Author: Justin Tan
Adversarial training for robustness against systematic error.
In [1]:
import tensorflow as tf
import numpy as np
import pandas as pd
import pyarrow.parquet as pq
import glob, time, os
import selu
from diagnostics import *
class config(object):
# Set network parameters
mode = 'adv-selu'
channel = 'B2Xsy'
keep_prob = 0.95
num_epochs = 512
batch_size = 256
n_layers = 7
adv_n_layers = 3
adv_keep_prob = 0.98
hidden_layer_nodes = [1024, 1024, 512, 512, 512, 256, 256]
adv_hidden_nodes = [256,512,256]
n_gaussians = 5
ema_decay = 0.999
learning_rate = 2.5e-5
adv_learning_rate = 0.001
adv_lambda = 1.0
cycles = 3 # Number of annealing cycles
n_classes = 2
adv_n_classes = 10
epsilon = 1e-8
builder = 'selu'
K = 3
adversary = True
pivots = ['deltae', 'mbc']
class directories(object):
train = '/data/projects/punim0011/jtan/spark/spark2tf/cont_b2sy_By_val.parquet'# '/home/jtan/gpu/jtan/spark/spark2tf/example_train.tfrecords'
test = '/data/projects/punim0011/jtan/spark/spark2tf/cont_b2sy_By_test.parquet'#'/home/jtan/gpu/jtan/spark/spark2tf/example_test.tfrecords'
val = '/data/projects/punim0011/jtan/spark/spark2tf/cont_b2sy_By_train.parquet'#'/home/jtan/gpu/jtan/spark/spark2tf/example_test.tfrecords'
tensorboard = 'tensorboard'
checkpoints = 'checkpoints'
best_checkpoints = 'best'
architecture = '{} - {} | Layers: {} | Dropout: {} | Base LR: {} | Epochs: {}'.format(
config.channel, config.mode, config.n_layers, config.keep_prob, config.learning_rate, config.num_epochs)
get_available_gpus()
def load_parquet(datasetName, subsample=False, dropFrac=0.75):
from sklearn.model_selection import train_test_split
excludeFeatures = ['labels', 'mbc', 'deltae', 'daughterInvM', 'nCands', 'evtNum', 'MCtype', 'channel']
dataset = pq.ParquetDataset(datasetName)
pdf = dataset.read(nthreads=4).to_pandas()
pdf = pdf.sample(frac=1).reset_index(drop=True)
features = pdf.drop(excludeFeatures, axis=1)
labels = pdf['labels'].astype(np.int32)
auxillary = pdf[['deltae', 'mbc']]
auxillary=auxillary.assign(dE_labels=pd.qcut(auxillary['deltae'], q=config.adv_n_classes, labels=False),
mbc_labels=pd.qcut(auxillary['mbc'], q=config.adv_n_classes, labels=False))
return features.values.astype(np.float32), labels.values, auxillary.values.astype(np.float32), pdf
In [2]:
features, labels, auxillary, pdf = load_parquet(directories.train)
featuresTest, labelsTest, auxillaryTest, pdf_test = load_parquet(directories.test)
config.nTrainExamples, config.nFeatures = features.shape[0], features.shape[-1]
config.steps_per_epoch = features.shape[0] // config.batch_size
In [3]:
def dataset_placeholder(features_placeholder, labels_placeholder, batchSize, numEpochs, training=True):
dataset = tf.contrib.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder))
dataset = dataset.shuffle(buffer_size=16384)
dataset = dataset.batch(batchSize)
dataset = dataset.repeat(numEpochs) if training else dataset
return dataset
def dataset_placeholder_aux(features_placeholder, labels_placeholder, auxillary_placeholder, batchSize, numEpochs, training=True, shuffle=True):
dataset = tf.contrib.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder, auxillary_placeholder))
if shuffle:
dataset = dataset.shuffle(buffer_size=16384)
dataset = dataset.batch(batchSize)
dataset = dataset.repeat(numEpochs) if training else dataset
return dataset
def dataset_placeholder_plot(features_placeholder, labels_placeholder, auxillary_placeholder, batchSize=51200, training=False):
dataset = tf.contrib.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder, auxillary_placeholder))
dataset = dataset.batch(int(features_placeholder.get_shape()[0]))
dataset = dataset.repeat()
return dataset
def dataset_single_epoch(features, labels, auxillary, batchSize, training=True):
dataset = tf.contrib.data.Dataset.from_tensor_slices((features, labels, auxillary))
dataset = dataset.shuffle(buffer_size=25600)
dataset = dataset.batch(batchSize)
return dataset
def dataset_train(dataDirectory, batchSize, numEpochs, nFeatures, training=True):
filenames = glob.glob('{}/part*'.format(dataDirectory))
dataset = tf.contrib.data.TFRecordDataset(filenames)
# Extract data from `tf.Example` protocol buffer
def parser(record, batchSize=128):
keys_to_features = {
"features": tf.FixedLenFeature([nFeatures], tf.float32),
"labels": tf.FixedLenFeature((), tf.float32,
default_value=tf.zeros([], dtype=tf.float32)),
# "mbc": tf.FixedLenFeature((), tf.float32, dtype=tf.float32),
# "deltae": tf.FixedLenFeature((), tf.float32, dtype=tf.float32)
}
parsed = tf.parse_single_example(record, keys_to_features)
label = tf.cast(parsed['labels'], tf.int32)
return parsed['features'], label
# Transform into feature, label tensor pair
dataset = dataset.map(parser)
dataset = dataset.shuffle(buffer_size=16384)
dataset = dataset.batch(batchSize)
dataset = dataset.repeat(numEpochs) if training else dataset
return dataset
def dense_builder(x, shape, name, keep_prob, training=True, actv=tf.nn.elu):
init=tf.contrib.layers.xavier_initializer()
kwargs = {'center': True, 'scale': True, 'training': training, 'fused': True, 'renorm': True}
with tf.variable_scope(name, initializer=init) as scope:
layer = tf.layers.dense(x, units=shape[1], activation=actv)
bn = tf.layers.batch_normalization(layer, **kwargs)
layer_out = tf.layers.dropout(bn, 1-keep_prob, training=training)
return layer_out
def selu_builder(x, shape, name, keep_prob, training=True):
init = tf.contrib.layers.variance_scaling_initializer(factor=1.0, mode='FAN_IN')
with tf.variable_scope(name) as scope:
W = tf.get_variable("weights", shape = shape, initializer=init)
b = tf.get_variable("biases", shape = [shape[1]], initializer=tf.random_normal_initializer(stddev=0.01))
actv = selu.selu(tf.add(tf.matmul(x, W), b))
layer_output = selu.dropout_selu(actv, rate=1-keep_prob, training=training)
return layer_output
def dense_model(x, n_layers, hidden_layer_nodes, keep_prob, n_input, n_classes, scope, builder=selu_builder, reuse=False, training=True):
# Extensible dense model
SELU_initializer = tf.contrib.layers.variance_scaling_initializer(factor=1.0, mode='FAN_IN')
init = SELU_initializer if builder==selu_builder else tf.contrib.layers.xavier_initializer()
assert n_layers == len(hidden_layer_nodes), 'Specified layer nodes and number of layers do not correspond.'
layers = [x]
with tf.variable_scope(scope, reuse=reuse):
hidden_0 = builder(x, shape=[n_input, hidden_layer_nodes[0]], name='hidden0',
keep_prob = keep_prob, training=training)
layers.append(hidden_0)
for n in range(0,n_layers-1):
hidden_n = builder(layers[-1], shape=[hidden_layer_nodes[n], hidden_layer_nodes[n+1]], name='hidden{}'.format(n+1),
keep_prob=keep_prob, training=training)
layers.append(hidden_n)
readout = tf.layers.dense(hidden_n, units=n_classes, kernel_initializer=init)
return readout
def custom_elu(x):
return tf.nn.elu(x) + 1.0
def log_sum_exp_trick(x, axis=1):
x_max = tf.reduce_max(x, axis=1, keep_dims=True)
lse = x_max + tf.log(tf.reduce_sum(tf.exp(x-x_max), axis=1, keep_dims=True))
return lse
def scope_variables(name):
with tf.variable_scope(name):
return tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name)
def adversary_dense(x, z_pivot, n_layers, hidden_nodes, keep_prob, name, builder=selu_builder, reuse=False, training=True, actv=selu.selu):
# Extensible dense model
SELU_initializer = tf.contrib.layers.variance_scaling_initializer(factor=1.0, mode='FAN_IN')
init = SELU_initializer if builder==selu_builder else tf.contrib.layers.xavier_initializer()
assert n_layers == len(hidden_nodes), 'Specified layer nodes and number of layers do not correspond.'
layers = [x]
with tf.variable_scope('adversary', reuse=reuse):
hidden_0 = builder(x, shape=[2, hidden_nodes[0]], name='hidden0',
keep_prob=keep_prob, training=training)
layers.append(hidden_0)
for n in range(0,n_layers-1):
hidden_n = builder(layers[-1], shape=[hidden_nodes[n], hidden_nodes[n+1]], name='hidden{}'.format(n+1),
keep_prob=keep_prob, training=training)
layers.append(hidden_n)
fc = tf.layers.dense(hidden_n, units=96, activation=tf.nn.tanh, kernel_initializer=init)
fc_logits, fc_mu, fc_sigma = tf.split(fc, 3, axis=1)
logits = tf.layers.dense(fc_logits, units=config.n_gaussians, activation=tf.identity, name='mixing_fractions')
centers = tf.layers.dense(fc_mu, units=config.n_gaussians, activation=tf.identity, name='means')
variances = tf.layers.dense(fc_sigma, units=config.n_gaussians, activation=custom_elu, name='variances')
mixing_coeffs = tf.nn.softmax(logits)
exponent = tf.log(mixing_coeffs) - 1/2 * tf.log(2*np.pi) - tf.log(variances) - tf.square(centers-tf.expand_dims(z_pivot,1))/(2*tf.square(variances))
# adversary_out = tf.reduce_sum(mixing_coeffs*(1/(tf.sqrt(2*np.pi*variances))*tf.exp(-tf.square(centers-tf.expand_dims(z_pivot,1)) /(2*tf.square(variances))), axis=1)
return log_sum_exp_trick(exponent)
def selu_adversary(x, z_pivot, n_layers, hidden_nodes, keep_prob, name, reuse=False, training=True, actv=selu.selu):
SELU_initializer = tf.contrib.layers.variance_scaling_initializer(factor=1.0, mode='FAN_IN')
xavier = tf.contrib.layers.xavier_initializer(uniform=True, seed=None, dtype=tf.float32)
with tf.variable_scope('adversary', reuse=reuse):
l0 = tf.layers.dense(x, units=hidden_nodes[0], activation=actv, kernel_initializer=SELU_initializer)
d0 = selu.dropout_selu(l0, rate=1-keep_prob, training=training)
l1 = tf.layers.dense(d0, units=hidden_nodes[1], activation=actv, kernel_initializer=SELU_initializer)
d1 = selu.dropout_selu(l1, rate=1-keep_prob, training=training)
fc = tf.layers.dense(d1, units=96, activation=tf.nn.tanh, kernel_initializer=xavier)
fc_logits, fc_mu, fc_sigma = tf.split(fc, 3, axis=1)
logits = tf.layers.dense(fc_logits, units=config.n_gaussians, activation=tf.identity, name='mixing_fractions')
centers = tf.layers.dense(fc_mu, units=config.n_gaussians, activation=tf.identity, name='means')
variances = tf.layers.dense(fc_sigma, units=config.n_gaussians, activation=custom_elu, name='variances')
mixing_coeffs = tf.nn.softmax(logits)
exponent = tf.log(mixing_coeffs) - 1/2 * tf.log(2*np.pi) - tf.log(variances) - tf.square(centers-tf.expand_dims(z_pivot,1))/(2*tf.square(variances))
# adversary_out = tf.reduce_sum(mixing_coeffs*(1/(tf.sqrt(2*np.pi*variances))*tf.exp(-tf.square(centers-tf.expand_dims(z_pivot,1)) /(2*tf.square(variances))), axis=1)
return log_sum_exp_trick(exponent)
In [4]:
class vanillaDNN():
# Builds the computational graph
def __init__(self, config, training=True, cyclical=False):
self.global_step = tf.Variable(0, trainable=False)
self.handle = tf.placeholder(tf.string, shape=[])
self.training_phase = tf.placeholder(tf.bool)
self.beta = tf.placeholder(tf.float32) if cyclical else config.learning_rate
# self.beta = tf.train.exponential_decay(config.learning_rate, self.global_step,
# decay_steps = config.steps_per_epoch, decay_rate = config.lr_epoch_decay, staircase=True)
self.features_placeholder = tf.placeholder(features.dtype, features.shape)
self.labels_placeholder = tf.placeholder(labels.dtype, labels.shape)
self.auxillary_placeholder = tf.placeholder(auxillary.dtype, auxillary.shape)
self.featuresTest_placeholder = tf.placeholder(features.dtype, featuresTest.shape)
self.labelsTest_placeholder = tf.placeholder(labels.dtype, labelsTest.shape)
self.auxillaryTest_placeholder = tf.placeholder(auxillary.dtype, auxillaryTest.shape)
preTrainDataset = dataset_single_epoch(self.features_placeholder, self.labels_placeholder, self.auxillary_placeholder, config.batch_size, training=True)
trainDataset = dataset_placeholder_aux(self.features_placeholder, self.labels_placeholder, self.auxillary_placeholder,
config.batch_size, config.num_epochs, training=True)
testDataset = dataset_placeholder_aux(self.featuresTest_placeholder, self.labelsTest_placeholder, self.auxillaryTest_placeholder,
config.batch_size, config.num_epochs, training=True)
plotDataset = dataset_placeholder_plot(self.featuresTest_placeholder, self.labelsTest_placeholder, self.auxillaryTest_placeholder)
self.iterator = tf.contrib.data.Iterator.from_string_handle(self.handle, trainDataset.output_types, trainDataset.output_shapes)
self.pretrain_iterator = preTrainDataset.make_initializable_iterator()
self.train_iterator = trainDataset.make_initializable_iterator()
self.test_iterator = testDataset.make_initializable_iterator()
self.plot_iterator = plotDataset.make_initializable_iterator()
self.example, self.label, self.ancillary = self.iterator.get_next()
# self.readout = dense_SELU(self.example, config.n_layers, [1024, 1024, 512, 512, 256], config.keep_prob, training=self.training_phase)
self.readout = dense_model(self.example, config.n_layers, config.hidden_layer_nodes, config.keep_prob, builder=selu_builder,
n_input=config.nFeatures, n_classes=config.n_classes, reuse=False, training=self.training_phase, scope='predictor')
# Introduce a separate adversary network for each pivot
for i in range(len(config.pivots)):
adversary_readout = dense_model(tf.nn.softmax(self.readout), config.adv_n_layers, config.adv_hidden_nodes, config.adv_keep_prob,
n_input=2, n_classes=config.adv_n_classes, builder=selu_builder, training=self.training_phase, scope='adversary_{}'.format(config.pivots[i]))
adversary_loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=adversary_readout,
labels=tf.cast(self.ancillary[:,i+2], tf.int32)))
tf.add_to_collection('adversary_losses', adversary_loss)
# Mask signal values in adversary loss
self.adversary_loss = tf.add_n(tf.get_collection('adversary_losses'), name='total_adversary_loss')
self.predictor_loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.readout, labels=self.label))
self.total_loss = self.predictor_loss - config.adv_lambda*self.adversary_loss
theta_f = scope_variables('predictor')
theta_r = scope_variables('adversary')
self.theta = theta_f, theta_r
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# Ensures that we execute the update_ops before performing the train_step
predictor_optimizer = tf.train.AdamOptimizer(config.learning_rate)
predictor_gs = tf.Variable(0, name='predictor_global_step', trainable=False)
self.predictor_train_op = predictor_optimizer.minimize(self.predictor_loss, name='predictor_opt', global_step=predictor_gs, var_list=theta_f)
predictor_optimize = predictor_optimizer.minimize(self.total_loss, name='predictor_opt', global_step=predictor_gs, var_list=theta_f)
# self.joint_train_op = predictor_optimizer.minimize(self.total_loss, name='joint_opt', global_step=predictor_gs, var_list=theta_f)
adversary_optimizer = tf.train.AdamOptimizer(config.adv_learning_rate)
adversary_gs = tf.Variable(0, name='adversary_global_step', trainable=False)
self.adversary_train_op = adversary_optimizer.minimize(self.adversary_loss, name='adversary_opt', global_step=adversary_gs, var_list=theta_r)
self.ema = tf.train.ExponentialMovingAverage(decay=config.ema_decay, num_updates=predictor_gs, name='predictor_ema')
maintain_predictor_averages_op = self.ema.apply(theta_f)
with tf.control_dependencies([predictor_optimize]):
self.joint_train_op = tf.group(maintain_predictor_averages_op)
# Evaluation metrics
self.cross_entropy = self.predictor_loss
self.p = tf.nn.softmax(self.readout)
self.transform = tf.log(self.p[:,1]/(1-self.p[:,1]+config.epsilon)+config.epsilon)
correct_prediction = tf.equal(tf.cast(tf.argmax(self.readout, 1), tf.int32), self.label)
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
_, self.auc_op = tf.metrics.auc(predictions = tf.argmax(self.readout,1), labels = self.label, num_thresholds = 1024)
self.pearson_dE, self.pearson_dE_op = tf.contrib.metrics.streaming_pearson_correlation(predictions=self.transform,
labels=self.ancillary[:,0], name='pearson_dE')
self.pearson_mbc, self.pearson_mbc_op = tf.contrib.metrics.streaming_pearson_correlation(predictions=self.transform,
labels=self.ancillary[:,1], name='pearson_mbc')
self.adv_xentropy = self.adversary_loss
adv_correct_prediction = tf.equal(tf.cast(tf.argmax(adversary_readout,1), tf.int32), tf.cast(self.ancillary[:,3], tf.int32))
self.adv_accuracy = tf.reduce_mean(tf.cast(adv_correct_prediction, tf.float32))
tf.summary.scalar('accuracy', self.accuracy)
tf.summary.scalar('auc', self.auc_op)
tf.summary.scalar('learning_rate', self.beta)
tf.summary.scalar('predictor_loss', self.predictor_loss)
tf.summary.scalar('adversary_loss', self.adversary_loss)
tf.summary.scalar('total_loss', self.total_loss)
tf.summary.scalar('pearson_dE', self.pearson_dE_op)
tf.summary.scalar('pearson_mbc', self.pearson_mbc_op)
self.merge_op = tf.summary.merge_all()
self.train_writer = tf.summary.FileWriter(
os.path.join(directories.tensorboard, 'train_{}'.format(time.strftime('%d-%m_%I:%M'))), graph = tf.get_default_graph())
self.test_writer = tf.summary.FileWriter(
os.path.join(directories.tensorboard, 'test_{}'.format(time.strftime('%d-%m_%I:%M'))))
def predict(self, ckpt):
pin_cpu = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True, device_count = {'GPU':0})
start_time = time.time()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = self.ema.variables_to_restore().update(dict((v, v) for v in scope_variables('adversary')))
saver = tf.train.Saver(variables_to_restore)
# saver = tf.train.Saver()
valDataset = dataset_placeholder_aux(self.featuresTest_placeholder, self.labelsTest_placeholder, self.auxillaryTest_placeholder,
config.batch_size, config.num_epochs, training=False, shuffle=False)
val_iterator = valDataset.make_initializable_iterator()
concatLabels = tf.cast(self.label, tf.int32)
concatPreds = tf.cast(tf.argmax(self.readout,1), tf.int32)
concatOutput = self.p[:,1]
with tf.Session(config=pin_cpu) as sess:
# Initialize variables
init_op = tf.global_variables_initializer()
sess.run(init_op)
sess.run(tf.local_variables_initializer())
assert (ckpt.model_checkpoint_path), 'Missing checkpoint file!'
saver.restore(sess, ckpt.model_checkpoint_path)
print('{} restored.'.format(ckpt.model_checkpoint_path))
val_handle = sess.run(val_iterator.string_handle())
labels, preds, outputs = [], [], []
sess.run(val_iterator.initializer, feed_dict={vDNN.featuresTest_placeholder: featuresTest,
vDNN.labelsTest_placeholder: labelsTest,
vDNN.auxillaryTest_placeholder: auxillaryTest})
while True:
try:
l, p, o = sess.run([concatLabels, concatPreds, concatOutput],
feed_dict = {vDNN.training_phase: False, vDNN.handle: val_handle})
labels.append(l), preds.append(p), outputs.append(o)
except tf.errors.OutOfRangeError:
labels, preds, outputs = np.concatenate(labels), np.concatenate(preds), np.concatenate(outputs)
break
acc = np.mean(np.equal(labels,preds))
print("Validation accuracy: {:.3f}".format(acc))
plot_ROC_curve(network_output=outputs, y_true=labels, identifier=config.mode+config.channel,
meta=architecture + ' | Test accuracy: {:.3f}'.format(acc))
delta_t = time.time() - start_time
print("Inference complete. Duration: %g s" %(delta_t))
return labels, preds, outputs
In [ ]:
def train(config, restore = False):
# Executes training operations
print('Architecture: {}'.format(architecture))
vDNN = vanillaDNN(config, training=True)
start_time = time.time()
global_step, v_auc_best = 0, 0.
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(directories.checkpoints)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)) as sess:
# Initialize variables
init_op = tf.global_variables_initializer()
sess.run(init_op)
pretrain_handle = sess.run(vDNN.pretrain_iterator.string_handle())
train_handle = sess.run(vDNN.train_iterator.string_handle())
test_handle = sess.run(vDNN.test_iterator.string_handle())
plot_handle = sess.run(vDNN.plot_iterator.string_handle())
if restore and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('{} restored.'.format(ckpt.model_checkpoint_path))
train_init = {vDNN.features_placeholder: features, vDNN.labels_placeholder: labels, vDNN.auxillary_placeholder: auxillary}
test_init = {vDNN.featuresTest_placeholder: featuresTest, vDNN.labelsTest_placeholder: labelsTest, vDNN.auxillaryTest_placeholder: auxillaryTest}
pretrain_feed = {vDNN.training_phase: True, vDNN.handle: pretrain_handle}
train_feed = {vDNN.training_phase: True, vDNN.handle: train_handle}
sess.run(vDNN.train_iterator.initializer, feed_dict=train_init)
sess.run(vDNN.test_iterator.initializer, feed_dict=test_init)
sess.run(vDNN.plot_iterator.initializer, feed_dict=test_init)
def pretrain(iterator, train_op, feed_dict_init, feed_dict_train, n_epochs):
for epoch in range(n_epochs):
sess.run(iterator.initializer, feed_dict=feed_dict_init)
while True:
try:
sess.run(train_op, feed_dict=feed_dict_train)
except tf.errors.OutOfRangeError:
break
plot_distributions(vDNN, epoch, sess, handle=plot_handle)
run_diagnostics(vDNN, config, directories, sess, saver, train_handle, test_handle, global_step,
config.nTrainExamples, start_time, v_auc_best, epoch)
# pretrain(vDNN.pretrain_iterator, train_op=vDNN.predictor_train_op, feed_dict_init=train_init, feed_dict_train=pretrain_feed, n_epochs=5)
# pretrain(vDNN.pretrain_iterator, train_op=vDNN.adversary_train_op, feed_dict_init=train_init, feed_dict_train=pretrain_feed, n_epochs=1)
while True:
try:
if config.adversary:
# adversary trains in inner loop
if global_step % config.K == 0:
sess.run(vDNN.joint_train_op, feed_dict=train_feed)
else:
sess.run(vDNN.adversary_train_op, feed_dict=train_feed)
global_step+=1
if global_step % (config.steps_per_epoch) == 0:
epoch, v_auc_best = run_adv_diagnostics(vDNN, config, directories, sess, saver, train_handle, test_handle,
global_step, config.nTrainExamples, start_time, v_auc_best)
plot_distributions(vDNN, epoch, sess, handle=plot_handle)
else:
# Run X steps on training dataset
sess.run(vDNN.predictor_train_op, feed_dict=train_feed)
global_step+=1
if global_step % (config.steps_per_epoch // 4) == 0:
epoch, v_auc_best = run_diagnostics(vDNN, config, directories, sess, saver, train_handle, test_handle,
global_step, config.nTrainExamples, start_time, v_auc_best)
except tf.errors.OutOfRangeError:
break
save_path = saver.save(sess, os.path.join(directories.checkpoints, 'vDNN_{}_{}_end.ckpt'.format(config.mode, config.channel)), global_step=epoch)
print("Training Complete. Model saved to file: {} Time elapsed: {:.3f} s".format(save_path, time.time()-start_time))
In [ ]:
train(config)#, restore=True)
In [6]:
import os
ckpt = tf.train.get_checkpoint_state('best')
vDNN = vanillaDNN(config, training = False)
labels, preds, output = vDNN.predict(ckpt)
# Add predictions to test set as a new column, save as HDF5
output = pd.Series(output, name='preds')
test = pd.concat([pdf_test, output], axis=1)
In [11]:
bkg = test[test['labels']<0.5]
sig = test[test['labels']==1]
In [16]:
post_cut = bkg[bkg['probs']>0.9]
post_sig = sig[sig['probs']>0.9]
In [19]:
import matplotlib.pyplot as plt
import seaborn as sns
sns.distplot(sig['probs'], kde=False)
plt.show()
In [14]:
test.to_hdf(os.path.join(directories.checkpoints, 'adv_both{}_preds.h5'.format(os.path.basename(directories.test))), key = 'df', format='t', data_columns=True)
In [11]:
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
plt.style.use('seaborn-darkgrid')
plt.style.use('seaborn-talk')
plt.style.use('seaborn-pastel')
In [8]:
def normPlot(variable, pdf, epoch, signal, nbins=50, bkg_rejection=0.95):
from scipy import stats
titles={'mbc': r'$M_{bc}$ (GeV)', 'deltae': r'$\Delta E$ (GeV)', 'daughterInvM': r'$M_{X_q}$ (GeV)'}
bkg = pdf[pdf['labels']<0.5]
post_bkg = bkg.nlargest(int(bkg.shape[0]*(1-bkg_rejection)), columns=['preds'])
threshold = post_bkg['preds'].min()
print('Post-selection:', post_bkg.shape[0])
if signal:
sig = pdf[pdf['labels']==1]
post_sig = pdf[(pdf['labels']==1) & (pdf['preds']>threshold)]
sns.distplot(post_sig[variable], hist=True, kde=True, label='Signal - {} BG rejection'.format(bkg_rejection), bins=nbins,
hist_kws=dict(edgecolor="0.85", linewidth=0.5, alpha = 0.8))
sns.distplot(sig[variable], hist=True, kde=True, label='Signal', bins=nbins,
hist_kws=dict(edgecolor="0.85", linewidth=0.5, alpha = 0.8))
else:
sns.distplot(post_bkg[variable], hist=True, kde=False, norm_hist=True, label='Background - {} BG rejection'.format(bkg_rejection), bins=nbins,# fit=stats.norm,
hist_kws=dict(edgecolor="0.85", linewidth=0.5, alpha = 0.8))
sns.distplot(bkg[variable], hist=True, kde=True, label='Background', bins=nbins,
hist_kws=dict(edgecolor="0.85", linewidth=0.5, alpha = 0.8))
plt.xlabel(r'{}'.format(titles[variable]))
plt.ylabel(r'Normalized events/bin')
plt.legend(loc = "best")
plt.savefig('best/{}_adv-ep{}-nb.pdf'.format(variable, epoch), bbox_inches='tight',format='pdf', dpi=1000)
plt.show()
plt.gcf().clear()
In [13]:
normPlot('deltae', test, signal=False, epoch='talk_bkg_mbc_adv-nn')
In [15]:
normPlot('mbc', test, signal=False, epoch='talk_bkg_mbc_adv-nn')
In [ ]: