def print_prediction(model, img_list, main_batch_size, run_name, output_prediction=False):
#
scores = model.evaluate_generator(GeneratorImgs(img_list, batch_size=main_batch_size, width=WIDTH, height=HEIGHT),
steps=len(img_list)/main_batch_size)
#
print('Test loss: {:.4f}'.format(scores[0]))
print('Test accuracy: {:.4f}'.format(scores[1]))
#
preds_proba = list()
preds = list()
conf_by_tumor = np.zeros((8,2))
labels = list()
class_count = np.array([0,0])
imgname = list()
predictions = list()
#
if(output_prediction):
fpred = open("predictions/{}".format(run_name), "w")
#
for x, y, z in ReadImgs(img_list, width=WIDTH, height=HEIGHT):
predictions.append(model.predict(np.array([x])).squeeze())
labels.append(y.argmax())
imgname.append(z)
#
predictions = np.array(predictions)
class_count_real, class_count_aug = classes_count(LoadBreakhisList(TRAIN_FILE))
predictions /= class_count_aug
predictions *= class_count_real
print("P_train: ",class_count_aug)
print("P_real: ",class_count_real)
#
for i in range(len(predictions)):
if(output_prediction):
fpred.write("{};{};".format(imgname[i].split("/")[-1], labels[i]))
class_count[labels[i]] += 1
preds.append(predictions[i].argmax())
preds_proba.append(predictions[i][labels[i]])
conf_by_tumor[TumorToLabel8(imgname[i])][np.argmax(predictions[i])] += 1
if(output_prediction):
for j in predictions[i]:
fpred.write("{:.4f};".format(j))
fpred.write("\n")
#
if(output_prediction):
fpred.close()
#
fpr, tpr, _ = roc_curve(labels, preds_proba, pos_label=0)
roc_auc = auc(fpr, tpr)
#
print("Test AUC 0: {:.4f}".format(roc_auc))
#
print(classification_report(labels, preds, target_names=["malign", "benign"]))
#
#fpr, tpr, _ = roc_curve(labels, preds_proba, pos_label=1)
#roc_auc = auc(fpr, tpr)
#
#print("Test AUC 1: {:.4f}".format(roc_auc))
a = confusion_matrix(labels, preds)
print("Confusion matrix:\n",a)
print("Total elements per class:", class_count)
if(a[0][0]+a[1][0] != 0 and a[0][1]+a[1][1] != 0):
print("Accuracy per class: {:.3f} {:.3f} {:.3f}\n".format(float(a[0][0])/(a[0][0]+a[0][1]), float(a[1][1])/(a[1][0]+a[1][1]), (float(a[0][0])/(a[0][0]+a[0][1])+float(a[1][1])/(a[1][0]+a[1][1]))/2))
else:
print("Zero predictions in one class.")
print("Confusion by tumor type: \n", conf_by_tumor)
print("Accuracy by patient: ", accuracy_by_patient(img_list, labels, preds))
print("Malign Benign")
print("M_LC\nM_MC\nM_PC\nM_DC\nB_TA\nB_A\nB_PT\nB_F")
if(i.find("rotat") == -1 and i.find("flip") == -1 and i.find("trans") == -1):
class_count_real += label
class_count_aug += label
return class_count_real.astype("float32")/(class_count_real[0]+class_count_real[1]), class_count_aug.astype("float32")/(class_count_aug[0]+class_count_aug[1])
#
#
#
model = build_cnn(int(sys.argv[1]))
model.summary()
print(model.optimizer.get_config())
print("Epochs: ",EPOCHS)
print("Batch-size: ", BATCH_SIZE)
#
#
#
train_imgs = LoadBreakhisList(TRAIN_FILE)
random.shuffle(train_imgs)
val_imgs = LoadBreakhisList(VAL_FILE)
random.shuffle(val_imgs)
#
main_batch_size = BATCH_SIZE
nr_batches_val = len(val_imgs)/main_batch_size
nr_batches = len(train_imgs)/main_batch_size
#
#
#
#for i,j in GeneratorImgs(train_imgs, batch_size=main_batch_size):
# for t in range(len(i)):
# print(j)
#
#for i,j in GeneratorImgs(val_imgs, batch_size=main_batch_size):