def run_repeated_training_binary(cfg_in):
# create the empty lists for holding all the data
vec_train_acc = []
vec_valid_acc = []
vec_test_acc = []
vec_y_true = []
vec_y_pred = []
vec_y_pred_prob = []
mat_vec_idx_absolute_test = []
# set the random seed for all the splits
temp = initialize_config_split(copy.deepcopy(cfg_in))
np.random.seed(temp.random_seed)
# initiate the repeated training
for i in range(cfg_in.n_repeats):
cfg = copy.deepcopy(cfg_in)
print("\n\nIteration: {}".format(i + 1))
# now load the preprocessed data and the label
f_data_handle = "preproc_data_{}_{}.pkl".format(
cfg.vec_idx_patient[0], cfg.vec_idx_patient[1])
f_data = cfg.d_preproc / f_data_handle
with open(str(f_data), 'rb') as handle:
X = pickle.load(handle)
y = generate_labels(cfg.str_feature, cfg, bool_append_csv_to_cfg=True)
# now prepare data for training
cfg = initialize_config_split(cfg)
X, y = correct_data_label(X, y, cfg)
Xs, ys = prepare_data_for_train(X, y, cfg)
print("\nx_train Angiography cube shape: {}".format(Xs[0][0].shape))
print("x_train Structure OCT cube shape: {}".format(Xs[0][1].shape))
print("x_train B scan shape: {}".format(Xs[0][2].shape))
print("y_train onehot shape: {}".format(ys[0].shape))
print("\nx_valid Angiography cube shape: {}".format(Xs[1][0].shape))
print("x_valid Structure OCT cube shape: {}".format(Xs[1][1].shape))
print("x_valid B scan shape: {}".format(Xs[1][2].shape))
print("y_valid onehot shape: {}".format(ys[1].shape))
print("\nx_test Angiography cube shape: {}".format(Xs[2][0].shape))
print("x_test Structure OCT cube shape: {}".format(Xs[2][1].shape))
print("x_test B scan shape: {}".format(Xs[2][2].shape))
print("y_test onehot shape: {}".format(ys[2].shape))
# finally set the training parameters
cfg = initialize_config_training(cfg, bool_debug=cfg.bool_debug)
model = get_model(cfg.str_arch, cfg)
callbacks = get_callbacks(cfg)
h = model.fit(Xs[0],
ys[0],
batch_size=cfg.batch_size,
epochs=cfg.n_epoch,
verbose=2,
callbacks=callbacks,
validation_data=(Xs[1], ys[1]),
shuffle=False,
validation_batch_size=Xs[1][0].shape[0])
# Now perform prediction
train_set_score = model.evaluate(Xs[0],
ys[0],
callbacks=callbacks,
verbose=0)
valid_set_score = model.evaluate(Xs[1],
ys[1],
callbacks=callbacks,
verbose=0)
test_set_score = model.evaluate(Xs[2],
ys[2],
callbacks=callbacks,
verbose=0)
vec_train_acc.append(train_set_score[1])
vec_valid_acc.append(valid_set_score[1])
vec_test_acc.append(test_set_score[1])
if cfg.num_classes == 2:
y_true = ys[-1]
y_pred_logits = model.predict(Xs[2])
y_pred = y_pred_logits.copy()
y_pred[y_pred >= 0.5] = 1
y_pred[y_pred < 0.5] = 0
y_pred = y_pred.reshape(-1)
# now transform the logits into a matrix of two class probabilities and append
y_pred_logits_both = np.concatenate(
[1 - y_pred_logits, y_pred_logits], axis=1)
else:
raise ValueError("The number of classes should be two")
vec_y_true.append(y_true)
vec_y_pred.append(y_pred)
vec_y_pred_prob.append(y_pred_logits_both)
mat_vec_idx_absolute_test.append(cfg.vec_idx_absolute[2])
Xs = []
ys = []
print('\n' + '=' * 80)
print("Average train set accuracy: {} + ".format(np.mean(vec_train_acc)),
np.std(vec_train_acc))
print("Average valid set accuracy: {} + ".format(np.mean(vec_valid_acc)),
np.std(vec_valid_acc))
print("Average test set accuracy: {} + ".format(np.mean(vec_test_acc)),
np.std(vec_test_acc))
# now concatenate the results together
y_true = np.concatenate(vec_y_true, axis=0)
y_pred = np.concatenate(vec_y_pred, axis=0)
y_pred_prob = np.concatenate(vec_y_pred_prob, axis=0)
vec_idx_absolute_test = np.concatenate(mat_vec_idx_absolute_test, axis=0)
# sanity check
if not y_true.shape[0] == y_pred.shape[0] == y_pred_prob.shape[
0] == vec_idx_absolute_test.shape[0]:
raise ValueError("These should be equal")
# print the overall test set accuracy
print("\nOverall test set accuracy: {}".format(
np.sum(y_true == y_pred) / len(y_true)))
# now load the preprocessed data and the label
f_data_handle = "preproc_data_{}_{}.pkl".format(cfg_in.vec_idx_patient[0],
cfg_in.vec_idx_patient[1])
f_data = cfg_in.d_preproc / f_data_handle
with open(str(f_data), 'rb') as handle:
X = pickle.load(handle)
y = generate_labels(cfg_in.str_feature,
cfg_in,
bool_append_csv_to_cfg=True)
# now prepare data for training
cfg_in = initialize_config_split(cfg_in)
X, y = correct_data_label(X, y, cfg_in)
_, _ = prepare_data_for_train(X, y, cfg_in)
X, y = None, None
# append the hyperparameters to the cfg structure
cfg_in.vec_train_acc = vec_train_acc
cfg_in.vec_valid_acc = vec_valid_acc
cfg_in.vec_test_acc = vec_test_acc
cfg_in.y_test_true = y_true
cfg_in.y_test_pred = y_pred
cfg_in.y_test_pred_prob = y_pred_prob
cfg_in.vec_idx_absolute_test = vec_idx_absolute_test
cfg_in.vec_y_true = vec_y_true
cfg_in.vec_y_pred = vec_y_pred
cfg_in.vec_y_pred_prob = vec_y_pred_prob
cfg_in.mat_vec_idx_absolute_test = mat_vec_idx_absolute_test
# now get the configuration right for saving the output
cfg_in.str_model = cfg_in.str_arch
cfg_in.f_model = "{}_{}".format(cfg_in.str_arch,
time.strftime("%Y%m%d_%H%M%S"))
cfg_in.p_figure = cfg_in.d_model / cfg_in.str_model / cfg_in.f_model
cfg_in.p_figure.mkdir(parents=True, exist_ok=True)
cfg_in.p_cfg = cfg_in.p_figure
# plot the confusion matrices
plot_raw_conf_matrix(y_true,
y_pred,
cfg_in,
save=True,
f_figure=cfg_in.str_arch)
plot_norm_conf_matrix(y_true,
y_pred,
cfg_in,
save=True,
f_figure=cfg_in.str_arch)
# save the cfg, which contains configurations and results
save_cfg(cfg_in, overwrite=True)
# save the mat file, which contains the binary results
save_mat(cfg_in, overwrite=True)
def run_training_binary(Xs, ys, cfg):
print("\nx_train Angiography cube shape: {}".format(Xs[0][0].shape))
print("x_train Structure OCT cube shape: {}".format(Xs[0][1].shape))
print("x_train B scan shape: {}".format(Xs[0][2].shape))
print("y_train onehot shape: {}".format(ys[0].shape))
print("\nx_valid Angiography cube shape: {}".format(Xs[1][0].shape))
print("x_valid Structure OCT cube shape: {}".format(Xs[1][1].shape))
print("x_valid B scan shape: {}".format(Xs[1][2].shape))
print("y_valid onehot shape: {}".format(ys[1].shape))
print("\nx_test Angiography cube shape: {}".format(Xs[2][0].shape))
print("x_test Structure OCT cube shape: {}".format(Xs[2][1].shape))
print("x_test B scan shape: {}".format(Xs[2][2].shape))
print("y_test onehot shape: {}".format(ys[2].shape))
# Get and train model
model = get_model(cfg.str_arch, cfg)
callbacks = get_callbacks(cfg)
h = model.fit(Xs[0],
ys[0],
batch_size=cfg.batch_size,
epochs=cfg.n_epoch,
verbose=2,
callbacks=callbacks,
validation_data=(Xs[1], ys[1]),
shuffle=False,
validation_batch_size=Xs[1][0].shape[0])
cfg.history = h.history
# save trained models
save_model(model, cfg, overwrite=True, save_format='tf')
# plotting training history
plot_training_loss(h, cfg, save=True)
plot_training_acc(h, cfg, save=True)
# Now perform prediction
train_set_score = model.evaluate(Xs[0],
ys[0],
callbacks=callbacks,
verbose=0)
valid_set_score = model.evaluate(Xs[1],
ys[1],
callbacks=callbacks,
verbose=0)
test_set_score = model.evaluate(Xs[2],
ys[2],
callbacks=callbacks,
verbose=0)
print("\nTrain set accuracy: {}".format(train_set_score[1]))
print("Valid set accuracy: {}".format(valid_set_score[1]))
print("Test set accuracy: {}".format(test_set_score[1]))
cfg.vec_acc = [train_set_score[1], valid_set_score[1], test_set_score[1]]
if cfg.num_classes == 2:
y_true = ys[-1]
y_pred_logits = model.predict(Xs[2])
y_pred = y_pred_logits.copy()
y_pred[y_pred >= 0.5] = 1
y_pred[y_pred < 0.5] = 0
y_pred = y_pred.reshape(-1)
# now transform the logits into a matrix of two class probabilities and append
y_pred_logits_both = np.concatenate([1 - y_pred_logits, y_pred_logits],
axis=1)
cfg.y_test_pred_prob = y_pred_logits_both
else:
raise ValueError("The number of classes should be two")
# Printing out true and pred labels for log reg
print('Test set: ground truth')
print(y_true)
print('Test set: prediction')
print(y_pred)
# Print out the patient IDs corresponding to the query
# Here for example
# if you are running 'disease' label and you set true_label_id = 0 and predicted_label_id = 2
# then you would get the patients who are normal/healthy and but falsely classified as NV AMD
# the true_label_id and predicted_label_id correspond to cfg.vec_str_labels defined above
print(
get_patient_id_by_label(y_true,
y_pred,
true_label_id=0,
predicted_label_id=1,
cfg=cfg))
# you can also print multiple of these at the same time
print(
get_patient_id_by_label(y_true,
y_pred,
true_label_id=1,
predicted_label_id=0,
cfg=cfg))
# Extra caveat: for feature labels since we don't have possible any more and since the classes
# are automatically recasted to get the FN (patient has the feature but network predicts not present)
# you need to do something like
# print(get_patient_id_by_label(y_true, y_pred, true_label_id=1, predicted_label_id=0, cfg=cfg))
cfg.y_test_true = y_true
cfg.y_test_pred = y_pred
# now compute the test set accuracies
test_acc = np.sum(y_true == y_pred) / len(y_true)
print("\nTest accuracy: {}".format(test_acc))
cfg.test_acc_full = test_acc
# plot the confusion matrices
plot_raw_conf_matrix(y_true, y_pred, cfg, save=True)
plot_norm_conf_matrix(y_true, y_pred, cfg, save=True)
# save the cfg, which contains configurations and results
save_cfg(cfg, overwrite=True)
# save the mat file, which contains the binary results
save_mat(cfg, overwrite=True)
vec_train_acc = []
vec_valid_acc = []
vec_test_acc = []
vec_y_true = []
vec_y_pred = []
vec_model = []
# Preprocessing
Xs, ys = preprocess(vec_idx_healthy, vec_idx_dry_amd, vec_idx_cnv, cfg)
for i in range(cfg.n_ensemble):
print("\n\nIteration: {}".format(i + 1))
model = get_model('arch_011', cfg)
callbacks = get_callbacks(cfg)
h = model.fit(Xs[0],
ys[0],
batch_size=cfg.batch_size,
epochs=cfg.n_epoch,
verbose=2,
callbacks=callbacks,
validation_data=(Xs[1], ys[1]),
shuffle=False,
validation_batch_size=Xs[1][0].shape[0])
# Now perform prediction
train_set_score = model.evaluate(Xs[0],
ys[0],
callbacks=callbacks,
def run_training_cv(vec_Xs, vec_ys, cfg):
vec_history = []
# find the aggregate results on the entire dataset
vec_y_true = []
vec_y_pred = []
vec_y_pred_prob = []
# also declare empty list for the validation set
vec_y_valid_true = []
vec_y_valid_pred = []
vec_y_valid_pred_prob = []
for idx_fold in range(len(vec_Xs)):
print('\n\nFold: {}\n'.format(idx_fold))
Xs = vec_Xs[idx_fold]
ys = vec_ys[idx_fold]
print("\nx_train Angiography cube shape: {}".format(Xs[0][0].shape))
print("x_train Structure OCT cube shape: {}".format(Xs[0][1].shape))
print("x_train B scan shape: {}".format(Xs[0][2].shape))
print("x_train 3D B scan shape: {}".format(Xs[0][3].shape))
print("y_train onehot shape: {}".format(ys[0].shape))
print("\nx_valid Angiography cube shape: {}".format(Xs[1][0].shape))
print("x_valid Structure OCT cube shape: {}".format(Xs[1][1].shape))
print("x_valid B scan shape: {}".format(Xs[1][2].shape))
print("x_valid 3D B scan shape: {}".format(Xs[1][3].shape))
print("y_valid onehot shape: {}".format(ys[1].shape))
print("\nx_test Angiography cube shape: {}".format(Xs[2][0].shape))
print("x_test Structure OCT cube shape: {}".format(Xs[2][1].shape))
print("x_test B scan shape: {}".format(Xs[2][2].shape))
print("x_test 3D B scan shape: {}".format(Xs[2][3].shape))
print("y_test onehot shape: {}".format(ys[2].shape))
# Get and train model
model_curr = get_model(cfg.str_arch, cfg)
callbacks_curr = get_callbacks(cfg)
h = model_curr.fit(Xs[0],
ys[0],
batch_size=cfg.batch_size,
epochs=cfg.n_epoch,
verbose=2,
callbacks=callbacks_curr,
validation_data=(Xs[1], ys[1]),
shuffle=False,
validation_batch_size=Xs[1][0].shape[0])
vec_history.append(h.history)
# save trained models
save_model(model_curr,
cfg,
overwrite=True,
save_format='tf',
idx_cv_fold=idx_fold)
# plotting training history
plot_training_loss(h, cfg, save=True)
plot_training_acc(h, cfg, save=True)
# Now perform prediction
train_set_score = model_curr.evaluate(Xs[0],
ys[0],
callbacks=callbacks_curr,
verbose=0)
valid_set_score = model_curr.evaluate(Xs[1],
ys[1],
callbacks=callbacks_curr,
verbose=0)
test_set_score = model_curr.evaluate(Xs[2],
ys[2],
callbacks=callbacks_curr,
verbose=0)
print("\nTrain set accuracy: {}".format(train_set_score[1]))
print("Valid set accuracy: {}".format(valid_set_score[1]))
print("Test set accuracy: {}".format(test_set_score[1]))
cfg.vec_acc = [
train_set_score[1], valid_set_score[1], test_set_score[1]
]
if cfg.num_classes == 2:
# make predictions for test set
y_true = ys[-1]
y_pred_logits = model_curr.predict(Xs[2])
y_pred = y_pred_logits.copy()
y_pred[y_pred >= 0.5] = 1
y_pred[y_pred < 0.5] = 0
y_pred = y_pred.reshape(-1)
# now transform the logits into a matrix of two class probabilities and append
y_pred_logits = np.concatenate([1 - y_pred_logits, y_pred_logits],
axis=1)
# make predictions for validation set
y_valid_true = ys[1]
y_valid_pred_logits = model_curr.predict(Xs[1])
y_valid_pred = y_valid_pred_logits.copy()
y_valid_pred[y_valid_pred >= 0.5] = 1
y_valid_pred[y_valid_pred < 0.5] = 0
y_valid_pred = y_valid_pred.reshape(-1)
# nwo transform the logits into two class probabilities and append for validation set
y_valid_pred_logits = np.concatenate(
[1 - y_valid_pred_logits, y_valid_pred_logits], axis=1)
else:
# make the predictions for the test set
y_true = np.argmax(ys[-1], axis=1)
y_pred = np.argmax(model_curr.predict(Xs[2]), axis=1)
y_pred_logits = model_curr.predict(Xs[2])
# make the predictions for the validation set
y_valid_true = np.argmax(ys[1], axis=1)
y_valid_pred = np.argmax(model_curr.predict(Xs[1]), axis=1)
y_valid_pred_logits = model_curr.predict(Xs[1])
# plot the confusion matrices
plot_raw_conf_matrix(y_true, y_pred, cfg, save=True)
plot_norm_conf_matrix(y_true, y_pred, cfg, save=True)
# now append the results to a list
vec_y_true.append(y_true)
vec_y_pred.append(y_pred)
vec_y_pred_prob.append(y_pred_logits)
# append the results from the validation set also
vec_y_valid_true.append(y_valid_true)
vec_y_valid_pred.append(y_valid_pred)
vec_y_valid_pred_prob.append(y_valid_pred_logits)
# Now we are outside of the loop
y_true_unsorted_all = np.concatenate(vec_y_true, axis=-1)
y_pred_unsorted_all = np.concatenate(vec_y_pred, axis=-1)
y_pred_prob_unsorted_all = np.concatenate(vec_y_pred_prob, axis=0)
# Now obtain the correct indices
vec_idx_absolute_test_all = []
for idx_fold in range(len(vec_Xs)):
vec_idx_test_curr = cfg.vec_idx_absolute[idx_fold][-1]
vec_idx_absolute_test_all.append(vec_idx_test_curr)
vec_idx_absolute_test_all = np.concatenate(vec_idx_absolute_test_all, -1)
# Now get all the test set data
idx_permutation_sort = np.argsort(vec_idx_absolute_test_all)
y_true_all = y_true_unsorted_all[idx_permutation_sort]
y_pred_all = y_pred_unsorted_all[idx_permutation_sort]
y_pred_prob_all = y_pred_prob_unsorted_all[idx_permutation_sort, ...]
cfg.y_test_true = y_true_all
cfg.y_test_pred = y_pred_all
cfg.y_test_pred_prob = y_pred_prob_all
# also generate the data for the validation set predictions
y_valid_true_unsorted_all = np.concatenate(vec_y_valid_true, axis=-1)
y_valid_pred_unsorted_all = np.concatenate(vec_y_valid_pred, axis=-1)
y_valid_pred_prob_unsorted_all = np.concatenate(vec_y_valid_pred_prob,
axis=0)
cfg.y_valid_true = y_valid_true_unsorted_all
cfg.y_valid_pred = y_valid_pred_unsorted_all
cfg.y_valid_pred_prob = y_valid_pred_prob_unsorted_all
test_acc_full = np.sum(y_true_all == y_pred_all) / len(y_true_all)
print("\nOverall accuracy: {}".format(test_acc_full))
cfg.test_acc_full = test_acc_full
# Print out the patient IDs corresponding to the query
# Here for example
# if you are running 'disease' label and you set true_label_id = 0 and predicted_label_id = 2
# then you would get the patients who are normal/healthy and but falsely classified as NV AMD
# the true_label_id and predicted_label_id correspond to cfg.vec_str_labels defined above
# print(get_patient_id_by_label(y_true_all, y_pred_all, true_label_id=0, predicted_label_id=2, cfg=cfg))
# you can also print multiple of these at the same time
# print(get_patient_id_by_label(y_true_all, y_pred_all, true_label_id=2, predicted_label_id=1, cfg=cfg))
# Extra caveat: for feature labels since we don't have possible any more and since the classes
# are automatically recasted to get the FN (patient has the feature but network predicts not present)
# you need to do something like
print(
get_patient_id_by_label(y_true_all,
y_pred_all,
true_label_id=1,
predicted_label_id=0,
cfg=cfg))
# Plot and save the final result
plot_raw_conf_matrix(y_true_all, y_pred_all, cfg, save=True, cv_all=True)
plot_norm_conf_matrix(y_true_all, y_pred_all, cfg, save=True, cv_all=True)
# append final training history
cfg.vec_history = vec_history
# save the output as a csv file also
save_csv(y_true_all, y_pred_all, cfg)
# save the cfg, which contains configurations and results
save_cfg(cfg, overwrite=True)
# save the mat file, which contains all useful output information
save_mat(cfg, overwrite=True, bool_save_valid=True)
return cfg