import tensorflow as tf
import math
from sklearn.metrics import roc_auc_score
import numpy as np
EPOCHS = 5
BATCH_SIZE = 2048
def train_test_model_demo(model, train_label_path, train_idx_path, train_value_path,
test_label_path, test_idx_path, test_value_path):
train_batch_dataset = get_batch_dataset(train_label_path, train_idx_path, train_value_path)
test_batch_dataset = get_batch_dataset(test_label_path, test_idx_path, test_value_path)
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-4)
for epoch in range(5):
train_model(model, train_batch_dataset, optimizer, epoch)
test_model(model, test_batch_dataset)
def get_batch_dataset(label_path, idx_path, value_path):
label = tf.data.TextLineDataset(label_path)
idx = tf.data.TextLineDataset(idx_path)
value = tf.data.TextLineDataset(value_path)
label = label.map(lambda x: tf.strings.to_number(tf.strings.split(x, sep='\t')), num_parallel_calls=12)
idx = idx.map(lambda x: tf.strings.to_number(tf.strings.split(x, sep='\t')), num_parallel_calls=12)
value = value.map(lambda x: tf.strings.to_number(tf.strings.split(x, sep='\t')), num_parallel_calls=12)
batch_dataset = tf.data.Dataset.zip((label, idx, value))
batch_dataset = batch_dataset.shuffle(buffer_size=20480)
batch_dataset = batch_dataset.batch(BATCH_SIZE)
batch_dataset = batch_dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
return batch_dataset
""" ************************************************************************************ """
""" Using Criteo DataSet to train/test Model """
""" ************************************************************************************ """
@tf.function
def cross_entropy_loss(y_true, y_pred):
return tf.reduce_mean(tf.losses.binary_crossentropy(y_true, y_pred))
@tf.function
def train_one_step(model, optimizer, idx, value, label):
with tf.GradientTape() as tape:
output = model(idx, value)
loss = cross_entropy_loss(y_true=label, y_pred=output)
reg_loss = []
for p in model.trainable_variables:
reg_loss.append(tf.nn.l2_loss(p))
reg_loss = tf.reduce_sum(tf.stack(reg_loss))
loss = loss + model.reg_l2 * reg_loss
grads = tape.gradient(loss, model.trainable_variables)
grads = [tf.clip_by_norm(g, 100) for g in grads]
optimizer.apply_gradients(grads_and_vars=zip(grads, model.trainable_variables))
return loss
def train_model(model, train_batch_dataset, optimizer, epoch):
train_item_count = 41120555
for batch_idx, (label, idx, value) in enumerate(train_batch_dataset):
if len(label) == 0:
break
loss = train_one_step(model, optimizer, idx, value, label)
if batch_idx % 1000 == 0:
print('Train Epoch: {} [{} / {} ({:.2f}%)]\tLoss:{:.6f}'.format(
epoch, batch_idx * len(idx), train_item_count,
100. * batch_idx / math.ceil(int(train_item_count / BATCH_SIZE)), loss.numpy()))
def test_model(model, test_batch_dataset):
pred_y, true_y = [], []
binaryloss = tf.keras.metrics.BinaryCrossentropy()
for batch_idx, (label, idx, value) in enumerate(test_batch_dataset):
if len(label) == 0:
break
output = model(idx, value, use_dropout=False)
binaryloss.update_state(y_true=label, y_pred=output)
pred_y.extend(list(output.numpy()))
true_y.extend(list(label.numpy()))
print('Roc AUC: %.5f' % roc_auc_score(y_true=np.array(true_y), y_score=np.array(pred_y)))
print('LogLoss: %.5f' % binaryloss.result())
# import tensorflow as tf
# import math
# from sklearn.metrics import roc_auc_score
# import numpy as np
#
# EPOCHS = 5
# BATCH_SIZE = 2048
#
#
# def get_batch_dataset(label_path, idx_path, value_path):
# label = tf.data.TextLineDataset(label_path)
# idx = tf.data.TextLineDataset(idx_path)
# value = tf.data.TextLineDataset(value_path)
#
# label = label.map(lambda x: tf.strings.to_number(tf.strings.split(x, sep='\t')), num_parallel_calls=12)
# idx = idx.map(lambda x: tf.strings.to_number(tf.strings.split(x, sep='\t')), num_parallel_calls=12)
# value = value.map(lambda x: tf.strings.to_number(tf.strings.split(x, sep='\t')), num_parallel_calls=12)
#
# batch_dataset = tf.data.Dataset.zip((label, idx, value))
# batch_dataset = batch_dataset.shuffle(buffer_size=20480)
# batch_dataset = batch_dataset.batch(BATCH_SIZE)
# batch_dataset = batch_dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
# return batch_dataset
#
#
# def train_test_model_demo(model, train_label_path, train_idx_path, train_value_path, test_label_path, test_idx_path, test_value_path):
# train_batch_dataset = get_batch_dataset(train_label_path, train_idx_path, train_value_path)
# test_batch_dataset = get_batch_dataset(test_label_path, test_idx_path, test_value_path)
#
# optimizer = tf.keras.optimizers.Adam(learning_rate=1e-4)
#
# for epoch in range(5):
# train_model(model, train_batch_dataset, optimizer, epoch)
# test_model(model, test_batch_dataset)
#
#
# """ ************************************************************************************ """
# """ Using Criteo DataSet to train/test Model """
# """ ************************************************************************************ """
# def train_model(model, train_batch_dataset, optimizer, epoch):
# binaryloss = tf.keras.metrics.BinaryCrossentropy()
# train_item_count = 41120555
# for batch_idx, (label, idx, value) in enumerate(train_batch_dataset):
# if len(label) == 0:
# break
#
# with tf.GradientTape() as tape:
# output = model(idx, value)
# loss = tf.keras.losses.binary_crossentropy(y_true=label, y_pred=output)
# binaryloss.update_state(y_true=label, y_pred=output)
#
# grads = tape.gradient(loss, model.trainable_variables)
# grads = [tf.clip_by_norm(g, 100) for g in grads]
# optimizer.apply_gradients(grads_and_vars=zip(grads, model.trainable_variables))
# if batch_idx % 1000 == 0:
# print('Train Epoch: {} [{} / {} ({:.2f}%)]\tLoss:{:.6f}'.format(epoch, batch_idx * len(idx), train_item_count,
# 100. * batch_idx / math.ceil(int(train_item_count / BATCH_SIZE)), binaryloss.result()))
#
#
# def test_model(model, test_batch_dataset):
# pred_y, true_y = [], []
# binaryloss = tf.keras.metrics.BinaryCrossentropy()
# for batch_idx, (label, idx, value) in enumerate(test_batch_dataset):
# if len(label) == 0:
# break
#
# output = model(idx, value)
# binaryloss.update_state(y_true=label, y_pred=output)
# pred_y.extend(list(output.numpy()))
# true_y.extend(list(label.numpy()))
# print('Roc AUC: %.5f' % roc_auc_score(y_true=np.array(true_y), y_score=np.array(pred_y)))
# print('LogLoss: %.5f' % binaryloss.result())
