|
|
|
|
@ -1,4 +1,4 @@
|
|
|
|
|
# Copyright (c) 2016 Baidu, Inc. All Rights Reserved
|
|
|
|
|
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
|
|
|
|
|
#
|
|
|
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
|
|
# you may not use this file except in compliance with the License.
|
|
|
|
|
@ -24,6 +24,7 @@ from paddle.trainer.config_parser import logger
|
|
|
|
|
import py_paddle.swig_paddle as api
|
|
|
|
|
import matplotlib.pyplot as plt
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def plot2DScatter(data, outputfile):
|
|
|
|
|
'''
|
|
|
|
|
Plot the data as a 2D scatter plot and save to outputfile
|
|
|
|
|
@ -41,9 +42,11 @@ def plot2DScatter(data, outputfile):
|
|
|
|
|
plt.scatter(x, y)
|
|
|
|
|
plt.savefig(outputfile, bbox_inches='tight')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def CHECK_EQ(a, b):
|
|
|
|
|
assert a == b, "a=%s, b=%s" % (a, b)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def copy_shared_parameters(src, dst):
|
|
|
|
|
'''
|
|
|
|
|
copy the parameters from src to dst
|
|
|
|
|
@ -52,11 +55,9 @@ def copy_shared_parameters(src, dst):
|
|
|
|
|
:param dst: the destination of the parameters
|
|
|
|
|
:type dst: GradientMachine
|
|
|
|
|
'''
|
|
|
|
|
src_params = [src.getParameter(i)
|
|
|
|
|
for i in xrange(src.getParameterSize())]
|
|
|
|
|
src_params = [src.getParameter(i) for i in xrange(src.getParameterSize())]
|
|
|
|
|
src_params = dict([(p.getName(), p) for p in src_params])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
for i in xrange(dst.getParameterSize()):
|
|
|
|
|
dst_param = dst.getParameter(i)
|
|
|
|
|
src_param = src_params.get(dst_param.getName(), None)
|
|
|
|
|
@ -68,14 +69,16 @@ def copy_shared_parameters(src, dst):
|
|
|
|
|
dst_value.copyFrom(src_value)
|
|
|
|
|
dst_param.setValueUpdated()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def print_parameters(src):
|
|
|
|
|
src_params = [src.getParameter(i)
|
|
|
|
|
for i in xrange(src.getParameterSize())]
|
|
|
|
|
src_params = [src.getParameter(i) for i in xrange(src.getParameterSize())]
|
|
|
|
|
|
|
|
|
|
print "***************"
|
|
|
|
|
for p in src_params:
|
|
|
|
|
print "Name is %s" % p.getName()
|
|
|
|
|
print "value is %s \n" % p.getBuf(api.PARAMETER_VALUE).copyToNumpyArray()
|
|
|
|
|
print "value is %s \n" % p.getBuf(api.PARAMETER_VALUE).copyToNumpyArray(
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def load_mnist_data(imageFile):
|
|
|
|
|
f = open(imageFile, "rb")
|
|
|
|
|
@ -93,6 +96,7 @@ def load_mnist_data(imageFile):
|
|
|
|
|
f.close()
|
|
|
|
|
return data.astype('float32')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def load_cifar_data(cifar_path):
|
|
|
|
|
batch_size = 10000
|
|
|
|
|
data = numpy.zeros((5 * batch_size, 32 * 32 * 3), dtype="float32")
|
|
|
|
|
@ -106,11 +110,13 @@ def load_cifar_data(cifar_path):
|
|
|
|
|
data = data / 255.0 * 2.0 - 1.0
|
|
|
|
|
return data
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# synthesize 2-D uniform data
|
|
|
|
|
def load_uniform_data():
|
|
|
|
|
data = numpy.random.rand(1000000, 2).astype('float32')
|
|
|
|
|
return data
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def merge(images, size):
|
|
|
|
|
if images.shape[1] == 28 * 28:
|
|
|
|
|
h, w, c = 28, 28, 1
|
|
|
|
|
@ -124,6 +130,7 @@ def merge(images, size):
|
|
|
|
|
((images[idx, :].reshape((h, w, c), order="F").transpose(1, 0, 2) + 1.0) / 2.0 * 255.0)
|
|
|
|
|
return img.astype('uint8')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def save_images(images, path):
|
|
|
|
|
merged_img = merge(images, [8, 8])
|
|
|
|
|
if merged_img.shape[2] == 1:
|
|
|
|
|
@ -132,13 +139,16 @@ def save_images(images, path):
|
|
|
|
|
im = Image.fromarray(merged_img, mode="RGB")
|
|
|
|
|
im.save(path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_real_samples(batch_size, data_np):
|
|
|
|
|
return data_np[numpy.random.choice(data_np.shape[0], batch_size,
|
|
|
|
|
replace=False),:]
|
|
|
|
|
return data_np[numpy.random.choice(
|
|
|
|
|
data_np.shape[0], batch_size, replace=False), :]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_noise(batch_size, noise_dim):
|
|
|
|
|
return numpy.random.normal(size=(batch_size, noise_dim)).astype('float32')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_fake_samples(generator_machine, batch_size, noise):
|
|
|
|
|
gen_inputs = api.Arguments.createArguments(1)
|
|
|
|
|
gen_inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(noise))
|
|
|
|
|
@ -147,12 +157,14 @@ def get_fake_samples(generator_machine, batch_size, noise):
|
|
|
|
|
fake_samples = gen_outputs.getSlotValue(0).copyToNumpyMat()
|
|
|
|
|
return fake_samples
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_training_loss(training_machine, inputs):
|
|
|
|
|
outputs = api.Arguments.createArguments(0)
|
|
|
|
|
training_machine.forward(inputs, outputs, api.PASS_TEST)
|
|
|
|
|
loss = outputs.getSlotValue(0).copyToNumpyMat()
|
|
|
|
|
return numpy.mean(loss)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def prepare_discriminator_data_batch_pos(batch_size, data_np):
|
|
|
|
|
real_samples = get_real_samples(batch_size, data_np)
|
|
|
|
|
labels = numpy.ones(batch_size, dtype='int32')
|
|
|
|
|
@ -161,6 +173,7 @@ def prepare_discriminator_data_batch_pos(batch_size, data_np):
|
|
|
|
|
inputs.setSlotIds(1, api.IVector.createVectorFromNumpy(labels))
|
|
|
|
|
return inputs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def prepare_discriminator_data_batch_neg(generator_machine, batch_size, noise):
|
|
|
|
|
fake_samples = get_fake_samples(generator_machine, batch_size, noise)
|
|
|
|
|
labels = numpy.zeros(batch_size, dtype='int32')
|
|
|
|
|
@ -169,6 +182,7 @@ def prepare_discriminator_data_batch_neg(generator_machine, batch_size, noise):
|
|
|
|
|
inputs.setSlotIds(1, api.IVector.createVectorFromNumpy(labels))
|
|
|
|
|
return inputs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def prepare_generator_data_batch(batch_size, noise):
|
|
|
|
|
label = numpy.ones(batch_size, dtype='int32')
|
|
|
|
|
inputs = api.Arguments.createArguments(2)
|
|
|
|
|
@ -193,10 +207,9 @@ def get_layer_size(model_conf, layer_name):
|
|
|
|
|
def main():
|
|
|
|
|
parser = argparse.ArgumentParser()
|
|
|
|
|
parser.add_argument("-d", "--data_source", help="mnist or cifar or uniform")
|
|
|
|
|
parser.add_argument("--use_gpu", default="1",
|
|
|
|
|
help="1 means use gpu for training")
|
|
|
|
|
parser.add_argument("--gpu_id", default="0",
|
|
|
|
|
help="the gpu_id parameter")
|
|
|
|
|
parser.add_argument(
|
|
|
|
|
"--use_gpu", default="1", help="1 means use gpu for training")
|
|
|
|
|
parser.add_argument("--gpu_id", default="0", help="the gpu_id parameter")
|
|
|
|
|
args = parser.parse_args()
|
|
|
|
|
data_source = args.data_source
|
|
|
|
|
use_gpu = args.use_gpu
|
|
|
|
|
@ -209,8 +222,9 @@ def main():
|
|
|
|
|
if not os.path.exists("./%s_params/" % data_source):
|
|
|
|
|
os.makedirs("./%s_params/" % data_source)
|
|
|
|
|
|
|
|
|
|
api.initPaddle('--use_gpu=' + use_gpu, '--dot_period=10', '--log_period=100',
|
|
|
|
|
'--gpu_id=' + args.gpu_id, '--save_dir=' + "./%s_params/" % data_source)
|
|
|
|
|
api.initPaddle('--use_gpu=' + use_gpu, '--dot_period=10',
|
|
|
|
|
'--log_period=100', '--gpu_id=' + args.gpu_id,
|
|
|
|
|
'--save_dir=' + "./%s_params/" % data_source)
|
|
|
|
|
|
|
|
|
|
if data_source == "uniform":
|
|
|
|
|
conf = "gan_conf.py"
|
|
|
|
|
@ -220,7 +234,8 @@ def main():
|
|
|
|
|
num_iter = 1000
|
|
|
|
|
|
|
|
|
|
gen_conf = parse_config(conf, "mode=generator_training,data=" + data_source)
|
|
|
|
|
dis_conf = parse_config(conf, "mode=discriminator_training,data=" + data_source)
|
|
|
|
|
dis_conf = parse_config(conf,
|
|
|
|
|
"mode=discriminator_training,data=" + data_source)
|
|
|
|
|
generator_conf = parse_config(conf, "mode=generator,data=" + data_source)
|
|
|
|
|
batch_size = dis_conf.opt_config.batch_size
|
|
|
|
|
noise_dim = get_layer_size(gen_conf.model_config, "noise")
|
|
|
|
|
@ -245,11 +260,9 @@ def main():
|
|
|
|
|
generator_machine = api.GradientMachine.createFromConfigProto(
|
|
|
|
|
generator_conf.model_config)
|
|
|
|
|
|
|
|
|
|
dis_trainer = api.Trainer.create(
|
|
|
|
|
dis_conf, dis_training_machine)
|
|
|
|
|
dis_trainer = api.Trainer.create(dis_conf, dis_training_machine)
|
|
|
|
|
|
|
|
|
|
gen_trainer = api.Trainer.create(
|
|
|
|
|
gen_conf, gen_training_machine)
|
|
|
|
|
gen_trainer = api.Trainer.create(gen_conf, gen_training_machine)
|
|
|
|
|
|
|
|
|
|
dis_trainer.startTrain()
|
|
|
|
|
gen_trainer.startTrain()
|
|
|
|
|
@ -272,21 +285,23 @@ def main():
|
|
|
|
|
noise = get_noise(batch_size, noise_dim)
|
|
|
|
|
data_batch_dis_pos = prepare_discriminator_data_batch_pos(
|
|
|
|
|
batch_size, data_np)
|
|
|
|
|
dis_loss_pos = get_training_loss(dis_training_machine, data_batch_dis_pos)
|
|
|
|
|
dis_loss_pos = get_training_loss(dis_training_machine,
|
|
|
|
|
data_batch_dis_pos)
|
|
|
|
|
|
|
|
|
|
data_batch_dis_neg = prepare_discriminator_data_batch_neg(
|
|
|
|
|
generator_machine, batch_size, noise)
|
|
|
|
|
dis_loss_neg = get_training_loss(dis_training_machine, data_batch_dis_neg)
|
|
|
|
|
dis_loss_neg = get_training_loss(dis_training_machine,
|
|
|
|
|
data_batch_dis_neg)
|
|
|
|
|
|
|
|
|
|
dis_loss = (dis_loss_pos + dis_loss_neg) / 2.0
|
|
|
|
|
|
|
|
|
|
# Do forward pass in generator to get the gen_loss
|
|
|
|
|
data_batch_gen = prepare_generator_data_batch(
|
|
|
|
|
batch_size, noise)
|
|
|
|
|
data_batch_gen = prepare_generator_data_batch(batch_size, noise)
|
|
|
|
|
gen_loss = get_training_loss(gen_training_machine, data_batch_gen)
|
|
|
|
|
|
|
|
|
|
if i % 100 == 0:
|
|
|
|
|
print "d_pos_loss is %s d_neg_loss is %s" % (dis_loss_pos, dis_loss_neg)
|
|
|
|
|
print "d_pos_loss is %s d_neg_loss is %s" % (dis_loss_pos,
|
|
|
|
|
dis_loss_neg)
|
|
|
|
|
print "d_loss is %s g_loss is %s" % (dis_loss, gen_loss)
|
|
|
|
|
|
|
|
|
|
# Decide which network to train based on the training history
|
|
|
|
|
@ -300,7 +315,8 @@ def main():
|
|
|
|
|
curr_strike = 1
|
|
|
|
|
dis_trainer.trainOneDataBatch(batch_size, data_batch_dis_neg)
|
|
|
|
|
dis_trainer.trainOneDataBatch(batch_size, data_batch_dis_pos)
|
|
|
|
|
copy_shared_parameters(dis_training_machine, gen_training_machine)
|
|
|
|
|
copy_shared_parameters(dis_training_machine,
|
|
|
|
|
gen_training_machine)
|
|
|
|
|
|
|
|
|
|
else:
|
|
|
|
|
if curr_train == "gen":
|
|
|
|
|
@ -311,7 +327,8 @@ def main():
|
|
|
|
|
gen_trainer.trainOneDataBatch(batch_size, data_batch_gen)
|
|
|
|
|
# TODO: add API for paddle to allow true parameter sharing between different GradientMachines
|
|
|
|
|
# so that we do not need to copy shared parameters.
|
|
|
|
|
copy_shared_parameters(gen_training_machine, dis_training_machine)
|
|
|
|
|
copy_shared_parameters(gen_training_machine,
|
|
|
|
|
dis_training_machine)
|
|
|
|
|
copy_shared_parameters(gen_training_machine, generator_machine)
|
|
|
|
|
|
|
|
|
|
dis_trainer.finishTrainPass()
|
|
|
|
|
@ -319,11 +336,14 @@ def main():
|
|
|
|
|
# At the end of each pass, save the generated samples/images
|
|
|
|
|
fake_samples = get_fake_samples(generator_machine, batch_size, noise)
|
|
|
|
|
if data_source == "uniform":
|
|
|
|
|
plot2DScatter(fake_samples, "./%s_samples/train_pass%s.png" % (data_source, train_pass))
|
|
|
|
|
plot2DScatter(fake_samples, "./%s_samples/train_pass%s.png" %
|
|
|
|
|
(data_source, train_pass))
|
|
|
|
|
else:
|
|
|
|
|
save_images(fake_samples, "./%s_samples/train_pass%s.png" % (data_source, train_pass))
|
|
|
|
|
save_images(fake_samples, "./%s_samples/train_pass%s.png" %
|
|
|
|
|
(data_source, train_pass))
|
|
|
|
|
dis_trainer.finishTrain()
|
|
|
|
|
gen_trainer.finishTrain()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
|
|
|
|
main()
|
|
|
|
|
|
wangyanfei01
9 years ago