Code fragments of main.py

• import os
• import os.path as osp
• import sys
• from random import randrange
• import numpy as np
• import pandas as pd
• import shutil
• import datetime
• import pytz
• import torch
• import torch.nn as nn
• from fhirpy import SyncFHIRClient
• from sklearn.preprocessing import StandardScaler
• from sklearn.metrics import confusion_matrix
• from keras.layers import Dense, Flatten, Embedding, Multiply, Concatenate, Input
• from keras.layers.advanced_activations import LeakyReLU
• from keras.models import Model, Sequential
• from keras.optimizers import adam_v2
• import csv
• import random
• from sklearn.preprocessing import MinMaxScaler
• from sklearn.model_selection import train_test_split
• # for the data generation part there are two variable
• # 1. is for if data generation method should be used or not
• # 2. is for to handle outliners like minus values which could be appear
• # the generated data can hold negative value, this value are biological not explainable, but from a informatics point of
• # view they are reasonable
• # so the station owner can decide if the methode should remove the "false" value
• # Get the env vars from station software
• # fhir_server = str(os.environ['FHIR_SERVER'])
• # fhir_port = str(os.environ['FHIR_PORT'])
• # station_name = str(os.environ['STATION_NAME'])
• # generation_synthetic_data = os.environ['GENERATION_SYNTHETIC_DATA']
• # autoremove = os.environ['AUTOREMOVE']
• # TODO: check if the enviroment variables can hold the item for synthetic-data generation and autoremove
• # TODO: comment out when deploy
• fhir_server, fhir_port = "10.50.9.90", "80"
• station_name = "UKL"
• generation_synthetic_data = False
• # autoremove = False
• assert station_name in ["UKA", "UKG", "UKK", "UKL", "IMISE", "Mittweida"]
• # Configurations
• seed = 7
• num_epochs = 10000
• input_dim = 30
• hidden_dim = 64
• num_classes = 2
• learning_rate = 0.01
• weight_decay = 0.0005
• station_name = station_name.lower()
• # Define directory of output
• here = osp.dirname(osp.abspath(__file__))
• out_dir = osp.join(here, 'output')
• if not os.path.exists(out_dir):
• os.makedirs(out_dir)
• # values for the build of the Discriminator and Generator of the cGAN
• # if we need to change them, we could adjust them here and not in every definition
• img_rows = 1
• img_cols = 31
• img_shape = (img_rows, img_cols)
• zdim = 100
• num_classes = 2
• class LogisticRegression(nn.Module):
• def __init__(self, input_dim, hidden_dim, num_classes):
• super(LogisticRegression, self).__init__()
• self.hidden_layer = nn.Linear(input_dim, hidden_dim)
• self.sigmoid = nn.Sigmoid()
• self.output_layer = nn.Linear(hidden_dim, num_classes)
• def forward(self, x):
• out = self.output_layer(self.sigmoid(self.hidden_layer(x)))
• return out
• def train(X_train, y_train, model, criterion, optimizer):
• inputs = torch.from_numpy(X_train).float()
• targets = torch.from_numpy(y_train).long()
• optimizer.zero_grad()
• outputs = model(inputs)
• loss = criterion(outputs, targets)
• loss.backward()
• optimizer.step()
• return loss.item()
• def valid(X_test, y_test, model, criterion):
• inputs = torch.from_numpy(X_test).float()
• targets = torch.from_numpy(y_test).long()
• outputs = model(inputs)
• val_loss = criterion(outputs, targets)
• _, predicted = torch.max(outputs, 1)
• cm = confusion_matrix(targets.numpy(), predicted.numpy())
• tn, fp, fn, tp = cm[0][0], cm[0][1], cm[1][0], cm[1][1]
• with np.errstate(divide='ignore', invalid='ignore'):
• val_acc = (tp + tn) / (tp + fp + fn + tn)
• val_ppv = tp / (tp + fp)
• val_precesion = tp / (tp + fp)
• val_recall = tp / (tp + fn)
• val_f1_score = 2 * tp / (2 * tp + fn + fp)
• return val_loss.item(), val_acc, val_precesion, val_recall, val_f1_score, val_ppv
• def shuffle_list(lst):
• lst2 = lst.copy()
• random.shuffle(lst2)
• return lst2
• def train_synthetic(data, iterations, batch_size, interval, number_of_row):
• Xtrainnew = data
• mydata = Xtrainnew.values.tolist()
• ytrain = []
• for j in mydata:
• ytrain.append(j[0])
• Xtrainnew = pd.DataFrame(data=mydata)
• Ytrainnew = np.array(ytrain)
• scaler = MinMaxScaler()
• scaled = scaler.fit_transform(Xtrainnew)
• Xtrain = scaled
• ytrain = Ytrainnew
• real = np.ones((batch_size, 1))
• fake = np.zeros((batch_size, 1))
• for iteration in range(iterations):
• ids = np.random.randint(0, Xtrain.shape[0], batch_size)
• imgs = Xtrain[ids]
• labels = ytrain[ids]
• z = np.random.normal(0, 1, (batch_size, 100))
• gen_imgs = gen_v.predict([z, labels])
• dloss_real = dis_v.train_on_batch([imgs, labels], real)
• dloss_fake = dis_v.train_on_batch([gen_imgs, labels], fake)
• dloss, accuracy = 0.5 * np.add(dloss_real, dloss_fake)
• z = np.random.normal(0, 1, (batch_size, 100))
• labels = np.random.randint(0, num_classes, batch_size).reshape(-1, 1)
• gloss = gan_v.train_on_batch([z, labels], real)
• if (iteration + 1) % interval == 0:
• losses.append((dloss, gloss))
• accuracies.append(100.0 * accuracy)
• iteration_checks.append(iteration + 1)
• print("%d [D loss: %f , acc: %.2f] [G loss: %f]" % (iteration + 1, dloss, 100.0 * accuracy, gloss))
• return show_data(gen_v, scaler, number_of_row)
• def savelist2csv(mynamefile, mylist):
• with open('./' + mynamefile, 'w') as myfile:
• wr = csv.writer(myfile, delimiter='\n', quoting=csv.QUOTE_MINIMAL)
• wr.writerow(mylist)
• def show_data(gen, scaler, number_of_rows):
• z = np.random.normal(0, 1, (number_of_rows, 100))
• labels = np.random.randint(2, size=number_of_rows)
• gen_imgs = gen.predict([z, labels])
• gen_imgs = scaler.inverse_transform(gen_imgs)
• for index in range(0, number_of_rows):
• gen_imgs[index] = np.around(gen_imgs[index], 4)
• gen_imgs[index][0] = np.around(gen_imgs[index][0], 0)
• return gen_imgs
• def build_gen(zdim):
• model = Sequential()
• model.add(Dense(31, input_dim=zdim))
• model.add(LeakyReLU(alpha=0.01))
• model.add(Dense(1 * 31, activation='tanh'))
• return model
• def build_cgen(zdim):
• z = Input(shape=(zdim,))
• lable = Input(shape=(1,), dtype='int32')
• lable_emb = Embedding(num_classes, zdim, input_length=1)(lable)
• lable_emb = Flatten()(lable_emb)
• joined_rep = Multiply()([z, lable_emb])
• gen_v = build_gen(zdim)
• c_img = gen_v(joined_rep)
• return Model([z, lable], c_img)
• def build_dis(img_shape):
• model = Sequential()
• model.add(Flatten(input_shape=img_shape))
• model.add(Dense(31))
• model.add(LeakyReLU(alpha=0.01))
• model.add(Dense(1, activation='sigmoid'))
• return model
• def build_cdis(img_shape):
• img = Input(shape=(img_cols,))
• lable = Input(shape=(1,), dtype='int32')
• lable_emb = Embedding(num_classes, np.prod((31)), input_length=1)(lable)
• lable_emb = Flatten()(lable_emb)
• # lable_emb=Reshape(img_shape)(lable_emb)
• concate_img = Concatenate(axis=-1)([img, lable_emb])
• dis_v = build_dis((img_rows, img_cols * 2))
• classification = dis_v(concate_img)
• return Model([img, lable], classification)
• def build_cgan(genrator, discriminator):
• z = Input(shape=(zdim,))
• lable = Input(shape=(1,), dtype='int32')
• f_img = genrator([z, lable])
• classification = discriminator([f_img, lable])
• model = Model([z, lable], classification)
• return model
• dis_v = build_cdis(img_shape)
• dis_v.compile(loss='binary_crossentropy', optimizer=adam_v2.Adam(), metrics=['accuracy'])
• gen_v = build_cgen(zdim)
• dis_v.trainable = False
• gan_v = build_cgan(gen_v, dis_v)
• gan_v.compile(loss='binary_crossentropy', optimizer=adam_v2.Adam() )
• losses = []
• accuracies = []
• iteration_checks = []
• """ @Laurenz, please replace here with your code for data extraction (FHIR -> DataFrame) """
• ###############################################################################################
• # data = pd.read_csv('../input/data.csv')
• # data = data.drop(['Unnamed: 32', 'id'], axis=1)
• ###############################################################################################
• X_FEATURES = ['radius_mean', 'texture_mean', 'perimeter_mean', 'area_mean', 'smoothness_mean', 'compactness_mean', 'concavity_mean', 'concave.points_mean', 'symmetry_mean', 'fractal_dimension_mean', 'radius_se', 'texture_se', 'perimeter_se', 'area_se', 'smoothness_se', 'compactness_se', 'concavity_se', 'concave.points_se', 'symmetry_se', 'fractal_dimension_se', 'radius_worst', 'texture_worst', 'perimeter_worst', 'area_worst', 'smoothness_worst', 'compactness_worst', 'concavity_worst', 'concave.points_worst', 'symmetry_worst', 'fractal_dimension_worst']
• Y_FEATURE = 'label'
• client = SyncFHIRClient('http://{}:{}/fhir'.format(fhir_server, fhir_port))
• # Return lazy search set
• patients = client.resources('Patient')
• patients_data = []
• for patient in patients:
• patient_birthDate = None
• try:
• patient_birthDate = patient.birthDate
• except: pass
• patients_data.append([patient.id, patient.gender, patient_birthDate])
• patients_df = pd.DataFrame(patients_data, columns=["patient_id", "gender", "birthDate"])
• patients_observation = {}
• observations = client.resources("Observation").include("Patient", "subject")
• for observation in observations:
• try:
• feature = observation["category"][0]["coding"][0]["code"]
• if feature in X_FEATURES:
• value = observation["valueQuantity"]["value"]
• patient_id_str = observation["subject"]["reference"]
• if patient_id_str[:7] == "Patient":
• patient_id = patient_id_str[8:]
• if patient_id not in patients_observation:
• patients_observation[patient_id] = {}
• patients_observation[patient_id][feature] = float(value)
• except KeyError: print("Key error encountered, skipping Observation...")
• for k in patients_observation.keys():
• patients_observation[k].update(patient_id=k)
• observation_df = pd.DataFrame.from_dict(patients_observation.values())
• observation_df.set_index(["patient_id"])
• patients_condition = []
• conditions = client.resources("Condition")
• for condition in conditions:
• try:
• label = condition["code"]["coding"][0]["code"]
• patient_id_str = condition["subject"]["reference"]
• if patient_id_str[:7] == "Patient":
• patient_id = patient_id_str[8:]
• patients_condition.append([patient_id, label])
• except KeyError: print("Key error encountered, skipping Condition...")
• condition_df = pd.DataFrame(patients_condition, columns=["patient_id", "label"])
• data = pd.merge(pd.read_csv("./input/{}.csv".format(station_name)), pd.merge(pd.merge(patients_df, observation_df, on="patient_id", how="outer"), condition_df, on="patient_id", how="outer"), on="patient_id", how="inner")
• data['label'] = data['label'].map(lambda x: "B" if x == "non-cancer" else "M")
• print(data.head(5))
• ###############################################################################################
• ## splitting the station-data into train and test
• # splitting will be performed as 60% training data and 40% test data
• # create a data frame dictionary to store your data frames
• DataFrameDict = {elem: pd.DataFrame for elem in ['M', 'B']}
• for key in DataFrameDict.keys():
• DataFrameDict[key] = data[:][data.label == key]
• data_label_B = DataFrameDict['B']
• data_label_M = DataFrameDict['M']
• data_train_B, data_test_B = train_test_split(data_label_B, test_size=0.4)
• data_train_M, data_test_M = train_test_split(data_label_M, test_size=0.4)
• data_train = data_train_B.append(data_train_M)
• data_test = data_test_B.append(data_test_M)
• ##############################################################################################
• ## Synthetic data generation
• # for the Generation we use a cGAN-Algorithm with an intuitive structure
• # in the cGAN structure there is a Discriminator and a Generator
• # the Generator produce synthetic data based on the real world data
• # the Discriminator try to separate the real world data from the synthetic data
• # the output of both get the input of the other module
• # for this case we use a static value to break the cycle
• # the number of iteration, batch_size, interval and number of rows can be set
• # iteration is the number of cycles
• # batch size is the size of the data which go from Discriminator and Generator
• # intervall is show of cycle data
• # number of generated row can be set
• # formating the training data of the cGAN
• # label need to be 0 and 1 for better transistion
• if generation_synthetic_data:
• labels_training = data_train[Y_FEATURE]
• labels_training = labels_training.map(dict(M=1, B=0))
• features_training = data_train[X_FEATURES]
• # merge the data to one dataframe again for the training of cGAN
• train_data_synthetic = pd.concat([labels_training, features_training], axis=1, join='inner')
• # main method of the training of the cGAN as explained above
• synthetic_data = train_synthetic(data=train_data_synthetic, iterations=5000, batch_size=128, interval=1000, number_of_row=data.shape[1])
• # the random generation for the patient-id, gender and birthday
• # label needs to be transformed to the original values
• label_synthetic = []
• patient_id_synthetic = []
• patient_gender_synthetic = []
• patient_birthday_synthetic = []
• # very simple approach for the missing data generation, note: can be improved
• for row in range(0, len(synthetic_data)):
• synthetic_data_row = synthetic_data[row,]
• if (synthetic_data_row[0] == 1):
• label_synthetic.append("M")
• else:
• label_synthetic.append("B")
• patient_id_synthetic.append(("bbmri" + str(row)))
• p_g = "female"
• p_b = "01.01.2000"
• patient_birthday_synthetic.append(p_b)
• patient_gender_synthetic.append(p_g)
• # use the X_FEATURES of the synthetic data
• synthetic_data = synthetic_data[:, 1:31]
• # write everything in a dataframe for representation
• synthetic_df = pd.DataFrame(np.c_[ patient_id_synthetic, patient_gender_synthetic, patient_birthday_synthetic, synthetic_data, label_synthetic], columns=["patient_id", "gender", "birthDate", "radius_mean", "texture_mean", "perimeter_mean", "area_mean", "smoothness_mean", "compactness_mean", "concavity_mean", "concave.points_mean", "symmetry_mean", "fractal_dimension_mean", "radius_se", "texture_se", "perimeter_se", "area_se", "smoothness_se", "compactness_se", "concavity_se", "concave.points_se", "symmetry_se", "fractal_dimension_se", "radius_worst", "texture_worst", "perimeter_worst", "area_worst", "smoothness_worst", "compactness_worst", "concavity_worst", "concave.points_worst", "symmetry_worst", "fractal_dimension_worst", "label"])
• """ @Toralf, please add the code for enlarging the dataset size. """
• ###############################################################################################
• ## Exploratory data analysis (EDA) and data normalization
• ## training data will be in this example the synthetic data generated by the GAN-Algorithm
• ## test data will be the real data
• if generation_synthetic_data:
• X_train = data_train[X_FEATURES].append(synthetic_df[X_FEATURES].astype(str).astype(float))
• y_train = data_train[Y_FEATURE].append(synthetic_df[Y_FEATURE])
• else:
• X_train = data_train[X_FEATURES]
• y_train = data_train[Y_FEATURE]
• X_test = data_test[X_FEATURES]
• y_test = data_test[Y_FEATURE]
• ## Saving statistical data for visualization of train data
• X_train_mean = X_train.mean(axis=0)
• X_train_std = X_train.std(axis=0)
• print("Mean: ", X_train_mean)
• print("Std: ", X_train_std)
• B = (y_train == 'B').sum()
• M = (y_train == 'M').sum()
• print('Number of Benign: ', B)
• print('Number of Malignant : ', M)
• stat_dir = os.path.join(out_dir, 'stat', station_name)
• if not os.path.exists(stat_dir):
• os.makedirs(stat_dir)
• df_X_train_mean = pd.DataFrame([X_train_mean], columns=X_FEATURES)
• df_X_train_std = pd.DataFrame([X_train_std])
• df_y_train_dist = pd.DataFrame([[B, M]], columns=['B', 'M'])
• df_X_train_mean.to_csv(osp.join(stat_dir, 'X_train_mean.csv'))
• df_X_train_std.to_csv(osp.join(stat_dir, 'X_train_std.csv'))
• # changed the name of the file
• df_y_train_dist.to_csv(osp.join(stat_dir, 'Y_train_Dist.csv'))
• X_test_mean = X_test.mean(axis=0)
• X_test_std = X_test.std(axis=0)
• print("Mean: ", X_test_mean)
• print("Std: ", X_test_std)
• B = (y_test == 'B').sum()
• M = (y_test == 'M').sum()
• print('Number of Benign: ', B)
• print('Number of Malignant : ', M)
• stat_dir = os.path.join(out_dir, 'stat', station_name)
• if not os.path.exists(stat_dir):
• os.makedirs(stat_dir)
• df_X_test_mean = pd.DataFrame([X_test_mean], columns=X_FEATURES)
• df_X_test_std = pd.DataFrame([X_test_std])
• df_y_test_dist = pd.DataFrame([[B, M]], columns=['B', 'M'])
• df_X_test_mean.to_csv(osp.join(stat_dir, 'X_test_mean.csv'))
• df_X_test_std.to_csv(osp.join(stat_dir, 'X_test_std.csv'))
• # changed the name of the file
• df_y_test_dist.to_csv(osp.join(stat_dir, 'Y_test_Dist.csv'))
• # normalization of the data
• scaler = StandardScaler()
• X_train = scaler.fit_transform(X_train)
• y_train.replace(to_replace=dict(M=1, B=0), inplace=True)
• y_train = y_train.to_numpy()
• X_test = scaler.fit_transform(X_test)
• y_test.replace(to_replace=dict(M=1, B=0), inplace=True)
• y_test = y_test.to_numpy()
• ## logistic regression model
• model = LogisticRegression(input_dim, hidden_dim, num_classes)
• criterion = nn.CrossEntropyLoss()
• optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
• best_acc = 0.0
• model_dir = os.path.join(out_dir, 'models')
• if not os.path.exists(model_dir):
• os.makedirs(model_dir)
• torch.save({ 'epoch': -1, 'optim_state_dict': optimizer.state_dict(), 'model_state_dict': model.state_dict(), 'best_acc': 0.0, }, osp.join(model_dir, 'dnn.pth.tar'))
• else:
• checkpoint = torch.load(osp.join(model_dir, "dnn.pth.tar"))
• model.load_state_dict(checkpoint['model_state_dict'])
• optimizer.load_state_dict(checkpoint['optim_state_dict'])
• best_acc = checkpoint['best_acc']
• training_dir = osp.join(out_dir, 'training')
• if not os.path.exists(training_dir):
• os.makedirs(training_dir)
• if not osp.exists(osp.join(training_dir, 'log.csv')):
• with open(osp.join(training_dir, 'log.csv'), 'w') as f:
• f.write(','.join(["epoch", "station_name", "train_loss", "val_loss", "val_acc", "val_prec", "val_rec", "val_f1", "time"]) + '\n')
• if generation_synthetic_data:
• filename_output_log = station_name + "synthetic_logs.csv"
• else:
• filename_output_log = station_name + "logs.csv"
• for epoch in range(num_epochs):
• perm = np.arange(X_train.shape[0])
• np.random.shuffle(perm)
• X_train = X_train[perm]
• y_train = y_train[perm]
• loss = train(X_train, y_train, model, criterion, optimizer)
• # changed
• val_loss, val_acc, val_precesion, val_recall, val_f1_score, val_ppv = valid(X_test, y_test, model, criterion)
• with open(osp.join(training_dir, filename_output_log), 'a') as f:
• log = map(str, [epoch, station_name, loss, val_loss, val_acc, val_precesion, val_recall, val_f1_score, str(datetime.datetime.now(pytz.timezone('Europe/Berlin')))])
• f.write(','.join(log) + '\n')
• torch.save({ 'epoch': epoch, 'optim_state_dict': optimizer.state_dict(), 'model_state_dict': model.state_dict(), 'best_acc': val_acc, }, osp.join(model_dir, 'checkpoint.pth.tar'))
• if val_acc > best_acc:
• shutil.copy(osp.join(model_dir, 'checkpoint.pth.tar'), osp.join(model_dir, 'dnn.pth.tar'))

Graph