def training(model, FLAGS, model_name, smi_total, prop_total):
np.set_printoptions(threshold=sys.maxsize)
print("Start Training XD")
stuff = open("C:/Users/Zac Hung/Desktop/masterthesis/experiment/baysept9thamesfirst.txt", 'w')
stuff1 = open("C:/Users/Zac Hung/Desktop/masterthesis/experiment/bayfirstmdames.txt", 'w')
stuff2 = open("C:/Users/Zac Hung/Desktop/masterthesis/experiment/bayamesaleat.txt", 'w')
stuff3 = open("C:/Users/Zac Hung/Desktop/masterthesis/experiment/bayamesepist.txt", 'w')
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/aug12/MC-Dropout_HIV.ckpt-31")
num_epochs = FLAGS.epoch_size
batch_size = FLAGS.batch_size
init_lr = FLAGS.init_lr
total_st = time.time()
smi_train, smi_eval, smi_test = split_train_eval_test(smi_total, 0, 1, 0)
prop_train, prop_eval, prop_test = split_train_eval_test(prop_total, 0, 1, 0)
prop_eval = np.asarray(prop_eval)
prop_test = np.asarray(prop_test)
num_train = len(smi_train)
num_eval = len(smi_eval)
num_test = len(smi_test)
smi_train = smi_train[:num_train]
prop_train = prop_train[:num_train]
num_batches_train = (num_train // batch_size) + 1
num_batches_eval = (num_eval // batch_size) + 1
num_batches_test = (num_test // batch_size) + 1
num_sampling = 20
total_iter = 0
print("Number of- training data:", num_train, "\t evaluation data:", num_eval, "\t test data:", num_test)
for epoch in range(num_epochs):
st = time.time()
lr = init_lr * 0.5 ** (epoch // 10)
model.assign_lr(lr)
#smi_train, prop_train = shuffle_two_list(smi_train, prop_train)
prop_train = np.asarray(prop_train)
# TRAIN
num = 0
train_loss = 0.0
Y_pred_total = np.array([])
Y_batch_total = np.array([])
for i in range(num_batches_train):
num += 1
st_i = time.time()
total_iter += 1
#tmp=smi_train[i * batch_size:(i + 1) * batch_size]
A_batch, X_batch = convert_to_graph(smi_train[i * batch_size:(i + 1) * batch_size], FLAGS.max_atoms)
Y_batch = prop_train[i * batch_size:(i + 1) * batch_size]
#mtr = np.abs(model.get_feature(A_batch, X_batch, Y_batch))
# print(np.shape(mtr))
#print(len(tmp))
#count = -1
# for i in tmp:
# count += 1
# iMol = Chem.MolFromSmiles(i.strip())
#
# start= (np.argpartition((mtr[count]),-10))
# start=np.array((start[start<len(Chem.rdmolops.GetAdjacencyMatrix(iMol))])).tolist()[0:9]
# #stuff.write(str(smi_test[count][start:end + 1]) + "\n")
# #print(len(Chem.rdmolops.GetAdjacencyMatrix(iMol)))
# print(start)
# print(rdkit.Chem.rdmolfiles.MolFragmentToSmiles(iMol,start))
Y_mean, _, loss = model.train(A_batch, X_batch, Y_batch)
train_loss += loss
Y_pred = np_sigmoid(Y_mean.flatten())
Y_pred_total = np.concatenate((Y_pred_total, Y_pred), axis=0)
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
et_i = time.time()
train_loss /= num
train_accuracy = accuracy_score(Y_batch_total, np.around(Y_pred_total).astype(int))
train_auroc = 0.0
try:
train_auroc = roc_auc_score(Y_batch_total, Y_pred_total)
except:
train_auroc = 0.0
# Eval
Y_pred_total = np.array([])
Y_batch_total = np.array([])
num = 0
eval_loss = 0.0
for i in range(num_batches_eval):
evalbatch=smi_eval[i * batch_size:(i + 1) * batch_size]
A_batch, X_batch = convert_to_graph(evalbatch, FLAGS.max_atoms)
Y_batch = prop_eval[i * batch_size:(i + 1) * batch_size]
# mtr_eval = np.abs(model.get_feature(A_batch, X_batch, Y_batch))
# print(np.shape(mtr_eval))
# count=-1
# print(len(evalbatch))
# for i in evalbatch:
# count += 1
# iMol = Chem.MolFromSmiles(i.strip())
#
# #start= (np.argpartition((mtr_eval[count]),-10))
# start=mtr_eval[count]
# start=start[start>0.1]
# start=np.array((start[start<len(Chem.rdmolops.GetAdjacencyMatrix(iMol))])).tolist()[0:9]
# #stuff.write(str(smi_test[count][start:end + 1]) + "\n")
# #print(len(Chem.rdmolops.GetAdjacencyMatrix(iMol)))
# print(start)
# print(rdkit.Chem.rdmolfiles.MolFragmentToSmiles(iMol,start))
# MC-sampling
P_mean = []
for n in range(1):
num += 1
Y_mean, _, loss = model.test(A_batch, X_batch, Y_batch)
eval_loss += loss
P_mean.append(Y_mean.flatten())
P_mean = np_sigmoid(np.asarray(P_mean))
mean = np.mean(P_mean, axis=0)
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
Y_pred_total = np.concatenate((Y_pred_total, mean), axis=0)
eval_loss /= num
eval_accuracy = accuracy_score(Y_batch_total, np.around(Y_pred_total).astype(int))
eval_auroc = 0.0
try:
eval_auroc = roc_auc_score(Y_batch_total, Y_pred_total)
except:
eval_auroc = 0.0
# Save network!
ckpt_path = 'save/' + model_name + '.ckpt'
model.save(ckpt_path, epoch)
et = time.time()
# Print Results
print("Time for", epoch, "-th epoch: ", et - st)
print("Loss Train:", round(train_loss, 3), "\t Evaluation:", round(eval_loss, 3))
print("Accuracy Train:", round(train_accuracy, 3), "\t Evaluation:", round(eval_accuracy, 3))
print("AUROC Train:", round(train_auroc, 3), "\t Evaluation:", round(eval_auroc, 3))
total_et = time.time()
print("Finish training! Total required time for training : ", (total_et - total_st))
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/aug15thames/MC-Dropout_HIV.ckpt-12")
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/amesaug16th/MC-Dropout_HIV.ckpt-7")
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/aug16-256/MC-Dropout_HIV.ckpt-4")
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/aug16thsummax/MC-Dropout_HIV.ckpt-4")
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/aug17japan/MC-Dropout_HIV.ckpt-9")
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/japanaug18bay/MC-Dropout_HIV.ckpt-8")
model.restore("./AmesBays/MC-Dropout_HIV.ckpt-11")
#model.restore("C:/Users/Zac Hung/PycharmProjects/mythesis/uq_molecule/japanfullbayfirstmodSept7th/MC-Dropout_HIV.ckpt-4")
# Test
test_st = time.time()
Y_pred_total = np.array([])
Y_batch_total = np.array([])
ale_unc_total = np.array([])
epi_unc_total = np.array([])
tot_unc_total = np.array([])
num = 0
test_loss = 0.0
count = -1
for i in range(num_batches_test):
num += 1
testBatch=smi_test[i * batch_size:(i + 1) * batch_size]
A_batch, X_batch = convert_to_graph(testBatch, FLAGS.max_atoms)
Y_batch = prop_test[i * batch_size:(i + 1) * batch_size]
mtr_test = np_sigmoid(model.get_feature(A_batch, X_batch, Y_batch))
#print(np.shape(mtr_test))
count = -1
#print(len(testBatch))
for j in testBatch:
count += 1
iMol = Chem.MolFromSmiles(j.strip())
adj_len=len(Chem.rdmolops.GetAdjacencyMatrix(iMol))
#start= (np.argpartition((mtr_test[count]),-10))
start=maxSum(mtr_test[count],adj_len,10)
#start = mtr_test[count]
#print("adj_len",adj_len)
#start = (np.squeeze(np.argwhere(start > 1)))
#print("this is start",start)
#print("adj len",adj_len)
#print(j)
#print(start)
#start = np.array((start[start < adj_len])).tolist()[0:9]
# stuff.write(str(smi_test[count][start:end + 1]) + "\n")
# print(len(Chem.rdmolops.GetAdjacencyMatrix(iMol)))
#print(start)
#print(rdkit.Chem.rdmolfiles.MolFragmentToSmiles(iMol, start))
#bondNum=Chem.rdchem.Mol.GetNumBonds(iMol)
#tmp=rdkit.Chem.rdmolfiles.MolFragmentToSmarts(iMol, atomsToUse=start,bondsToUse=list(range(1,bondNum)),isomericSmarts=False)
#print(tmp)
#tmp4=(rdkit.Chem.rdmolfiles.MolFragmentToSmarts(iMol, atomsToUse=start))
#print(tmp4)
stuff.write(processUnit(iMol,start,i,batch_size,count,mtr_test[count],adj_len,10)+ "\n")
#stuff.write(tmp4+ "\n")
#uncomment this for the drawing.
#fig = Draw.MolToFile(iMol, "./amesfirstmodImg3/"+str(i*batch_size+count)+'.png', size=size, highlightAtoms=start)
# MC-sampling
P_mean = []
for n in range(5):
Y_mean, _, loss = model.test(A_batch, X_batch, Y_batch)
P_mean.append(Y_mean.flatten())
# mtr = np.abs(model.get_feature(A_batch, X_batch, Y_batch))
# print(np.shape(mtr))
# for j in range(len(Y_batch)):
# count += 1
#
# start, end = maxSum(mtr[j], 503, 15)
# stuff.write(str(smi_test[count][start:end + 1]) + "\n")
P_mean = np_sigmoid(np.asarray(P_mean))
mean = np.mean(P_mean, axis=0)
ale_unc = np.mean(P_mean * (1.0 - P_mean), axis=0)
epi_unc = np.mean(P_mean ** 2, axis=0) - np.mean(P_mean, axis=0) ** 2
tot_unc = ale_unc + epi_unc
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
Y_pred_total = np.concatenate((Y_pred_total, mean), axis=0)
ale_unc_total = np.concatenate((ale_unc_total, ale_unc), axis=0)
epi_unc_total = np.concatenate((epi_unc_total, epi_unc), axis=0)
tot_unc_total = np.concatenate((tot_unc_total, tot_unc), axis=0)
stuff1.write(str(np.around(Y_pred_total)) + "\n")
stuff2.write(str(ale_unc_total) + "\n")
stuff3.write(str(epi_unc_total) + "\n")
#print("ale:",ale_unc_total)
#print("epi",epi_unc_total)
True_positive = 0
False_postive = 0
True_negative = 0
False_negative = 0
Exp = Y_batch_total
Pred = np.around(Y_pred_total)
for i in range(len(Exp)):
if (Exp[i] == Pred[i] and Exp[i] == 1):
True_positive += 1
if (Exp[i] != Pred[i] and Exp[i] == 0):
False_postive += 1
if (Exp[i] == Pred[i] and Exp[i] == 0):
True_negative += 1
if (Exp[i] != Pred[i] and Exp[i] == 1):
False_negative += 1
count_TP = True_positive
print("True Positive:", count_TP)
count_FP = False_postive
print("False Positive", count_FP)
count_FN = False_negative
print("False Negative:", count_FN)
count_TN = True_negative
print("True negative:", count_TN)
Accuracy = (count_TP + count_TN) / (count_TP + count_FP + count_FN + count_TN)
print("Accuracy:", Accuracy)
import math
MCC = (count_TP * count_TN - count_FP * count_FN) / math.sqrt(abs((count_TN
+ count_FP)
* (count_TN + count_FN)
* (count_TP + count_FP)
* (count_TP + count_FN)))
print("MCC", MCC)
Specificity = count_TN / (count_TN + count_FP)
print("Specificity:", Specificity)
Precision = count_TP / (count_TP + count_FP)
print("Precision:", Precision)
# sensitivity
Recall = count_TP / (count_TP + count_FN)
print("Recall:", Recall)
# F1
Fmeasure = (2 * count_TP) / (2 * count_TP + count_FP + count_FN)
print("Fmeasure", Fmeasure)
test_et = time.time()
print("Finish Testing, Total time for test:", (test_et - test_st))
return
def training(model, FLAGS, model_name, smi_total, prop_total):
print("Start Training XD")
num_epochs = FLAGS.epoch_size
batch_size = FLAGS.batch_size
init_lr = FLAGS.init_lr
total_st = time.time()
smi_train, smi_eval, smi_test = split_train_eval_test(
smi_total, 0.8, 0.2, 0.1)
prop_train, prop_eval, prop_test = split_train_eval_test(
prop_total, 0.8, 0.2, 0.1)
prop_eval = np.asarray(prop_eval)
prop_test = np.asarray(prop_test)
num_train = len(smi_train)
num_eval = len(smi_eval)
num_test = len(smi_test)
smi_train = smi_train[:num_train]
prop_train = prop_train[:num_train]
num_batches_train = (num_train // batch_size) + 1
num_batches_eval = (num_eval // batch_size) + 1
num_batches_test = (num_test // batch_size) + 1
num_sampling = 20
total_iter = 0
print("Number of- training data:", num_train, "\t evaluation data:",
num_eval, "\t test data:", num_test)
for epoch in range(num_epochs):
st = time.time()
lr = init_lr * 0.5**(epoch // 10)
model.assign_lr(lr)
smi_train, prop_train = shuffle_two_list(smi_train, prop_train)
prop_train = np.asarray(prop_train)
# TRAIN
num = 0
train_loss = 0.0
Y_pred_total = np.array([])
Y_batch_total = np.array([])
for i in range(num_batches_train):
num += 1
st_i = time.time()
total_iter += 1
A_batch, X_batch = convert_to_graph(
smi_train[i * batch_size:(i + 1) * batch_size],
FLAGS.max_atoms)
Y_batch = prop_train[i * batch_size:(i + 1) * batch_size]
Y_noise = np.random.normal(0.0, FLAGS.noise, Y_batch.shape[0])
Y_batch = Y_batch + Y_noise
Y_mean, Y_logvar, loss = model.train(A_batch, X_batch, Y_batch)
train_loss += loss
Y_pred = Y_mean.flatten()
Y_pred_total = np.concatenate((Y_pred_total, Y_pred), axis=0)
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
et_i = time.time()
#print ("train_iter : ", total_iter, ", epoch : ", epoch, ", loss : ", loss, "\t Time:", (et_i-st_i))
train_loss /= num
train_mae = np.mean(np.abs(Y_batch_total - Y_pred_total))
#Eval
Y_pred_total = np.array([])
Y_batch_total = np.array([])
num = 0
eval_loss = 0.0
for i in range(num_batches_eval):
A_batch, X_batch = convert_to_graph(
smi_eval[i * batch_size:(i + 1) * batch_size], FLAGS.max_atoms)
Y_batch = prop_eval[i * batch_size:(i + 1) * batch_size]
# MC-sampling
P_mean = []
P_logvar = []
for n in range(3):
num += 1
Y_mean, Y_logvar, loss = model.test(A_batch, X_batch, Y_batch)
eval_loss += loss
P_mean.append(Y_mean.flatten())
P_mean = np.asarray(P_mean)
mean = np.mean(P_mean, axis=0)
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
Y_pred_total = np.concatenate((Y_pred_total, mean), axis=0)
eval_loss /= num
eval_mae = np.mean(np.abs(Y_batch_total - Y_pred_total))
# Save network!
ckpt_path = 'save/' + model_name + '.ckpt'
model.save(ckpt_path, epoch)
et = time.time()
# Print Results
print("Time for", epoch, "-th epoch: ", et - st)
print("Loss Train:", round(train_loss, 3), "\t Evaluation:",
round(eval_loss, 3))
print("MAE Train:", round(train_mae, 3), "\t Evaluation:",
round(eval_mae, 3))
total_et = time.time()
print("Finish training! Total required time for training : ",
(total_et - total_st))
#Test
test_st = time.time()
Y_pred_total = np.array([])
Y_batch_total = np.array([])
ale_unc_total = np.array([])
epi_unc_total = np.array([])
tot_unc_total = np.array([])
num = 0
test_loss = 0.0
for i in range(num_batches_test):
num += 1
A_batch, X_batch = convert_to_graph(
smi_test[i * batch_size:(i + 1) * batch_size], FLAGS.max_atoms)
Y_batch = prop_test[i * batch_size:(i + 1) * batch_size]
# MC-sampling
P_mean = []
P_logvar = []
for n in range(num_sampling):
Y_mean, Y_logvar, loss = model.test(A_batch, X_batch, Y_batch)
P_mean.append(Y_mean.flatten())
P_logvar.append(Y_logvar.flatten())
P_mean = np.asarray(P_mean)
P_logvar = np.exp(np.asarray(P_logvar))
mean = np.mean(P_mean, axis=0)
ale_unc = np.mean(P_logvar, axis=0)
epi_unc = np.var(P_mean, axis=0)
tot_unc = ale_unc + epi_unc
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
Y_pred_total = np.concatenate((Y_pred_total, mean), axis=0)
ale_unc_total = np.concatenate((ale_unc_total, ale_unc), axis=0)
epi_unc_total = np.concatenate((epi_unc_total, epi_unc), axis=0)
tot_unc_total = np.concatenate((tot_unc_total, tot_unc), axis=0)
np.save('./statistics/' + model_name + '_mc_truth.npy', Y_batch_total)
np.save('./statistics/' + model_name + '_mc_pred.npy', Y_pred_total)
np.save('./statistics/' + model_name + '_mc_epi_unc.npy', epi_unc_total)
np.save('./statistics/' + model_name + '_mc_ale_unc.npy', ale_unc_total)
np.save('./statistics/' + model_name + '_mc_tot_unc.npy', tot_unc_total)
test_et = time.time()
print("Finish Testing, Total time for test:", (test_et - test_st))
return
def training(model, FLAGS, model_name, smi_total, prop_total):
print ("Start Training XD")
num_epochs = FLAGS.epoch_size
batch_size = FLAGS.batch_size
init_lr = FLAGS.init_lr
total_st = time.time()
smi_train, smi_eval, smi_test = split_train_eval_test(smi_total, 0.8, 0.2, 0.1)
prop_train, prop_eval, prop_test = split_train_eval_test(prop_total, 0.8, 0.2, 0.1)
prop_eval = np.asarray(prop_eval)
prop_test = np.asarray(prop_test)
num_train = len(smi_train)
num_eval = len(smi_eval)
num_test = len(smi_test)
smi_train = smi_train[:num_train]
prop_train = prop_train[:num_train]
num_batches_train = (num_train//batch_size) + 1
num_batches_eval = (num_eval//batch_size) + 1
num_batches_test = (num_test//batch_size) + 1
num_sampling = 20
total_iter = 0
print("Number of- training data:", num_train, "\t evaluation data:", num_eval, "\t test data:", num_test)
for epoch in range(num_epochs):
st = time.time()
lr = init_lr * 0.5**(epoch//10)
model.assign_lr(lr)
smi_train, prop_train = shuffle_two_list(smi_train, prop_train)
prop_train = np.asarray(prop_train)
# TRAIN
num = 0
train_loss = 0.0
Y_pred_total = np.array([])
Y_batch_total = np.array([])
for i in range(num_batches_train):
num += 1
st_i = time.time()
total_iter += 1
A_batch, X_batch = convert_to_graph(smi_train[i*batch_size:(i+1)*batch_size], FLAGS.max_atoms)
Y_batch = prop_train[i*batch_size:(i+1)*batch_size]
Y_mean, _, loss = model.train(A_batch, X_batch, Y_batch)
train_loss += loss
Y_pred = np_sigmoid(Y_mean.flatten())
Y_pred_total = np.concatenate((Y_pred_total, Y_pred), axis=0)
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
et_i = time.time()
train_loss /= num
train_accuracy = accuracy_score(Y_batch_total, np.around(Y_pred_total).astype(int))
train_auroc = 0.0
try:
train_auroc = roc_auc_score(Y_batch_total, Y_pred_total)
except:
train_auroc = 0.0
#Eval
Y_pred_total = np.array([])
Y_batch_total = np.array([])
num = 0
eval_loss = 0.0
for i in range(num_batches_eval):
A_batch, X_batch = convert_to_graph(smi_eval[i*batch_size:(i+1)*batch_size], FLAGS.max_atoms)
Y_batch = prop_eval[i*batch_size:(i+1)*batch_size]
# MC-sampling
P_mean = []
for n in range(3):
num += 1
Y_mean, _, loss = model.test(A_batch, X_batch, Y_batch)
eval_loss += loss
P_mean.append(Y_mean.flatten())
P_mean = np_sigmoid(np.asarray(P_mean))
mean = np.mean(P_mean, axis=0)
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
Y_pred_total = np.concatenate((Y_pred_total, mean), axis=0)
eval_loss /= num
eval_accuracy = accuracy_score(Y_batch_total, np.around(Y_pred_total).astype(int))
eval_auroc = 0.0
try:
eval_auroc = roc_auc_score(Y_batch_total, Y_pred_total)
except:
eval_auroc = 0.0
# Save network!
ckpt_path = 'save/'+model_name+'.ckpt'
model.save(ckpt_path, epoch)
et = time.time()
# Print Results
print ("Time for", epoch, "-th epoch: ", et-st)
print ("Loss Train:", round(train_loss,3), "\t Evaluation:", round(eval_loss,3))
print ("Accuracy Train:", round(train_accuracy,3), "\t Evaluation:", round(eval_accuracy,3))
print ("AUROC Train:", round(train_auroc,3), "\t Evaluation:", round(eval_auroc,3))
total_et = time.time()
print ("Finish training! Total required time for training : ", (total_et-total_st))
#Test
test_st = time.time()
Y_pred_total = np.array([])
Y_batch_total = np.array([])
ale_unc_total = np.array([])
epi_unc_total = np.array([])
tot_unc_total = np.array([])
num = 0
test_loss = 0.0
for i in range(num_batches_test):
num += 1
A_batch, X_batch = convert_to_graph(smi_test[i*batch_size:(i+1)*batch_size], FLAGS.max_atoms)
Y_batch = prop_test[i*batch_size:(i+1)*batch_size]
# MC-sampling
P_mean = []
for n in range(num_sampling):
Y_mean, _, loss = model.test(A_batch, X_batch, Y_batch)
P_mean.append(Y_mean.flatten())
P_mean = np_sigmoid(np.asarray(P_mean))
mean = np.mean(P_mean, axis=0)
ale_unc = np.mean(P_mean*(1.0-P_mean), axis=0)
epi_unc = np.mean(P_mean**2, axis=0) - np.mean(P_mean, axis=0)**2
tot_unc = ale_unc + epi_unc
Y_batch_total = np.concatenate((Y_batch_total, Y_batch), axis=0)
Y_pred_total = np.concatenate((Y_pred_total, mean), axis=0)
ale_unc_total = np.concatenate((ale_unc_total, ale_unc), axis=0)
epi_unc_total = np.concatenate((epi_unc_total, epi_unc), axis=0)
tot_unc_total = np.concatenate((tot_unc_total, tot_unc), axis=0)
np.save('./statistics/'+model_name+'_mc_truth.npy', Y_batch_total)
np.save('./statistics/'+model_name+'_mc_pred.npy', Y_pred_total)
np.save('./statistics/'+model_name+'_mc_epi_unc.npy', epi_unc_total)
np.save('./statistics/'+model_name+'_mc_ale_unc.npy', ale_unc_total)
np.save('./statistics/'+model_name+'_mc_tot_unc.npy', tot_unc_total)
test_et = time.time()
print ("Finish Testing, Total time for test:", (test_et-test_st))
return