# -*- coding: utf-8 -*-
"""
@author: Laura C. Kühle, Soraya Terrab (sorayaterrab)
Code-Style: E226, W503
Docstring-Style: D200, D400
TODO: Test new ANN set-up with Soraya
TODO: Remove object set-up (for more flexibility) -> Done
(decided against it to keep easy test set-up)
TODO: Add documentation
TODO: Allow multiple approximations in one config -> Done
TODO: Split workflow into multiple modules -> Done
TODO: Remove unnecessary instance variables -> Done
TODO: Add option to change 'num_iterations' for model testing -> Done
TODO: Save model evaluation data -> Done
TODO: Add README for ANN training
TODO: Fix random seed
TODO: Discuss whether to outsource scripts into separate directory
TODO: Discuss whether comparison between datasets is wanted
TODO: Discuss whether default model state is useful
"""
import numpy as np
import time
import matplotlib
from matplotlib import pyplot as plt
import os
import torch
import json
from torch.utils.data import TensorDataset, DataLoader, random_split
from sklearn.model_selection import KFold
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, roc_auc_score
import ANN_Model
from Plotting import plot_classification_accuracy, plot_boxplot
matplotlib.use('Agg')
class ModelTrainer(object):
def __init__(self, config):
self._reset(config)
def _reset(self, config):
self._batch_size = config.pop('batch_size', 500)
self._num_epochs = config.pop('num_epochs', 1000)
self._threshold = config.pop('threshold', 1e-5)
model = config.pop('model', 'ThreeLayerReLu')
model_config = config.pop('model_config', {})
loss_function = config.pop('loss_function', 'BCELoss')
loss_config = config.pop('loss_config', {})
optimizer = config.pop('optimizer', 'Adam')
optimizer_config = config.pop('optimizer_config', {})
# Set learning rate
learning_rate = config.pop('learning_rate', 1e-2)
optimizer_config['lr'] = learning_rate
if not hasattr(ANN_Model, model):
raise ValueError('Invalid model: "%s"' % model)
if not hasattr(torch.nn.modules.loss, loss_function):
raise ValueError('Invalid loss function: "%s"' % loss_function)
if not hasattr(torch.optim, optimizer):
raise ValueError('Invalid optimizer: "%s"' % optimizer)
self._model = getattr(ANN_Model, model)(model_config)
self._loss_function = getattr(torch.nn.modules.loss, loss_function)(
**loss_config)
self._optimizer = getattr(torch.optim, optimizer)(
self._model.parameters(), **optimizer_config)
self._validation_loss = torch.zeros(self._num_epochs//10)
def epoch_training(self, dataset, num_epochs=None, verbose=True):
tic = time.perf_counter()
if num_epochs is None:
num_epochs = self._num_epochs
# Split data into training and validation set
num_samples = len(dataset)
if verbose:
print('Splitting data randomly into training and validation set.')
train_ds, valid_ds = random_split(dataset, [round(num_samples*0.8), round(num_samples*0.2)])
# Load sets
train_dl = DataLoader(train_ds, batch_size=self._batch_size, shuffle=True)
valid_dl = DataLoader(valid_ds, batch_size=self._batch_size * 2)
# Training with Validation
if verbose:
print('\nTraining model...')
print('Number of epochs:', num_epochs)
tic_train = time.perf_counter()
for epoch in range(num_epochs):
self._model.train()
for x_batch, y_batch in train_dl:
pred = self._model(x_batch.float())
loss = self._loss_function(pred, y_batch.float()).mean()
# Run back propagation, update the weights, and zero gradients for next epoch
loss.backward()
self._optimizer.step()
self._optimizer.zero_grad()
self._model.eval()
with torch.no_grad():
valid_loss = sum(
self._loss_function(self._model(x_batch_valid.float()), y_batch_valid.float())
for x_batch_valid, y_batch_valid in valid_dl)
if (epoch+1) % 100 == 0:
self._validation_loss[int((epoch+1) / 100)-1] = valid_loss / len(valid_dl)
if verbose:
print(epoch+1, 'epochs completed. Loss:', valid_loss / len(valid_dl))
if valid_loss / len(valid_dl) < self._threshold:
break
toc_train = time.perf_counter()
if verbose:
print('Finished training model!')
print(f'Training time: {toc_train-tic_train:0.4f}s\n')
toc = time.perf_counter()
if verbose:
print(f'Total runtime: {toc-tic:0.4f}s\n')
def test_model(self, training_set, test_set):
self.epoch_training(training_set, num_epochs=100, verbose=False)
self._model.eval()
x_test, y_test = test_set
model_score = self._model(x_test.float())
model_output = torch.argmax(model_score, dim=1)
y_true = y_test.detach().numpy()[:, 1]
y_pred = model_output.detach().numpy()
accuracy = accuracy_score(y_true, y_pred)
precision, recall, f_score, support = precision_recall_fscore_support(y_true, y_pred,
zero_division=0)
auroc = roc_auc_score(y_true, y_pred)
return {'Precision_Smooth': precision[0], 'Precision_Troubled': precision[1],
'Recall_Smooth': recall[0], 'Recall_Troubled': recall[1],
'F-Score_Smooth': f_score[0], 'F-Score_Troubled': f_score[1],
'Accuracy': accuracy, 'AUROC': auroc}
def save_model(self, directory, model_name='test_model'):
# Set paths for files if not existing already
model_dir = directory + '/trained models'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
# Save model and loss
torch.save(self._model.state_dict(), model_dir + '/model__' + model_name + '.pt')
torch.save(self._validation_loss, model_dir + '/loss__' + model_name + '.pt')
def read_training_data(directory, normalized=True):
# Get training dataset from saved file and map to Torch tensor and dataset
input_file = directory + ('/normalized_input_data.npy' if normalized else '/input_data.npy')
output_file = directory + '/output_data.npy'
return TensorDataset(*map(torch.tensor, (np.load(input_file), np.load(output_file))))
def evaluate_models(models, directory, num_iterations=100, colors=None,
compare_normalization=False):
tic = time.perf_counter()
if colors is None:
colors = {'Accuracy': 'magenta', 'Precision_Smooth': 'red',
'Precision_Troubled': '#8B0000', 'Recall_Smooth': 'blue',
'Recall_Troubled': '#00008B', 'F-Score_Smooth': 'green',
'F-Score_Troubled': '#006400', 'AUROC': 'yellow'}
print('Read normalized training data.')
datasets = {'normalized': read_training_data(directory)}
if compare_normalization:
print('Read raw, non-normalized training data.')
datasets['raw'] = read_training_data(directory, False)
classification_stats = {measure: {model + ' (' + dataset + ')': [] for model in models
for dataset in datasets} for measure in colors}
print('\nTraining models with 5-fold cross validation...')
print('Number of iterations:', num_iterations)
tic_train = time.perf_counter()
for iteration in range(num_iterations):
for train_index, test_index in KFold(
n_splits=5, shuffle=True).split(datasets['normalized']):
for dataset in datasets.keys():
training_set = TensorDataset(*datasets[dataset][train_index])
test_set = datasets[dataset][test_index]
for model in models:
result = models[model].test_model(training_set, test_set)
for measure in colors:
classification_stats[measure][model + ' (' + dataset + ')'].append(
result[measure])
if iteration+1%max(10, 10*(num_iterations//100)):
print(iteration+1, 'iterations completed.')
toc_train = time.perf_counter()
print('Finished training models with 5-fold cross validation!')
print(f'Training time: {toc_train - tic_train:0.4f}s\n')
with open(directory + '/' + '_'.join(models.keys()) + '.json', 'w') as json_file:
json_file.write(json.dumps(classification_stats))
with open(directory + '/' + '_'.join(models.keys()) + '.json') as json_file:
classification_stats = json.load(json_file)
print('Plotting evaluation of trained models.')
plot_boxplot(classification_stats, colors)
classification_stats = {measure: {model + ' (' + dataset + ')': np.array(
classification_stats[measure][model + ' (' + dataset + ')']).mean() for model in models
for dataset in datasets} for measure in colors}
plot_classification_accuracy(classification_stats, colors)
# Set paths for plot files if not existing already
plot_dir = directory + '/model evaluation'
if not os.path.exists(plot_dir):
os.makedirs(plot_dir)
# Save plots
print('Saving plots.')
for identifier in plt.get_figlabels():
# Set path for figure directory if not existing already
if not os.path.exists(plot_dir + '/' + identifier):
os.makedirs(plot_dir + '/' + identifier)
plt.figure(identifier)
plt.savefig(plot_dir + '/' + identifier + '/' + '_'.join(models.keys()) + '.pdf')
toc = time.perf_counter()
print(f'Total runtime: {toc - tic:0.4f}s')