Created
January 2, 2020 08:22
-
-
Save VyBui/2e300e57b2860f3ca0360d341c39fefc to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
from __future__ import absolute_import, division, print_function, unicode_literals | |
import logging.config | |
import tensorflow as tf | |
import os | |
import numpy as np | |
from tensorflow import ConfigProto | |
from argument_parse import args | |
from module.discriminator_vgg19 import Discriminator | |
from module.generator import Generator | |
from module.losses import l1_loss | |
from config import cfg | |
from data_tools.parse_records_dataset import input_fn | |
from calculate_average_gradients import get_perturbed_batch, average_gradients | |
from preprocessing.dataset import pre_processing | |
from utils import shuffle | |
if args.mode in ['train', 'test', 'val']: | |
params = {'batch_size': cfg.train_batch_size, | |
'tfrecords_path': cfg.tfrecords_path} | |
train_dataset = input_fn(args.mode, params) | |
else: | |
raise ValueError("mode must be via ( train, test or val).") | |
if not any([isinstance(args.num_gpus, int), isinstance(args.batch_size, int)]): | |
raise ValueError("num gpus or batch size must be type integer.") | |
if args.mode in ['train', 'test', 'val']: | |
params = {'batch_size': cfg.train_batch_size, | |
'tfrecords_path': cfg.tfrecords_path} | |
train_dataset = input_fn(args.mode, params) | |
# get TF logger | |
# load logging confoguration and create log object | |
logging.config.fileConfig('logging.conf') | |
logging.basicConfig(filename='skin_generator.log', level=logging.DEBUG) | |
log = logging.getLogger('TensorFlow') | |
log.setLevel(logging.DEBUG) | |
# create formatter and add it to the handlers | |
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') | |
# create file handler which logs even debug messages | |
fh = logging.FileHandler('module_2.log') | |
fh.setLevel(logging.DEBUG) | |
fh.setFormatter(formatter) | |
log.addHandler(fh) | |
if __name__ == '__main__': | |
train_iterator = train_dataset.make_initializable_iterator() | |
batch_data = train_iterator.get_next() | |
image_label, body_parts, seg_parts, top_and_bottom = batch_data | |
train_batch_size_step = args.batch_size // args.num_gpus | |
# output of D for real images | |
D_real, D_real_logits = Discriminator(image_label).feed_forward() | |
# output of D for fake images | |
gen, end_points = Generator(body_parts, seg_parts, top_and_bottom).feed_forward() | |
D_fake, D_fake_logits = Discriminator(gen).feed_forward() | |
label_input_perturbed = get_perturbed_batch = get_perturbed_batch(image_label) | |
# get loss for discriminator | |
with tf.name_scope('D_loss'): | |
d_loss_real = tf.reduce_mean( | |
tf.nn.sigmoid_cross_entropy_with_logits(logits=D_real_logits, labels=tf.ones_like(D_real))) | |
d_loss_fake = tf.reduce_mean( | |
tf.nn.sigmoid_cross_entropy_with_logits(logits=D_fake_logits, labels=tf.zeros_like(D_fake))) | |
d_loss = d_loss_real + d_loss_fake | |
alpha = tf.random_uniform(shape=tf.shape(image_label), minval=0., maxval=1.) | |
differences = label_input_perturbed - image_label # This is different from WGAN-GP | |
interpolates = image_label + (alpha * differences) | |
_, D_inter = Discriminator(interpolates).feed_forward() | |
gradients = tf.gradients(D_inter, [interpolates])[0] | |
slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1])) | |
gradient_penalty = tf.reduce_mean((slopes - 1.) ** 2) | |
lambd = 0.1 | |
d_loss += lambd * gradient_penalty | |
# get loss for generator | |
with tf.name_scope('G_loss'): | |
g_mse_lambda = 100 | |
g_mse_loss = tf.keras.losses.MSE(y_true=image_label, y_pred=gen) | |
g_mse_loss = g_mse_loss * g_mse_lambda | |
gen_loss = g_loss = tf.reduce_mean( | |
tf.nn.sigmoid_cross_entropy_with_logits(logits=D_fake_logits, labels=tf.ones_like(D_fake))) + g_mse_loss | |
# Training: divide trainable variables into a group for D and a group for G | |
t_vars = tf.trainable_variables() | |
d_vars = [var for var in t_vars if "discriminator" in var.name] | |
g_vars = [var for var in t_vars if "generator" in var.name] | |
# Optimizers | |
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)): | |
d_train_opt = tf.train.AdamOptimizer(learning_rate=cfg.lr, beta1=0.5).minimize(d_loss, var_list=d_vars) | |
g_train_opt = tf.train.AdamOptimizer(learning_rate=cfg.lr, beta1=0.5).minimize(g_loss, var_list=g_vars) | |
top = top_and_bottom[:, :, :, 0:3] | |
# Summary | |
gen__image_sum = tf.summary.image("fake", gen[:, :, :, ::-1], max_outputs=1) | |
real_image_sum = tf.summary.image("real", image_label[:, :, :, ::-1], max_outputs=1) | |
top_sum = tf.summary.image("top", top[:, :, :, ::-1], max_outputs=1) | |
d_loss_real_sum = tf.summary.scalar("d_loss_real", d_loss_real, family="D_loss") | |
d_loss_fake_sum = tf.summary.scalar("d_loss_fake", d_loss_fake, family="D_loss") | |
d_loss_sum = tf.summary.scalar("d_loss", d_loss, family="D_loss") | |
g_loss_l1_sum = tf.summary.scalar("g_mse_loss", g_mse_loss, family="G_loss") | |
g_loss_sum = tf.summary.scalar("g_loss", g_loss, family="G_loss") | |
# final summary operations | |
g_sum = tf.summary.merge([d_loss_fake_sum, g_loss_sum, g_loss_l1_sum, gen__image_sum, real_image_sum, top_sum]) | |
d_sum = tf.summary.merge([d_loss_real_sum, d_loss_sum, gen__image_sum, real_image_sum]) | |
sess = tf.Session() | |
coord = tf.train.Coordinator() | |
threads = tf.train.start_queue_runners(sess, coord) | |
summary_writer = tf.summary.FileWriter(cfg.log_dir, sess.graph) | |
saver = tf.train.Saver() | |
sess.run(tf.global_variables_initializer()) | |
sess.run(train_iterator.initializer) | |
saver.restore(sess, tf.train.latest_checkpoint(cfg.path_save_model)) | |
print("restore successfully !! " * 100) | |
try: | |
for epoch in range(cfg.epoch_size): | |
# Training | |
for itr in range(cfg.dataset_size // cfg.train_batch_size): | |
# noise_label = get_perturbed_batch(image_label) | |
# Update Dicriminator | |
d_loss_val, summary_str, opt_d = sess.run([d_loss, d_sum, d_train_opt]) | |
# Update Generator | |
g_loss_val, summary_str, opt_g = sess.run([g_loss, g_sum, g_train_opt]) | |
if itr % 50 == 0: | |
print("epoch - {} | iter - {} | d-loss - {}".format(epoch, itr, d_loss_val)) | |
summary_writer.add_summary(summary_str, itr) | |
print("epoch - {} | iter - {} | g-loss - {}".format(epoch, itr, g_loss_val)) | |
summary_writer.add_summary(summary_str, itr) | |
# summary_writer.add_summary(clothes_sumarry, itr) | |
saver.save(sess, cfg.path_save_model) | |
print("Successful !!!") | |
except Exception as es: | |
log.debug(es) | |
pass | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment