def train(self, target_input_values, target_output_values):
l_index = len(self.layers) - 1
print(target_input_values)
print(target_output_values)
self.set_target_values(target_output_values, l_index)
self.get_values(target_input_values)
self.process_error()
for layer in self.layers:
layer.train()
def train(self, training_set):
l_index = len(self.layers) - 1
q_training = training_set.training_iterations
training_iter = 0
while training_iter < q_training:
training_iter += 1
self.set_target_values(training_set.output, l_index)
self.get_values(training_set.input)
self.process_error()
for layer in self.layers:
layer.train()
def main(args=None):
z = tf.placeholder(tf.float32, [flags.batch_size, flags.noise_size], name="z")
image = tf.placeholder(tf.float32, [flags.batch_size, flags.image_height, flags.image_width, 3])
label = tf.placeholder(tf.uint8, [flags.batch_size, 1])
source_logits_real, class_logits_real, source_logits_fake, class_logits_fake, generate_image = layer.inference(
image, label, z)
d_loss, g_loss = layer.loss(label,
source_logits_real=source_logits_real, class_logits_real=class_logits_real,
source_logits_fake=source_logits_fake, class_logits_fake=class_logits_fake
)
d_optimizer, g_optimazer = layer.train(d_loss, g_loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
reader = read_records.ReadRecords(train_path=flags.buckets + '*.train.tfrecords',
test_path=flags.buckets + '*.test.tfrecords')
tf.train.start_queue_runners(sess=sess)
g_loss_save = tf.summary.scalar('g_loss', g_loss)
d_loss_save = tf.summary.scalar('d_loss', d_loss)
saver = tf.train.Saver(var_list=[tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="discriminator"),
tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="generator")],
keep_checkpoint_every_n_hours=0.5)
# saver.restore(sess=sess, save_path=flags.checkpointDir)
summary = tf.summary.FileWriter(flags.summaryDir)
for step in xrange(flags.train_steps):
random_z = np.random.uniform(-1, 1, size=[flags.batch_size, flags.noise_size]).astype(np.float32)
image_read, label_read = sess.run(reader.read())
sess.run([d_optimizer, g_optimazer, g_optimazer], feed_dict={z: random_z, image: image_read, label: label_read})
if step % 50 == 0:
image_save = tf.summary.image('generate_image', generate_image, max_outputs=50)
merged = tf.summary.merge([image_save, g_loss_save, d_loss_save])
else:
merged = tf.summary.merge([g_loss_save, d_loss_save])
merged = sess.run(merged, feed_dict={z: random_z, image: image_read, label: label_read})
summary.add_summary(merged, step)
print "train steep: {}".format(step)
if step % 1000 == 0:
saver.save(sess=sess, save_path=flags.checkpointDir, global_step=step)
def train(model_checkpoint_path = 'log/model_dump/model_fix_input_bn_ema.ckpt',
has_bn=True, qweight=False, qactivation=False, image_norm=False):
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False, name='global_step')
learning_rate = tf.placeholder(dtype=tf.float32)
with tf.device('/cpu:0'):
train_iter = dataset.make_train_dataset()
image_batch, label_batch = train_iter.get_next()
print('images:', image_batch.shape, image_batch.dtype)
# Build inference Graph.
scale = None
if qweight or qactivation:
with open('log/scale', 'rb') as f:
scale = pickle.load(f)
logits = net.inference(image_batch, phase_train=True, has_bn=has_bn, image_norm=image_norm,
qactivation=qactivation, qweight=qweight, scale=scale)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
total_loss, softmax_loss, acc = layer.loss(logits, label_batch)
tf.summary.scalar('total_loss', total_loss)
tf.summary.scalar('softmax_loss', softmax_loss)
tf.summary.scalar('acc', acc)
train_op = layer.train(total_loss, learning_rate, global_step)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(cfg.log_path, tf.get_default_graph())
pre_saver = tf.train.Saver(tf.get_collection('params'))
# Create a saver.
saver = tf.train.Saver(max_to_keep=5000)
init = tf.global_variables_initializer()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False,
gpu_options=tf.GPUOptions(allow_growth=True)))
sess.run(init)
if model_checkpoint_path is not None:
pre_saver.restore(sess, model_checkpoint_path)
print('init model from {}'.format(model_checkpoint_path))
best_val_acc = evaluate(model_checkpoint_path, has_bn=True,
qweight=True, qactivation=True, image_norm=False)
if cfg.timeline_log:
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
else:
options = None
run_metadata = None
start_time = time.time()
train_num_batch = cfg.train_num // cfg.batch_size
train_log = open('log/work_log.txt', 'w')
for epoch in range(cfg.max_epoch):
for step in range(0, train_num_batch):
lr = get_learning_rate(epoch, step, train_num_batch)
feed_dict = {learning_rate: lr}
if step % cfg.log_step == 0:
_, _total_loss, _softmax_loss, _acc, _summary = \
sess.run([train_op, total_loss, softmax_loss, acc, summary_op],
feed_dict=feed_dict,
options=options,
run_metadata=run_metadata
)
duration = float(time.time() - start_time) / cfg.log_step
examples_per_sec = cfg.batch_size / duration
log_line = "%s: Epoch=%d/%d, Step=%d/%d, lr=%.7f, total_loss=%.3f, softmax_loss=%.3f, " \
"acc=%.2f%%(%.1f examples/sec; %.3f sec/batch)" \
% (datetime.now().strftime('%m-%d %H:%M:%S'), epoch, cfg.max_epoch,
step, train_num_batch, lr,
_total_loss, _softmax_loss, _acc,
examples_per_sec, duration)
train_log.write(log_line + '\n')
print(log_line)
summary_writer.add_summary(_summary, global_step=step)
start_time = time.time()
if cfg.timeline_log:
tl = timeline.Timeline(run_metadata.step_stats)
ctf = tl.generate_chrome_trace_format()
with open('timeline.json', 'w') as wd:
wd.write(ctf)
else:
_ = sess.run(train_op, feed_dict=feed_dict)
saver.save(sess, '{}/model.ckpt'.format(cfg.model_path), global_step=epoch)
_val_acc = evaluate('{}/model.ckpt-{}'.format(cfg.model_path, epoch), has_bn=True,
qweight=True, qactivation=True, image_norm=False)
if _val_acc > best_val_acc:
pre_saver.save(sess, '{}/best_model.ckpt'.format(cfg.model_path), write_meta_graph=False,
write_state=False, global_step=None)
best_val_acc = _val_acc
val_log = 'epoch=%d, val_acc=%.3f%%, best_val_acc=%.3f%%' \
% (epoch, _val_acc, best_val_acc)
train_log.write(val_log + '\n')
print(val_log)
elif y == 0 and x > 0:
output.append([1, 0, 0, 1])
elif y == 0 and x < 0:
output.append([0, 1, 1, 0])
elif x == 0 and y == 0:
output.append([0, 0, 0, 0])
input = np.array(input)
output = np.array(output)
layer1 = Layer(3, None, 10, input=input, inputLayer=True)
layer2 = Layer(3, layer1, .1)
layer3 = Layer(3, layer2, .001)
layer4 = Layer(4, layer3, None, outputLayer=True)
network.train(output, iterations=90000)
testCase = np.array([[1, 100, 100]])
print(getClassification(network.getHypothesis(testCase)))
testCase = np.array([[1, -100, 100]])
print(getClassification(network.getHypothesis(testCase)))
testCase = np.array([[1, -100, -100]])
print(getClassification(network.getHypothesis(testCase)))
testCase = np.array([[1, 100, -100]])
print(getClassification(network.getHypothesis(testCase)))
print(" ")
testCase = np.array([[1, 0, 100]])
print(getClassification(network.getHypothesis(testCase)))
testCase = np.array([[1, 0, -100]])
f = open('im_data.txt', 'r')
datax = sj.loads(f.read())
f.close()
arr = v.vec_data('test.jpg')
if len(arr) != 0:
for z in range(15):
datax.append(arr)
datay = []
for z in range(30):
if z < 15:
arr = [0]
datay.append(arr)
else:
arr = [1]
datay.append(arr)
net = l.init_network(128, 6, 1)
l.train(net, datax, datay, .1, 1000)
f = open('shihab_data.txt', 'w')
sj.dump(net, f)
f.close()
else:
print 'no face found'
