_image.labels = prediction
return training_image
def run_nn_models(sp,
n_folds,
combined_sbjs,
lp,
roi_proj_loadpath,
pats_ids_in=[
'EC01', 'EC02', 'EC03', 'EC04', 'EC05', 'EC06', 'EC07',
'EC08', 'EC09', 'EC10', 'EC11', 'EC12'
],
n_evs_per_sbj=500,
test_day=None,
tlim=[-1, 1],
n_chans_all=140,
dipole_dens_thresh=0.2,
rem_bad_chans=True,
models=['eegnet_hilb', 'eegnet', 'rf'],
save_suffix='',
n_estimators=150,
max_depth=8,
overwrite=True,
dropoutRate=0.25,
kernLength=32,
F1=8,
D=2,
F2=16,
dropoutType='Dropout',
kernLength_sep=16,
rand_seed=1337,
loss='categorical_crossentropy',
optimizer='adam',
patience=5,
early_stop_monitor='val_loss',
do_log=False,
n_test=1,
n_val=4,
custom_rois=True,
n_train=7,
epochs=20,
compute_val='power',
ecog_srate=500,
half_n_evs_test='nopad',
trim_n_chans=True):
'''
Main function that prepares data and aggregates accuracy values from model fitting.
Note that overwrite variable no longer does anything.
Also note that ecog_srate is only needed for frequency sliding computation in neural net (if compute_val=='freqslide')
'''
# Ensure pats_ids_in and models variables are lists
if not isinstance(pats_ids_in, list):
pats_ids_in = [pats_ids_in]
if not isinstance(models, list):
models = [models]
# Save pickle file with dictionary of input parameters (useful for reproducible dataset splits and model fitting)
params_dict = {
'sp': sp,
'n_folds': n_folds,
'combined_sbjs': combined_sbjs,
'lp': lp,
'pats_ids_in': pats_ids_in,
'n_evs_per_sbj': n_evs_per_sbj,
'test_day': test_day,
'tlim': tlim,
'n_chans_all': n_chans_all,
'dipole_dens_thresh': dipole_dens_thresh,
'rem_bad_chans': rem_bad_chans,
'models': models,
'save_suffix': save_suffix,
'n_estimators': n_estimators,
'max_depth': max_depth,
'overwrite': overwrite,
'dropoutRate': dropoutRate,
'kernLength': kernLength,
'F1': F1,
'D': D,
'F2': F2,
'dropoutType': dropoutType,
'kernLength_sep': kernLength_sep,
'rand_seed': rand_seed,
'loss': loss,
'optimizer': optimizer,
'patience': patience,
'early_stop_monitor': early_stop_monitor,
'do_log': do_log,
'n_test': n_test,
'n_val': n_val,
'n_train': n_train,
'epochs': epochs,
'compute_val': compute_val,
'ecog_srate': ecog_srate,
'trim_n_chans': trim_n_chans
}
f = open(sp + 'param_file.pkl', 'wb')
pickle.dump(params_dict, f)
f.close()
# Set random seed
np.random.seed(rand_seed)
# Perform different procedures depending on whether or not multiple subjects are being fit together
if combined_sbjs:
# For multi-subject fits, obtain projection matrix and good regions of interest for each subject
if custom_rois:
custom_roi_inds = get_custom_motor_rois(
) # load custom roi's from precentral, postcentral, and inf parietal (AAL2)
else:
custom_roi_inds = None
print("Determining ROIs")
proj_mat_out, good_ROIs, chan_ind_vals_all = proj_mats_good_rois(
pats_ids_in,
n_chans_all=n_chans_all,
rem_bad_chans=rem_bad_chans,
dipole_dens_thresh=dipole_dens_thresh,
custom_roi_inds=custom_roi_inds,
chan_cut_thres=n_chans_all,
roi_proj_loadpath=roi_proj_loadpath)
nROIs = len(good_ROIs)
print("ROIs found")
# Retain only the electrodes with nonzero data (initially padded because number of electrodes varies across subjects)
# proj_mat_out : (len(pats_ids_in) x len(good_ROIs) x n_chans_all)
if trim_n_chans:
n_chans_all = len(
np.nonzero(
proj_mat_out.reshape(
-1, proj_mat_out.shape[-1]).mean(axis=0))[0])
proj_mat_out = proj_mat_out[..., :n_chans_all]
np.save(sp + "proj_mat_out", proj_mat_out)
# Load ECoG data (if test_day is None, then X_test_orig, y_test_orig, and sbj_order_test_load will be empty)
X, y, X_test_orig, y_test_orig, sbj_order, sbj_order_test_load = load_data(
pats_ids_in,
lp,
n_chans_all=n_chans_all,
test_day=test_day,
tlim=tlim)
X[np.isnan(X)] = 0 # set all NaN's to 0
# Identify the number of unique labels (or classes) present
nb_classes = len(np.unique(y))
# Choose which subjects for training/validation/testing for every fold (splits are based on random seed)
sbj_inds_all_train, sbj_inds_all_val, sbj_inds_all_test = folds_choose_subjects(
n_folds, pats_ids_in, n_test=n_test, n_val=n_val, n_train=n_train)
# Iterate across all model types specified
labels_unique = np.unique(y)
if isinstance(n_evs_per_sbj, str):
half_n_evs = n_evs_per_sbj
else:
half_n_evs = n_evs_per_sbj // len(labels_unique)
# half_n_evs_test = 'nopad' #avoid duplicating events for test set (okay for train/val sets where it is more important to balance trials across subjects)
train_inds_folds, val_inds_folds, test_inds_folds = [], [], []
for k, modeltype in enumerate(models):
accs = np.zeros([n_folds, 3]) # accuracy table for all NN models
last_epochs = np.zeros([n_folds, 2])
# For the number of folds, pick the events to use
for i in tqdm(range(n_folds)):
test_sbj = sbj_inds_all_test[i]
val_sbj = sbj_inds_all_val[i]
train_sbj = sbj_inds_all_train[i]
# Only need to determine train/val/test inds for first modeltype used
if k == 0:
# Find train/val/test indices (test inds differ depending on if test_day is specified or not)
# Note that subject_data_inds will balance number of trials across classes
train_inds, val_inds, test_inds = [], [], []
if test_day is None:
test_inds = subject_data_inds(np.full(1, test_sbj),
sbj_order, labels_unique,
i, 'test_inds',
half_n_evs_test, y, sp,
n_folds, test_inds,
overwrite)
else:
test_inds = subject_data_inds(
np.full(1, test_sbj), sbj_order_test_load,
labels_unique, i, 'test_inds', half_n_evs_test,
y_test_orig, sp, n_folds, test_inds, overwrite)
val_inds = subject_data_inds(val_sbj, sbj_order,
labels_unique, i, 'val_inds',
half_n_evs, y, sp, n_folds,
val_inds, overwrite)
train_inds = subject_data_inds(train_sbj, sbj_order,
labels_unique, i,
'train_inds', half_n_evs, y,
sp, n_folds, train_inds,
overwrite)
train_inds_folds.append(train_inds)
val_inds_folds.append(val_inds)
test_inds_folds.append(test_inds)
else:
train_inds = train_inds_folds[i]
val_inds = val_inds_folds[i]
test_inds = test_inds_folds[i]
# Now that we have the train/val/test event indices, generate the data for the models
X_train = X[train_inds, ...]
Y_train = y[train_inds]
sbj_order_train = sbj_order[train_inds]
X_validate = X[val_inds, ...]
Y_validate = y[val_inds]
sbj_order_validate = sbj_order[val_inds]
if test_day is None:
X_test = X[test_inds, ...]
Y_test = y[test_inds]
sbj_order_test = sbj_order[test_inds]
else:
X_test = X_test_orig[test_inds, ...]
Y_test = y_test_orig[test_inds]
sbj_order_test = sbj_order_test_load[test_inds]
if modeltype == 'rf':
# For random forest, project data from electrodes to ROIs in advance
X_train_proj = roi_proj_rf(X_train, sbj_order_train, nROIs,
proj_mat_out)
X_validate_proj = roi_proj_rf(X_validate,
sbj_order_validate, nROIs,
proj_mat_out)
X_test_proj = roi_proj_rf(X_test, sbj_order_test, nROIs,
proj_mat_out)
# Create Random Forest classifier model
model = RandomForestClassifier(n_estimators=n_estimators,
max_depth=max_depth,
class_weight="balanced",
random_state=rand_seed,
n_jobs=1,
oob_score=True)
# Fit model and store train/val/test accuracies
t_fit_start = time.time()
clf = model.fit(X_train_proj, Y_train.ravel())
last_epochs[i, 1] = time.time() - t_fit_start
accs[i, 0] = accuracy_score(Y_train.ravel(),
clf.predict(X_train_proj))
accs[i, 1] = accuracy_score(Y_validate.ravel(),
clf.predict(X_validate_proj))
accs[i, 2] = accuracy_score(Y_test.ravel(),
clf.predict(X_test_proj))
del X_train_proj, X_validate_proj, X_test_proj
# Save model
chckpt_path = sp + modeltype + '_fold' + str(
i) + save_suffix + '.sav'
pickle.dump(clf, open(chckpt_path, 'wb'))
elif modeltype == 'riemann':
# Project data from electrodes to ROIs in advance
X_train_proj = roi_proj_rf(X_train, sbj_order_train, nROIs,
proj_mat_out)
X_validate_proj = roi_proj_rf(X_validate,
sbj_order_validate, nROIs,
proj_mat_out)
X_test_proj = roi_proj_rf(X_test, sbj_order_test, nROIs,
proj_mat_out)
# Reshape into 3 dimensions
X_train_proj2 = X_train_proj.reshape(
(X_train.shape[0], -1, X_train.shape[-1]))
X_validate_proj2 = X_validate_proj.reshape(
(X_validate.shape[0], -1, X_validate.shape[-1]))
X_test_proj2 = X_test_proj.reshape(
(X_test.shape[0], -1, X_test.shape[-1]))
# Find any events where std is 0
train_inds_bad = np.nonzero(
X_train_proj2.std(axis=-1).max(axis=-1) == 0)[0]
val_inds_bad = np.nonzero(
X_validate_proj2.std(axis=-1).max(axis=-1) == 0)[0]
test_inds_bad = np.nonzero(
X_test_proj2.std(axis=-1).max(axis=-1) == 0)[0]
if not not train_inds_bad.tolist():
first_good_ind = np.setdiff1d(
np.arange(X_train_proj2.shape[0]),
train_inds_bad)[0]
X_train_proj2[train_inds_bad,
...] = X_train_proj2[(train_inds_bad *
0) +
first_good_ind, ...]
if not not val_inds_bad.tolist():
first_good_ind = np.setdiff1d(
np.arange(X_validate_proj2.shape[0]),
val_inds_bad)[0]
X_validate_proj2[val_inds_bad, ...] = X_validate_proj2[
(val_inds_bad * 0) + first_good_ind, ...]
if not not test_inds_bad.tolist():
first_good_ind = np.setdiff1d(
np.arange(X_test_proj2.shape[0]), test_inds_bad)[0]
X_test_proj2[test_inds_bad,
...] = X_test_proj2[(test_inds_bad * 0) +
first_good_ind, ...]
# Estimate covariances matrices
cov_data_train = pyriemann.estimation.Covariances(
'lwf').fit_transform(X_train_proj2)
cov_data_val = pyriemann.estimation.Covariances(
'lwf').fit_transform(X_validate_proj2)
cov_data_test = pyriemann.estimation.Covariances(
'lwf').fit_transform(X_test_proj2)
# Create MDM model
mdm = pyriemann.classification.MDM()
# Fit model and store train/val/test accuracies
t_fit_start = time.time()
clf = mdm.fit(cov_data_train, Y_train.ravel())
last_epochs[i, 1] = time.time() - t_fit_start
accs[i, 0] = accuracy_score(Y_train.ravel(),
clf.predict(cov_data_train))
accs[i, 1] = accuracy_score(Y_validate.ravel(),
clf.predict(cov_data_val))
accs[i, 2] = accuracy_score(Y_test.ravel(),
clf.predict(cov_data_test))
del X_train_proj, X_validate_proj, X_test_proj
# Save model
chckpt_path = sp + modeltype + '_fold' + str(
i) + save_suffix + '.sav'
pickle.dump(clf, open(chckpt_path, 'wb'))
else:
# Reformat data size for NN fitting
Y_train = np_utils.to_categorical(Y_train - 1)
X_train = np.expand_dims(X_train, 1)
Y_validate = np_utils.to_categorical(Y_validate - 1)
X_validate = np.expand_dims(X_validate, 1)
Y_test = np_utils.to_categorical(Y_test - 1)
X_test = np.expand_dims(X_test, 1)
proj_mat_out2 = np.expand_dims(proj_mat_out, 1)
# Fit NN model using Keras
chckpt_path = sp + 'checkpoint_gen_' + modeltype + '_fold' + str(
i) + save_suffix + '.h5'
accs_lst, last_epoch_tmp = cnn_model(
X_train,
Y_train,
X_validate,
Y_validate,
X_test,
Y_test,
chckpt_path,
modeltype,
proj_mat_out2,
sbj_order_train,
sbj_order_validate,
sbj_order_test,
nROIs=nROIs,
nb_classes=nb_classes,
dropoutRate=dropoutRate,
kernLength=kernLength,
F1=F1,
D=D,
F2=F2,
dropoutType=dropoutType,
kernLength_sep=kernLength_sep,
loss=loss,
optimizer=optimizer,
patience=patience,
early_stop_monitor=early_stop_monitor,
do_log=do_log,
epochs=epochs,
compute_val=compute_val,
ecog_srate=ecog_srate)
# Store train/val/test accuracies, and last epoch
for ss in range(3):
accs[i, ss] = accs_lst[ss]
last_epochs[i, :] = last_epoch_tmp
# Save accuracies for all folds for one type of model
np.save(
sp + 'acc_gen_' + modeltype + '_' + str(n_folds) +
save_suffix + '.npy', accs)
np.save(
sp + 'last_training_epoch_gen_tf' + modeltype + '_' +
str(n_folds) + save_suffix + '.npy', last_epochs)
# Returns average validation accuracy for hyperparameter tuning (will be for last model_type only)
return accs[:, 1].mean()
else:
# Single subject model fitting
for pat_id_curr in pats_ids_in:
# Load ECoG data
X, y, X_test, y_test, sbj_order, sbj_order_test = load_data(
pat_id_curr,
lp,
n_chans_all=n_chans_all,
test_day=test_day,
tlim=tlim)
X[np.isnan(X)] = 0 # set all NaN's to 0
# Identify the number of unique labels (or classes) present
nb_classes = len(np.unique(y))
# Randomize event order (random seed facilitates consistency)
order_inds = np.arange(len(y))
np.random.shuffle(order_inds)
X = X[order_inds, ...]
y = y[order_inds]
order_inds_test = np.arange(len(y_test))
np.random.shuffle(order_inds_test)
X_test = X_test[order_inds_test, ...]
y_test = y_test[order_inds_test]
# Iterate across all model types specified
for modeltype in models:
# Reformat data based on model
if modeltype == 'rf':
y2 = y.copy()
y_test2 = y_test.copy()
X2 = X.copy()
X_test2 = X_test.copy()
elif modeltype == 'riemann':
y2 = y.copy()
y_test2 = y_test.copy()
X2 = X.copy()
X_test2 = X_test.copy()
else:
y2 = np_utils.to_categorical(y - 1)
y_test2 = np_utils.to_categorical(y_test - 1)
X2 = np.expand_dims(X, 1)
X_test2 = np.expand_dims(X_test, 1)
# Create splits for train/val and fit model
split_len = X2.shape[0] // n_folds
accs = np.zeros([n_folds, 3])
last_epochs = np.zeros([n_folds, 2])
for frodo in range(n_folds):
val_inds = np.arange(0, split_len) + (frodo * split_len)
train_inds = np.setdiff1d(
np.arange(X2.shape[0]),
val_inds) #take all events not in val set
# Split data and labels into train/val sets
X_train = X2[train_inds, ...]
Y_train = y2[train_inds]
X_validate = X2[val_inds, ...]
Y_validate = y2[val_inds]
if modeltype == 'rf':
# For random forest, combine electrodes and time dimensions
X_train_rf = X_train.reshape(X_train.shape[0], -1)
X_validate_rf = X_validate.reshape(
X_validate.shape[0], -1)
X_test2_rf = X_test2.reshape(X_test2.shape[0], -1)
# Create random forest model
model = RandomForestClassifier(
n_estimators=n_estimators,
max_depth=max_depth,
class_weight="balanced",
random_state=rand_seed,
n_jobs=1,
oob_score=True)
# Fit model and store accuracies
t_fit_start = time.time()
clf = model.fit(X_train_rf, Y_train.ravel())
last_epochs[frodo, 1] = time.time() - t_fit_start
accs[frodo,
0] = accuracy_score(Y_train.ravel(),
clf.predict(X_train_rf))
accs[frodo,
1] = accuracy_score(Y_validate.ravel(),
clf.predict(X_validate_rf))
accs[frodo,
2] = accuracy_score(y_test2.ravel(),
clf.predict(X_test2_rf))
# Save model
chckpt_path = sp+modeltype+'_'+pat_id_curr+'_testday_'+\
str(test_day)+'_fold'+str(frodo)+save_suffix+'.sav'
pickle.dump(clf, open(chckpt_path, 'wb'))
elif modeltype == 'riemann':
# Find any events where std is 0
train_inds_bad = np.nonzero(
X_train.std(axis=-1).max(axis=-1) == 0)[0]
val_inds_bad = np.nonzero(
X_validate.std(axis=-1).max(axis=-1) == 0)[0]
test_inds_bad = np.nonzero(
X_test2.std(axis=-1).max(axis=-1) == 0)[0]
if not not train_inds_bad.tolist():
first_good_ind = np.setdiff1d(
np.arange(X_train.shape[0]), train_inds_bad)[0]
X_train[train_inds_bad,
...] = X_train[(train_inds_bad * 0) +
first_good_ind, ...]
if not not val_inds_bad.tolist():
first_good_ind = np.setdiff1d(
np.arange(X_validate.shape[0]),
val_inds_bad)[0]
X_validate[val_inds_bad,
...] = X_validate[(val_inds_bad * 0) +
first_good_ind, ...]
if not not test_inds_bad.tolist():
first_good_ind = np.setdiff1d(
np.arange(X_test2.shape[0]), test_inds_bad)[0]
X_test2[test_inds_bad,
...] = X_test2[(test_inds_bad * 0) +
first_good_ind, ...]
# Estimate covariances matrices
cov_data_train = pyriemann.estimation.Covariances(
'lwf').fit_transform(X_train)
cov_data_val = pyriemann.estimation.Covariances(
'lwf').fit_transform(X_validate)
cov_data_test = pyriemann.estimation.Covariances(
'lwf').fit_transform(X_test2)
# Create MDM model
mdm = pyriemann.classification.MDM()
# Fit model and store train/val/test accuracies
t_fit_start = time.time()
clf = mdm.fit(cov_data_train, Y_train.ravel())
last_epochs[frodo, 1] = time.time() - t_fit_start
accs[frodo,
0] = accuracy_score(Y_train.ravel(),
clf.predict(cov_data_train))
accs[frodo,
1] = accuracy_score(Y_validate.ravel(),
clf.predict(cov_data_val))
accs[frodo,
2] = accuracy_score(y_test2.ravel(),
clf.predict(cov_data_test))
# Save model
chckpt_path = sp+modeltype+'_'+pat_id_curr+'_testday_'+\
str(test_day)+'_fold'+str(frodo)+save_suffix+'.sav'
pickle.dump(clf, open(chckpt_path, 'wb'))
else:
# Fit NN model and store accuracies
chckpt_path = sp+'checkpoint_'+modeltype+'_'+pat_id_curr+'_testday_'+\
str(test_day)+'_fold'+str(frodo)+save_suffix+'.h5'
accs_lst, last_epoch_tmp = cnn_model(
X_train,
Y_train,
X_validate,
Y_validate,
X_test2,
y_test2,
chckpt_path,
modeltype,
nb_classes=nb_classes,
dropoutRate=dropoutRate,
kernLength=kernLength,
F1=F1,
D=D,
F2=F2,
dropoutType=dropoutType,
kernLength_sep=kernLength_sep,
loss=loss,
optimizer=optimizer,
patience=patience,
early_stop_monitor=early_stop_monitor,
do_log=do_log,
epochs=epochs,
compute_val=compute_val,
ecog_srate=ecog_srate)
for ss in range(3):
accs[frodo, ss] = accs_lst[ss]
last_epochs[frodo, :] = last_epoch_tmp
# Save accuracies (train/val/test)
np.save(
sp + 'acc_' + modeltype + '_' + pat_id_curr + '_testday_' +
str(test_day) + save_suffix + '.npy', accs)
np.save(
sp + 'last_training_epoch_gen_tf' + modeltype + '_' +
pat_id_curr + '_testday_' + str(test_day) + save_suffix +
'.npy', last_epochs)
# Return validation accuracy for hyperparameter tuning (assumes only 1 model and 1 subject)
return accs[:, 1].mean()
# very bad
'lstm': lstm,
# very bad
'gru': gru,
# very good
'bidirectional_lstm': bidirectional_lstm,
'bidirectional_gru': bidirectional_gru
}
def create_model():
return models_available[MODEL_NAME]()
print('loading the data')
data_train = model_utils.load_data()
texts = []
labels = []
intents = model_utils.load_labels()
intents_lookup = model_utils.get_intents_lookup(intents)
inputs = np.zeros((len(data_train), MAX_SEQUENCE_LENGTH, EMBEDDING_DIM))
for idx, (text, intent) in enumerate(data_train):
encoded = model_utils.encode_sentence(text)
# copy the values, equivalent of padding
inputs[idx, :encoded.shape[0], :encoded.
shape[1]] = encoded[:MAX_SEQUENCE_LENGTH, :]
# append the id of the intent
labels.append(intents_lookup[intent])
def main():
parser = argparse.ArgumentParser()
add_test_args(parser)
add_common_args(parser)
args = parser.parse_args()
x_train, y_train_biden, y_train_trump, mask_train, x_dev, y_dev_biden, y_dev_trump, mask_dev, container = model_utils.load_data(
args.dataset_dir, dev_frac=args.dev_frac, max_entries=args.dataset_cap)
model = model_utils.load_model(
models.get_model(args.model)(), args.load_path)
model.eval()
# Change this line to hear other kinds of samples.
dataset = x_train
for i in range(dataset.shape[0]):
print(f"Playing Combined {i}...")
container.data = dataset[i, 0].numpy()
play(container.invert())
y_b, y_t = model(dataset[i:i + 1].abs())
container.data = (torch.clamp(y_b / dataset[i:i + 1].abs(), 0, 1) *
dataset[i:i + 1]).detach().numpy()[0, 0]
print(f"Playing model output {i}...")
play(container.invert())
type=str,
help="Model Architecture")
parser.add_argument("--learning_rate",
default=0.005,
type=float,
help="Learning Rate")
parser.add_argument("--hidden_units",
default=512,
type=int,
help="Hidden units")
parser.add_argument("--epochs", default=10, type=int, help="Epochs")
parser.add_argument("--gpu",
action='store_true',
default=False,
help="GPU")
args = parser.parse_args()
print('---------Parameters----------')
print('gpu = {!r}'.format(args.gpu))
print('epoch(s) = {!r}'.format(args.epochs))
print('arch = {!r}'.format(args.arch))
print('learning_rate = {!r}'.format(args.learning_rate))
print('hidden_units = {!r}'.format(args.hidden_units))
print('-----------------------------')
train_loader, valid_loader, _, class_to_idx = load_data(args.data_dir)
best_model = training(args.arch, args.hidden_units, args.learning_rate,
args.epochs, args.save_dir, train_loader,
valid_loader, class_to_idx)
def main():
[[X_train, y_train], [X_valid, y_valid],
[X_test, y_test]] = util.load_data(DATA_PATH)
print('Merging data...')
X = X_train.append(X_valid).append(X_test)
y = y_train.append(y_valid).append(y_test)
del X_train
del X_valid
del X_test
del y_train
del y_valid
del y_test
X = make_squared(X, SQ_FIELDS)
X = make_interactions(X, INTERACTIONS)
full = X.copy()
full['beetle'] = y['beetle']
X = X.drop([
'studyArea', 'x', 'y', 'elev_srtm30', 'year',
'varPrecip_growingSeason', 'precip_OctSep:varPrecip_growingSeason'
],
axis=1)
predictors = list(X)
print('Fitting model to full data set...')
scaler = StandardScaler()
X = scaler.fit_transform(X)
y = y['beetle'].values.reshape(-1)
logistic_clf = LogisticRegression(C=0.001, penalty='l2')
logistic_clf.fit(X, y)
coefs = pd.DataFrame(
[[pred, coef]
for pred, coef in zip(predictors, logistic_clf.coef_[0])],
columns=['predictor', 'coef'])
coefs['abs'] = np.abs(coefs.coef)
coefs = coefs.sort_values('abs', ascending=False)
coefs = coefs.drop(['abs'], axis=1)
print('Model Coefficients:\n', coefs)
x_range, y_range = get_ranges(full, verbose=True)
historic_years = range(1903, 2000)
year = 1999
next_year_data = full.loc[full.year == (year + 1), :]
while year >= historic_years[0]:
hist_data = pd.read_csv(HISTORIC_DATA_PATH + 'clean_%d.csv' % year)
hist_data = mask_data(hist_data, x_range, y_range, verbose=False)
hist_data = make_squared(hist_data, SQ_FIELDS)
hist_data = make_interactions(hist_data, INTERACTIONS)
print('\nBeginning predictions for', year)
xy = next_year_data.apply(lambda row: str(row['x']) + str(row['y']),
axis=1)
print(' Reducing %d data to study area...' % year)
extras = find_extra_rows(hist_data, xy)
hist_data = hist_data.drop(extras, axis=0)
hist_data = hist_data.rename(
columns={'precipPreious_OctSep': 'precipPrevious_OctSep'})
if year == historic_years[-1]:
hist_merge = hist_data[['x', 'y']]
print(' Ascertaining rows are aligned...')
assert list(hist_data.x) == list(next_year_data.x)
assert list(hist_data.y) == list(next_year_data.y)
hist_data.index = next_year_data.index
hist_merge.index = hist_data.index
hist_essentials = pd.DataFrame(hist_data[predictors[0]])
print(' Keeping essentials...')
for p in predictors[1:]:
hist_essentials[p] = hist_data[p]
hist_essentials = scaler.fit_transform(hist_essentials)
print(' Predicting...')
probs = logistic_clf.predict_proba(hist_essentials)
probs = np.array([prob[1] for prob in probs])
hist_merge.loc[:, 'probs_%d' % year] = probs
hist_merge.loc[:, 'preds_%d' % year] = list(
map(lambda x: 1 if x >= OPTIMAL_THRESHOLD else 0, probs))
print(' Saving data so far....')
hist_merge.to_csv(HISTORIC_DATA_PATH + 'predictions_no_beetle.csv')
year -= 1
next_year_data = hist_data
if args.model == models.LSTM_NAME:
config.set("LSTM", "timesteps", args.timesteps)
if args.model == models.CNN_NAME:
config.set("CNN", "timesteps", args.timesteps)
if args.model == models.LG_NAME:
config.set("MLR", "timesteps", args.timesteps)
if args.model == models.MLP_NAME:
config.set("MLP", "timesteps", args.timesteps)
config.set(args.model, "external_embeddings", embeddings)
train_conf = dict(config.items("Settings"))
print 'Loading training data...',
words, labels = model_utils.load_data(args.training,
path_spells,
train=True)
print "[OK]"
path_weights = args.model_weights
for n in range(1, args.S + 1):
args.model_weights = path_weights.replace(".hdf5",
"_" + str(n) + ".hdf5")
#Instantiating the model
print('Initializing model'), args.model
if args.model == models.LG_NAME:
m = models.LogisticRegressionHP(conf=dict(config.items('MLR')),
forms=words,
labels=labels,
parser.add_argument("--learning_rate",
type=float,
default=0.001,
help="set learning rate")
# hidden_units is irrelevant
parser.add_argument("--hidden_units",
type=int,
default=1024,
help="set hidden units")
parser.add_argument("--epochs", type=int, default=1, help="set epochs")
parser.add_argument("--gpu",
action="store_const",
const="cuda",
default="cpu",
help="use gpu")
parser.add_argument("--save_dir", help="save model")
args = parser.parse_args()
cat_to_name = read_json(args.category_names)
trainloader, testloader, validloader, train_data = load_data(args.data_dir)
model, history = make_NN(n_hidden=[args.hidden_units], n_epoch=args.epochs, labelsdict=cat_to_name, lr=args.learning_rate, device=args.gpu, \
model_name=args.arch, trainloader=trainloader, validloader=validloader, train_data=train_data)
with open('loss_history.pickle', 'wb') as f:
pickle.dump(history, f)
if args.save_dir:
save_whole_model(model, args.save_dir)
"""
Augment and save the dataset. Augments are rotation and flipping. Fake colors are added too. Saved as .tif files.
"""
if __name__ == '__main__':
source_path = sys.argv[1]
dest_path = sys.argv[2]
# Load dataset
image_paths = glob.glob(os.path.join(source_path, "*.tif"))
for image_path in image_paths:
image = model_utils.load_data(image_path, image_path.replace("images", "labels"))
# Do preprocessing and image augmentation
train_x, train_y = model_utils.convert_training_images_to_numpy_arrays([image], normalize=False)
train_y = model_utils.replace_class(train_y, class_id=5)
train_x = model_utils.fake_colors(train_x)
train_x = model_utils.image_augmentation(train_x)
train_y = model_utils.image_augmentation(train_y)
# Save the images
for i in range(train_x.shape[0]):
augmented_image_x = train_x[i, :, :, :]
augmented_image_y = train_y[i, :, :, :]
augmented_image = data_processing.TrainingImage(augmented_image_x, augmented_image_y,
geo_transform=image.geo_transform,
projection=image.projection)
data_output_path = os.path.join(dest_path, "images",
def main():
parser = argparse.ArgumentParser()
add_train_args(parser)
add_common_args(parser)
args = parser.parse_args()
add_experiment(args)
device = model_utils.get_device()
# Load dataset from disk
x_train, y_train_biden, y_train_trump, mask_train, x_dev, y_dev_biden, y_dev_trump, mask_dev, container = model_utils.load_data(
args.dataset_dir, dev_frac=args.dev_frac, max_entries=args.dataset_cap)
train_dl = data.DataLoader(
data.TensorDataset(x_train, y_train_biden, y_train_trump, mask_train),
batch_size=args.train_batch_size,
shuffle=True,
)
dev_dl = data.DataLoader(
data.TensorDataset(x_dev, y_dev_biden, y_dev_trump, mask_dev),
batch_size=args.val_batch_size,
shuffle=False,
)
# Initialize a model
model = models.get_model(args.model)()
# load from checkpoint if path specified
if args.load_path is not None:
model = model_utils.load_model(model, args.load_path)
# Move model to GPU if necessary
model.to(device)
# Initialize optimizer
optimizer = optim.Adam(
model.parameters(),
lr=args.learning_rate,
weight_decay=args.weight_decay,
)
# Scheduler
scheduler = optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode='min',
factor=0.5,
patience=30,
verbose=True,
)
os.makedirs(f'{args.save_path}/{args.experiment}')
print(f'Created new experiment: {args.experiment}')
save_arguments(args, f'{args.save_path}/{args.experiment}/args.txt')
# Train!
trained_model = train_model(
train_dl,
dev_dl,
model,
optimizer,
scheduler,
args,
)
# Save trained model
filename = f'{args.save_path}/{args.experiment}/{model.__class__.__name__}_trained.checkpoint'
model_utils.save_model(trained_model, filename)
train_path = '/nethome/zyu336/dl_catdog/data/train/'
img_shape = (3, 224, 224)
batch_size = 16
n_epochs = 20
sample_interval = 400
cuda = torch.cuda.is_available()
latent_dim = 256
train_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
dataloader, _, n_classes, _, _, _, _ = load_data('breeds_cat', train_transform,
train_transform, batch_size)
n_classes = len(n_classes)
class Flatten(nn.Module):
def forward(self, x):
N, C, H, W = x.size()
return x.view(N, -1)
class Unflatten(nn.Module):
def __init__(self, N=-1, C=3, H=224, W=224):
super(Unflatten, self).__init__()
self.N = N
self.C = C
def transfer_learn_nn(lp, sp, model_type = 'eegnet_hilb', layers_to_finetune = None,
n_train_trials = 50, per_train_trials = 0.6, n_val_trials = 50, per_val_trials = 0.3,
n_test_trials = 50, use_per_vals = False,loss='categorical_crossentropy', optimizer='adam',
patience=5,early_stop_monitor='val_loss',norm_rate=0.25,use_prev_opt_early_params=True,
single_sub=False, compute_val='power', ecog_srate=500, epochs = 20,
data_lp=None, pats_ids_in=None, test_day=None, n_train_sbj=None, n_folds=None,
proj_mat_lp=None):
'''
Main script for performing transfer learning across folds. Matches code from run_nn_models.py.
If doing test_day = 'last', only need to specify train and val trials/percent because test set is known.
'''
# Ensure layers_to_finetune is a list
if (layers_to_finetune is not None) and (not isinstance(layers_to_finetune, list)):
layers_to_finetune = [layers_to_finetune]
# Create suffix for saving files (so can save results from different train/val sizes to same folder)
if use_per_vals:
suffix_trials = '_ptra'+str(int(per_train_trials*100))+'_pval'+str(int(per_val_trials*100))
else:
suffix_trials = '_ntra'+str(n_train_trials)+'_nval'+str(n_val_trials)+'_ntes'+str(n_test_trials)
# Load param file from pre-trained model
file_pkl = open(lp+'param_file.pkl', 'rb')
params_dict = pickle.load(file_pkl)
file_pkl.close()
# Extract appropriate parameters from param file
tlim = params_dict['tlim']
if test_day==None:
test_day = params_dict['test_day']
if pats_ids_in==None:
pats_ids_in = params_dict['pats_ids_in']
rand_seed = params_dict['rand_seed']
n_test_sbj = params_dict['n_test']
n_val_sbj = params_dict['n_val']
if n_folds==None:
n_folds = params_dict['n_folds']
save_suffix = params_dict['save_suffix']
do_log = params_dict['do_log']
if data_lp==None:
data_lp = params_dict['lp']
if n_train_sbj==None:
if 'n_train' in list(params_dict.keys()):
n_train_sbj = params_dict['n_train']
else:
n_train_sbj = 7
if 'epochs' in list(params_dict.keys()):
epochs = params_dict['epochs']
compute_val = params_dict['compute_val']
ecog_srate = params_dict['ecog_srate']
if use_prev_opt_early_params:
# Use model fitting parameters from pre-trained model
loss = params_dict['loss']
optimizer = params_dict['optimizer']
patience = params_dict['patience']
early_stop_monitor = params_dict['early_stop_monitor']
# Load in hyperparameters
dropoutRate = params_dict['dropoutRate']
kernLength = params_dict['kernLength']
F1 = params_dict['F1']
D = params_dict['D']
F2 = params_dict['F2']
dropoutType = params_dict['dropoutType']
kernLength_sep = params_dict['kernLength_sep']
# Find pathnames of models from all folds
model_fnames = natsort.natsorted(glob.glob(lp + 'checkpoint_gen_'+model_type+'_fold*.h5'))
# Set random seed
np.random.seed(rand_seed)
# Load projection matrix (electrodes to ROI's) from pre-trained model
if proj_mat_lp==None:
proj_mat_out = np.load(lp+'proj_mat_out.npy')
n_chans_all = len(np.nonzero(proj_mat_out.reshape(-1,proj_mat_out.shape[-1]).mean(axis=0))[0])
else:
proj_mat_out = np.load(proj_mat_lp+'proj_mat_out.npy')
n_chans_all = params_dict['n_chans_all']
if proj_mat_out.shape[-1]>n_chans_all:
proj_mat_out = proj_mat_out[...,:n_chans_all]
elif proj_mat_out.shape[-1]<n_chans_all:
proj_mat_out_tmp = proj_mat_out.copy()
proj_sh = [val for val in proj_mat_out_tmp.shape]
proj_sh[-1] = n_chans_all
proj_mat_out = np.zeros(proj_sh)
proj_mat_out[...,:proj_mat_out_tmp.shape[-1]] = proj_mat_out_tmp
# Load ECoG data for all subjects
if test_day == 'last':
# If test day is 'last', load in last day's data for all subjects
X_all,y_all,X_test_last,y_test_last,sbj_order_all,sbj_order_test_last = load_data(pats_ids_in, data_lp,
n_chans_all=n_chans_all,
test_day=test_day, tlim=tlim)
else:
X_all,y_all,_,_,sbj_order_all,_ = load_data(pats_ids_in, data_lp,
n_chans_all=n_chans_all,
test_day=None, tlim=tlim)
# Identify the number of unique labels (or classes) present
nb_classes = len(np.unique(y_all))
# Choose subjects for training/validation/testing for every fold (random seed keeps this consistent to pre-trained data)
sbj_inds_all_train, sbj_inds_all_val, sbj_inds_all_test = folds_choose_subjects(n_folds, pats_ids_in,
n_test=n_test_sbj, n_val=n_val_sbj,
n_train=n_train_sbj)
print("Subject indices are: ", sbj_inds_all_test, len(sbj_inds_all_test))
# Determine train/val/test inds for every fold
labels_unique = np.unique(y_all)
nb_classes = len(labels_unique)
half_n_evs_test = 'nopad' #avoids duplicating events (will take all available events)
acc_pretrain = np.zeros([n_folds,3])
acc_trained = acc_pretrain.copy()
acc_single_sub = acc_pretrain.copy()
acc_single_sub_0 = acc_single_sub.copy()
last_epochs_TL = np.zeros([n_folds,2])
last_epochs_SS = np.zeros([n_folds,2])
for i in range(n_folds):
# Determine subjects in train/val/test sets for current fold
test_sbj = sbj_inds_all_test[i]
val_sbj = sbj_inds_all_val[i]
train_sbj = sbj_inds_all_train[i]
# First, find indices for all events associated with test subject
other_inds = subject_data_inds(np.full(1, test_sbj), sbj_order_all, labels_unique, i,
'test_inds', half_n_evs_test, y_all, sp, n_folds, [])
trainval_inds = np.asarray(list(set(other_inds)))
if test_day == 'last':
# Find all events for last day for test subject
test_inds = subject_data_inds(np.full(1, test_sbj), sbj_order_test_last, labels_unique, i,
'test_inds', half_n_evs_test, y_test_last, sp, n_folds, [])
# Determine number of train and val events (trials) to obtain
if use_per_vals:
n_train = int(len(trainval_inds) * per_train_trials)
n_val = int(len(trainval_inds) * per_val_trials)
else:
n_train = int(n_train_trials)
n_val = int(n_val_trials)
# Find train event indices
train_inds_tmp = subject_data_inds(np.full(1, test_sbj), sbj_order_all[trainval_inds], labels_unique, i,
'train_inds', n_train//nb_classes, y_all[trainval_inds], sp, n_folds, [])
#I think above is supposed to be 'train_inds'
train_inds = trainval_inds[train_inds_tmp] #convert back to original inds
# Remove train events and choose val inds from remaining events
# Note: if n_train is larger than available events for training data, finding validation events
# will throw an error because there are no remaining events to choose from
remain_inds = np.asarray(list(set(trainval_inds) - set(train_inds))) # remove train inds
if len(remain_inds) == 0:
sys.exit("Error: No data to pick from for validation set!")
val_inds_tmp = subject_data_inds(np.full(1, test_sbj), sbj_order_all[remain_inds], labels_unique, i,
'val_inds', n_val//nb_classes, y_all[remain_inds], sp, n_folds, [])
val_inds = remain_inds[val_inds_tmp] # convert back to original inds
else:
# If test_day is not last, then determine number of train, val, and test events (trials) to obtain
if use_per_vals:
n_train = int(len(other_inds) * per_train_trials)
n_val = int(len(other_inds) * per_val_trials)
n_test = int(len(other_inds) * (1-per_train_trials-per_val_trials))
else:
n_train = int(n_train_trials)
n_val = int(n_val_trials)
n_test = int(n_test_trials)
# Find train event indices
train_inds_tmp = subject_data_inds(np.full(1, test_sbj), sbj_order_all[trainval_inds], labels_unique, i,
'train_inds', n_train//nb_classes, y_all[trainval_inds], sp, n_folds, [])
train_inds = trainval_inds[train_inds_tmp] # convert back to original inds
# Remove train events and choose val inds from remaining events
# Note: if n_train is larger than available events for training data, finding validation events
# will throw an error because there are no remaining events to choose from
valtest_inds = np.asarray(list(set(other_inds) - set(train_inds))) #remove train inds
if len(valtest_inds) == 0:
sys.exit("Error: No data to pick from for validation and test sets!")
val_inds_tmp = subject_data_inds(np.full(1, test_sbj), sbj_order_all[valtest_inds], labels_unique, i,
'val_inds', n_val//nb_classes, y_all[valtest_inds], sp, n_folds, [])
val_inds = valtest_inds[val_inds_tmp] # convert back to original inds
# Remove val events and choose test inds from remaining events
# Note: if n_train+n_val is larger than available events for training data, finding test events
# will throw an error because there are no remaining events to choose from
remain_inds = np.asarray(list(set(valtest_inds) - set(val_inds))) # remove train inds
if len(remain_inds) == 0:
sys.exit("Error: No data to pick from for test set!")
test_inds_tmp = subject_data_inds(np.full(1, test_sbj), sbj_order_all[remain_inds], labels_unique, i,
'test_inds', n_test//nb_classes, y_all[remain_inds], sp, n_folds, [])
test_inds = remain_inds[test_inds_tmp] # convert back to original inds
# Generate train/val/test data based on event indices for each fold
X_train = X_all[train_inds,...]
Y_train = y_all[train_inds]
sbj_order_train = sbj_order_all[train_inds] # important for projection matrix input
X_validate = X_all[val_inds,...]
Y_validate = y_all[val_inds]
sbj_order_validate = sbj_order_all[val_inds] # important for projection matrix input
if test_day is None:
X_test = X_all[test_inds,...]
Y_test = y_all[test_inds]
sbj_order_test = sbj_order_all[test_inds] # important for projection matrix input
else:
# If test_day is last, use loaded data from last days only
X_test = X_test_last[test_inds,...]
Y_test = y_test_last[test_inds]
sbj_order_test = sbj_order_test_last[test_inds] # important for projection matrix input
# Reformat data size for NN
Y_train = np_utils.to_categorical(Y_train-1)
X_train = np.expand_dims(X_train,1)
Y_validate = np_utils.to_categorical(Y_validate-1)
X_validate = np.expand_dims(X_validate,1)
Y_test = np_utils.to_categorical(Y_test-1)
X_test = np.expand_dims(X_test,1)
proj_mat_out2 = np.expand_dims(proj_mat_out,1)
# Run transfer learning
str_len = len('checkpoint_gen_')
curr_mod_fname = model_fnames[i].split('/')[-1][:-3]
chckpt_path = sp+'checkpoint_gen_tf_'+curr_mod_fname[str_len:]+suffix_trials+'.h5'
acc_pretrain_tmp, acc_trained_tmp, last_epoch_tmp = run_transfer_learning(model_fnames[i], sbj_order_train,
X_train, Y_train, sbj_order_validate,
X_validate, Y_validate, sbj_order_test,
X_test, Y_test,proj_mat_out2, chckpt_path,
layers_to_finetune = layers_to_finetune,
norm_rate = norm_rate, loss=loss,
optimizer=optimizer, patience = patience,
early_stop_monitor = early_stop_monitor,
do_log=do_log, nb_classes = nb_classes,
epochs = epochs)
# Here need to run the single subject on the same amount of training and val data
if single_sub:
test_sbj_name = pats_ids_in[int(test_sbj)]
chckpt_path = sp+'checkpoint_gen_tf_single_sub_'+test_sbj_name+suffix_trials+'.h5'
acc_single_sub_tmp, last_epoch_single_tmp, acc_single_sub_tmp_0 = run_single_sub_percent_compare(sbj_order_train, X_train, Y_train,
sbj_order_validate, X_validate, Y_validate, sbj_order_test,
X_test, Y_test, chckpt_path, norm_rate = norm_rate,
loss=loss, optimizer=optimizer, patience = patience,
early_stop_monitor = early_stop_monitor, do_log=do_log, nb_classes = nb_classes,
compute_val=compute_val, ecog_srate=ecog_srate, epochs = epochs,
dropoutRate = dropoutRate, kernLength = kernLength, F1 = F1, D = D, F2 = F2,
dropoutType = dropoutType,kernLength_sep = kernLength_sep)
acc_single_sub[i,:] = acc_single_sub_tmp
last_epochs_SS[i,:] = last_epoch_single_tmp
acc_single_sub_0[i,:] = acc_single_sub_tmp_0
# Save train/val/test accuracies for every fold
acc_pretrain[i,:] = acc_pretrain_tmp
acc_trained[i,:] = acc_trained_tmp
last_epochs_TL[i,:] = last_epoch_tmp
# Save accuracies across all folds (adds suffix for number/percentage of trials)
np.save(sp+'acc_gen_tf_pretrain_'+model_type+'_'+str(n_folds)+save_suffix+suffix_trials+'.npy',acc_pretrain)
np.save(sp+'acc_gen_tf_trained_'+model_type+'_'+str(n_folds)+save_suffix+suffix_trials+'.npy',acc_trained)
np.save(sp+'last_training_epoch_gen_tf'+model_type+'_'+str(n_folds)+save_suffix+suffix_trials+'.npy', last_epochs_TL)
if single_sub:
np.save(sp+'acc_gen_tf_singlesub_'+model_type+'_'+str(n_folds)+save_suffix+suffix_trials+'.npy',acc_single_sub)
np.save(sp+'acc_gen_tf_singlesub0_'+model_type+'_'+str(n_folds)+save_suffix+suffix_trials+'.npy',acc_single_sub_0)
np.save(sp+'last_training_epoch_gen_tf_singlesub_'+model_type+'_'+str(n_folds)
+save_suffix+suffix_trials+'.npy', last_epochs_SS)
def transfer_learn_nn_eeg(lp, sp, eeg_data_lp,
model_type = 'eegnet_hilb', layers_to_finetune = None,
n_train_trials = 50, per_train_trials = 0.6, n_val_trials = 50, per_val_trials = 0.3,
n_test_trials = 50, use_per_vals = False,loss='categorical_crossentropy', optimizer='adam',
patience=5,early_stop_monitor='val_loss',norm_rate=0.25,use_prev_opt_early_params=True,
single_sub=False, compute_val='power', ecog_srate=500, epochs = 20):
'''
Main script for performing transfer learning across folds. Matches code from run_nn_models.py.
If doing test_day = 'last', only need to specify train and val trials/percent because test set is known.
'''
# Parameters for projection matrix
custom_rois = True
n_chans_eeg = 61
n_chans_ecog = 126 # number of channels in ecog data (expected by model)
per_test_trials = 0.2 # percentage of EEG data to use for test set
# Ensure layers_to_finetune is a list
if (layers_to_finetune is not None) and (not isinstance(layers_to_finetune, list)):
layers_to_finetune = [layers_to_finetune]
# Create suffix for saving files (so can save results from different train/val sizes to same folder)
if use_per_vals:
suffix_trials = '_ptra'+str(int(per_train_trials*100))+'_pval'+str(int(per_val_trials*100))
else:
suffix_trials = '_ntra'+str(n_train_trials)+'_nval'+str(n_val_trials)+'_ntes'+str(n_test_trials)
# Load param file from pre-trained model
file_pkl = open(lp+'param_file.pkl', 'rb')
params_dict = pickle.load(file_pkl)
file_pkl.close()
# Extract appropriate parameters from param file
tlim = params_dict['tlim']
test_day = params_dict['test_day']
# pats_ids_in = params_dict['pats_ids_in']
pats_ids_in = ['EE'+str(val).zfill(2) for val in np.arange(1,16).tolist()]
rand_seed = params_dict['rand_seed']
n_test_sbj = params_dict['n_test']
n_val_sbj = params_dict['n_val']
n_folds = params_dict['n_folds']
save_suffix = params_dict['save_suffix']
do_log = params_dict['do_log']
# data_lp = params_dict['lp']
if 'n_train' in list(params_dict.keys()):
n_train_sbj = params_dict['n_train']
else:
n_train_sbj = 7
if 'epochs' in list(params_dict.keys()):
epochs = params_dict['epochs']
compute_val = params_dict['compute_val']
ecog_srate = params_dict['ecog_srate']
if use_prev_opt_early_params:
# Use model fitting parameters from pre-trained model
loss = params_dict['loss']
optimizer = params_dict['optimizer']
patience = params_dict['patience']
early_stop_monitor = params_dict['early_stop_monitor']
# Load in hyperparameters
dropoutRate = params_dict['dropoutRate']
kernLength = params_dict['kernLength']
F1 = params_dict['F1']
D = params_dict['D']
F2 = params_dict['F2']
dropoutType = params_dict['dropoutType']
kernLength_sep = params_dict['kernLength_sep']
# Find pathnames of models from all folds
model_fnames = natsort.natsorted(glob.glob(lp + 'checkpoint_gen_'+model_type+'_fold*.h5'))
# Set random seed
np.random.seed(rand_seed)
# Load projection matrix (electrodes to ROI's) for EEG data
if custom_rois:
custom_roi_inds = get_custom_motor_rois() # load custom roi's from precentral, postcentral, and inf parietal (AAL2)
else:
custom_roi_inds = None
print("Determining ROIs")
proj_mat_out,good_ROIs,chan_ind_vals_all = proj_mats_good_rois(['EE01_bH'],
n_chans_all = n_chans_eeg,
rem_bad_chans=False,
dipole_dens_thresh=None,
custom_roi_inds=custom_roi_inds,
chan_cut_thres=n_chans_eeg,
roi_proj_loadpath= eeg_data_lp+'proj_mat/')
nROIs = len(good_ROIs)
print("ROIs found")
n_chans_all = n_chans_eeg
# Load EEG data for each subject and fit model
X_all,y_all,_,_,sbj_order_all,_ = load_data(pats_ids_in, eeg_data_lp, test_day=None, tlim=tlim, n_chans_all=n_chans_eeg)
X_all[np.isnan(X_all)] = 0 # set all NaN's to 0
for pat_ind,curr_pat in enumerate(pats_ids_in):
# Identify the number of unique labels (or classes) present
nb_classes = len(np.unique(y_all))
# Determine train/val/test inds for every fold
labels_unique = np.unique(y_all)
nb_classes = len(labels_unique)
half_n_evs_test = 'nopad' #avoids duplicating events (will take all available events)
acc_pretrain = np.zeros([n_folds,3])
acc_trained = acc_pretrain.copy()
acc_single_sub = acc_pretrain.copy()
acc_single_sub_0 = acc_single_sub.copy()
last_epochs_TL = np.zeros([n_folds,2])
last_epochs_SS = np.zeros([n_folds,2])
for i in range(n_folds):
# First, find indices for all events associated with test subject
other_inds = subject_data_inds(np.full(1, pat_ind), sbj_order_all, labels_unique, i,
'test_inds', half_n_evs_test, y_all, sp, n_folds, [])
trainval_inds = np.asarray(list(set(other_inds)))
# If test_day is not last, then determine number of train, val, and test events (trials) to obtain
if use_per_vals:
n_train = int(len(other_inds) * per_train_trials*(1-per_test_trials))
n_val = int(len(other_inds) * per_val_trials*(1-per_test_trials))
n_test = int(len(other_inds) * per_test_trials) #(1-per_train_trials-per_val_trials))
else:
n_train = int(n_train_trials)
n_val = int(n_val_trials)
n_test = int(n_test_trials)
# Find train event indices
test_inds_tmp = subject_data_inds(np.full(1, pat_ind), sbj_order_all[trainval_inds], labels_unique, i,
'test_inds', n_test//nb_classes, y_all[trainval_inds], sp, n_folds, [])
test_inds = trainval_inds[test_inds_tmp] # convert back to original inds
# Remove train events and choose val inds from remaining events
# Note: if n_train is larger than available events for training data, finding validation events
# will throw an error because there are no remaining events to choose from
valtest_inds = np.asarray(list(set(other_inds) - set(test_inds))) #remove train inds
if len(valtest_inds) == 0:
sys.exit("Error: No data to pick from for validation and test sets!")
val_inds_tmp = subject_data_inds(np.full(1, pat_ind), sbj_order_all[valtest_inds], labels_unique, i,
'val_inds', n_val//nb_classes, y_all[valtest_inds], sp, n_folds, [])
val_inds = valtest_inds[val_inds_tmp] # convert back to original inds
# Remove val events and choose test inds from remaining events
# Note: if n_train+n_val is larger than available events for training data, finding test events
# will throw an error because there are no remaining events to choose from
remain_inds = np.asarray(list(set(valtest_inds) - set(val_inds))) # remove train inds
if len(remain_inds) == 0:
sys.exit("Error: No data to pick from for test set!")
train_inds_tmp = subject_data_inds(np.full(1, pat_ind), sbj_order_all[remain_inds], labels_unique, i,
'train_inds', n_train//nb_classes, y_all[remain_inds], sp, n_folds, [])
train_inds = remain_inds[train_inds_tmp] # convert back to original inds
# Append train/val/test event indices for each fold
# train_inds_folds.append(train_inds)
# val_inds_folds.append(val_inds)
# test_inds_folds.append(test_inds)
# Reformat data size for NN
Y_all = np_utils.to_categorical(y_all-1)
X_all_tmp = np.expand_dims(X_all,1)
proj_mat_out2 = np.tile(proj_mat_out,[X_all_tmp.shape[0],1,1])
proj_mat_out2 = np.expand_dims(proj_mat_out2,1)
# Pad channel dimension to match ECoG data
X_all_sh = list(X_all_tmp.shape)
X_all_sh[2] = n_chans_ecog
X_all_resh = np.zeros(X_all_sh)
X_all_resh[...,:n_chans_eeg,:] = X_all_tmp
proj_mat_out3 = np.zeros(list(proj_mat_out2.shape[:-1])+[n_chans_ecog])
proj_mat_out3[...,:n_chans_eeg] = proj_mat_out2
# Generate train/val/test data based on event indices for each fold
X_train = X_all_resh[train_inds,...]
Y_train = Y_all[train_inds,...]
sbj_order_train = np.zeros(len(train_inds)).astype('int') # important for projection matrix input
X_validate = X_all_resh[val_inds,...]
Y_validate = Y_all[val_inds,...]
sbj_order_validate = np.zeros(len(val_inds)).astype('int') # important for projection matrix input
X_test = X_all_resh[test_inds,...]
Y_test = Y_all[test_inds,...]
sbj_order_test = np.zeros(len(test_inds)).astype('int') # important for projection matrix input
# Note that sbj_order doesn't matter here because all EEG subjects have same electrode locations
# Run transfer learning
str_len = len('checkpoint_gen_')
curr_mod_fname = model_fnames[i].split('/')[-1][:-3]
chckpt_path = sp+'checkpoint_gen_tf_'+curr_pat+'_'+curr_mod_fname[str_len:]+suffix_trials+'.h5'
acc_pretrain_tmp, acc_trained_tmp, last_epoch_tmp = run_transfer_learning(model_fnames[i], sbj_order_train,
X_train, Y_train, sbj_order_validate,
X_validate, Y_validate, sbj_order_test,
X_test, Y_test,proj_mat_out3, chckpt_path,
layers_to_finetune = layers_to_finetune,
norm_rate = norm_rate, loss=loss,
optimizer=optimizer, patience = patience,
early_stop_monitor = early_stop_monitor,
do_log=do_log, nb_classes = nb_classes,
epochs = epochs)
# Here need to run the single subject on the same amount of training and val data
if single_sub:
chckpt_path = sp+'checkpoint_gen_tf_single_sub_'+curr_pat+'_'+suffix_trials+'.h5'
acc_single_sub_tmp, last_epoch_single_tmp, acc_single_sub_tmp_0 = run_single_sub_percent_compare(sbj_order_train, X_train, Y_train,
sbj_order_validate, X_validate, Y_validate, sbj_order_test,
X_test, Y_test, chckpt_path, norm_rate = norm_rate,
loss=loss, optimizer=optimizer, patience = patience,
early_stop_monitor = early_stop_monitor, do_log=do_log, nb_classes = nb_classes,
compute_val=compute_val, ecog_srate=ecog_srate, epochs = epochs,
dropoutRate = dropoutRate, kernLength = kernLength, F1 = F1, D = D, F2 = F2,
dropoutType = dropoutType,kernLength_sep = kernLength_sep)
acc_single_sub[i,:] = acc_single_sub_tmp
last_epochs_SS[i,:] = last_epoch_single_tmp
acc_single_sub_0[i,:] = acc_single_sub_tmp_0
# Save train/val/test accuracies for every fold
acc_pretrain[i,:] = acc_pretrain_tmp
acc_trained[i,:] = acc_trained_tmp
last_epochs_TL[i,:] = last_epoch_tmp
# Save accuracies across all folds (adds suffix for number/percentage of trials)
np.save(sp+'acc_gen_tf_pretrain_'+curr_pat+'_'+model_type+'_'+str(n_folds)+save_suffix+\
suffix_trials+'.npy',acc_pretrain)
np.save(sp+'acc_gen_tf_trained_'+curr_pat+'_'+model_type+'_'+str(n_folds)+\
save_suffix+suffix_trials+'.npy',acc_trained)
np.save(sp+'last_training_epoch_gen_tf'+curr_pat+'_'+model_type+'_'+str(n_folds)+\
save_suffix+suffix_trials+'.npy', last_epochs_TL)
if single_sub:
np.save(sp+'acc_gen_tf_singlesub_'+curr_pat+'_'+model_type+'_'+str(n_folds)+\
save_suffix+suffix_trials+'.npy',acc_single_sub)
np.save(sp+'acc_gen_tf_singlesub0_'+curr_pat+'_'+model_type+'_'+str(n_folds)+\
save_suffix+'.npy',acc_single_sub_0)
np.save(sp+'last_training_epoch_gen_tf_singlesub_'+curr_pat+'_'+model_type+'_'+str(n_folds)
+save_suffix+suffix_trials+'.npy', last_epochs_SS)