def __init__(self, train_conf, dict_file, model_dir, label_file,
predicate_dict_file):
"""
train_conf: trainer configure.
dict_file: word dictionary file name.
model_dir: directory of model.
"""
self.dict = {}
self.labels = {}
self.predicate_dict = {}
self.labels_reverse = {}
self.load_dict_label(dict_file, label_file, predicate_dict_file)
len_dict = len(self.dict)
len_label = len(self.labels)
len_pred = len(self.predicate_dict)
conf = parse_config(
train_conf, 'dict_len=' + str(len_dict) + ',label_len=' +
str(len_label) + ',pred_len=' + str(len_pred) + ',is_predict=True')
self.network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
self.network.loadParameters(model_dir)
slots = [
integer_value_sequence(len_dict), integer_value_sequence(len_dict),
integer_value_sequence(len_dict), integer_value_sequence(len_dict),
integer_value_sequence(len_dict), integer_value_sequence(len_dict),
integer_value_sequence(len_pred), integer_value_sequence(2)
]
self.converter = DataProviderConverter(slots)
def __init__(self, train_conf, dict_file, model_dir, label_file):
"""
train_conf: trainer configure.
dict_file: word dictionary file name.
model_dir: directory of model.
"""
self.dict = {}
self.labels = {}
self.labels_reverse = {}
self.load_dict_label(dict_file, label_file)
len_dict = len(self.dict)
len_label = len(self.labels)
conf = parse_config(
train_conf, 'dict_len=' + str(len_dict) + ',label_len=' +
str(len_label) + ',is_predict=True')
self.network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
self.network.loadParameters(model_dir)
slots = [
integer_value_sequence(len_dict),
integer_value_sequence(len_dict),
integer_value_sequence(len_dict),
integer_value_sequence(len_dict),
integer_value_sequence(len_dict),
integer_value_sequence(2)
]
self.converter = DataProviderConverter(slots)
def main():
conf = parse_config("./mnist_model/trainer_config.conf.norm", "")
print conf.data_config.load_data_args
network = swig_paddle.GradientMachine.createFromConfigProto(conf.model_config)
assert isinstance(network, swig_paddle.GradientMachine) # For code hint.
network.loadParameters("./mnist_model/")
converter = DataProviderWrapperConverter(False, [DenseSlot(784)])
inArg = converter(TEST_DATA)
print network.forwardTest(inArg)
def __init__(self, train_conf, model_dir=None,
resize_dim=256, crop_dim=224,
mean_file=None,
output_layer=None,
oversample=False, is_color=True):
"""
train_conf: network configure.
model_dir: string, directory of model.
resize_dim: int, resized image size.
crop_dim: int, crop size.
mean_file: string, image mean file.
oversample: bool, oversample means multiple crops, namely five
patches (the four corner patches and the center
patch) as well as their horizontal reflections,
ten crops in all.
"""
self.train_conf = train_conf
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.resize_dim = resize_dim
self.crop_dims = [crop_dim, crop_dim]
self.oversample = oversample
self.is_color = is_color
self.output_layer = output_layer
if self.output_layer:
assert isinstance(self.output_layer, basestring)
self.output_layer = self.output_layer.split(",")
self.transformer = image_util.ImageTransformer(is_color = is_color)
self.transformer.set_transpose((2,0,1))
self.transformer.set_channel_swap((2,1,0))
self.mean_file = mean_file
if self.mean_file is not None:
mean = np.load(self.mean_file)['data_mean']
mean = mean.reshape(3, self.crop_dims[0], self.crop_dims[1])
self.transformer.set_mean(mean) # mean pixel
else:
# if you use three mean value, set like:
# this three mean value is calculated from ImageNet.
self.transformer.set_mean(np.array([103.939,116.779,123.68]))
conf_args = "is_test=1,use_gpu=1,is_predict=1"
conf = parse_config(train_conf, conf_args)
swig_paddle.initPaddle("--use_gpu=1")
self.network = swig_paddle.GradientMachine.createFromConfigProto(conf.model_config)
assert isinstance(self.network, swig_paddle.GradientMachine)
self.network.loadParameters(self.model_dir)
data_size = 3 * self.crop_dims[0] * self.crop_dims[1]
slots = [DenseSlot(data_size)]
is_sequence = False
self.converter = util.DataProviderWrapperConverter(is_sequence, slots)
def main():
conf = parse_config("./mnist_model/trainer_config.py", "")
print conf.data_config.load_data_args
network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(network, swig_paddle.GradientMachine) # For code hint.
network.loadParameters("./mnist_model/")
converter = DataProviderConverter([dense_vector(784)])
inArg = converter(TEST_DATA)
print network.forwardTest(inArg)
def __init__(self,
train_conf,
use_gpu=True,
model_dir=None,
resize_dim=None,
crop_dim=None,
mean_file=None,
oversample=False,
is_color=False):
"""
train_conf: 网络配置文件
model_dir: 模型路径
resize_dim: 设为原图大小
crop_dim: 图像裁剪大小，一般设为原图大小
oversample: bool, oversample表示多次裁剪,这里禁用
"""
self.train_conf = train_conf
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.resize_dim = resize_dim
self.crop_dims = [crop_dim, crop_dim]
self.oversample = oversample
self.is_color = is_color
self.transformer = image_util.ImageTransformer(is_color = is_color)
self.transformer.set_transpose((2,0,1))
self.mean_file = mean_file
mean = np.load(self.mean_file)['data_mean']
mean = mean.reshape(1, self.crop_dims[0], self.crop_dims[1])
self.transformer.set_mean(mean) # mean pixel
gpu = 1 if use_gpu else 0
conf_args = "is_test=1,use_gpu=%d,is_predict=1" % (gpu)
#使用 parse_config() 解析训练时的配置文件
conf = parse_config(train_conf, conf_args)
#PaddlePaddle目前使用Swig对其常用的预测接口进行了封装，使在Python环境下的预测接口更加简单
#使用 swig_paddle.initPaddle() 传入命令行参数初始化 PaddlePaddle
swig_paddle.initPaddle("--use_gpu=%d" % (int(use_gpu)))
#使用 swig_paddle.GradientMachine.createFromConfigproto() 根据上一步解析好的配置创建神经网络
self.network = swig_paddle.GradientMachine.createFromConfigProto(conf.model_config)
assert isinstance(self.network, swig_paddle.GradientMachine)
#从模型文件加载参数
self.network.loadParameters(self.model_dir)
data_size = 1 * self.crop_dims[0] * self.crop_dims[1]
slots = [dense_vector(data_size)]
'''
创建一个 DataProviderConverter 对象converter。
swig_paddle接受的原始数据是C++的Matrix，也就是直接写内存的float数组。 这个接口并不用户友好。所以，我们提供了一个工具类DataProviderConverter。 这个工具类接收和PyDataProvider2一样的输入数据
'''
self.converter = DataProviderConverter(slots)
def main():
options = parse_arguments()
api.initPaddle("--use_gpu=%s" % options.use_gpu,
"--trainer_count=%s" % options.trainer_count)
word_dict = load_dict(options.dict_file)
train_dataset = list(load_data(options.train_data, word_dict))
if options.test_data:
test_dataset = list(load_data(options.test_data, word_dict))
else:
test_dataset = None
trainer_config = parse_config(options.config,
"dict_file=%s" % options.dict_file)
# No need to have data provider for trainer
trainer_config.ClearField('data_config')
trainer_config.ClearField('test_data_config')
# create a GradientMachine from the model configuratin
model = api.GradientMachine.createFromConfigProto(
trainer_config.model_config)
# create a trainer for the gradient machine
trainer = api.Trainer.create(trainer_config, model)
# create a data converter which converts data to PaddlePaddle
# internal format
input_types = [
integer_value_sequence(len(word_dict))
if options.seq else sparse_binary_vector(len(word_dict)),
integer_value(2)
]
converter = DataProviderConverter(input_types)
batch_size = trainer_config.opt_config.batch_size
trainer.startTrain()
for train_pass in xrange(options.num_passes):
trainer.startTrainPass()
random.shuffle(train_dataset)
for pos in xrange(0, len(train_dataset), batch_size):
batch = itertools.islice(train_dataset, pos, pos + batch_size)
size = min(batch_size, len(train_dataset) - pos)
trainer.trainOneDataBatch(size, converter(batch))
trainer.finishTrainPass()
if test_dataset:
trainer.startTestPeriod()
for pos in xrange(0, len(test_dataset), batch_size):
batch = itertools.islice(test_dataset, pos, pos + batch_size)
size = min(batch_size, len(test_dataset) - pos)
trainer.testOneDataBatch(size, converter(batch))
trainer.finishTestPeriod()
trainer.finishTrain()
def main():
options = parse_arguments()
api.initPaddle("--use_gpu=%s" % options.use_gpu,
"--trainer_count=%s" % options.trainer_count)
word_dict = load_dict(options.dict_file)
train_dataset = list(load_data(options.train_data, word_dict))
if options.test_data:
test_dataset = list(load_data(options.test_data, word_dict))
else:
test_dataset = None
trainer_config = parse_config(options.config,
"dict_file=%s" % options.dict_file)
# No need to have data provider for trainer
trainer_config.ClearField('data_config')
trainer_config.ClearField('test_data_config')
# create a GradientMachine from the model configuratin
model = api.GradientMachine.createFromConfigProto(
trainer_config.model_config)
# create a trainer for the gradient machine
trainer = api.Trainer.create(trainer_config, model)
# create a data converter which converts data to PaddlePaddle
# internal format
input_types = [
integer_value_sequence(len(word_dict)) if options.seq else
sparse_binary_vector(len(word_dict)), integer_value(2)
]
converter = DataProviderConverter(input_types)
batch_size = trainer_config.opt_config.batch_size
trainer.startTrain()
for train_pass in xrange(options.num_passes):
trainer.startTrainPass()
random.shuffle(train_dataset)
for pos in xrange(0, len(train_dataset), batch_size):
batch = itertools.islice(train_dataset, pos, pos + batch_size)
size = min(batch_size, len(train_dataset) - pos)
trainer.trainOneDataBatch(size, converter(batch))
trainer.finishTrainPass()
if test_dataset:
trainer.startTestPeriod()
for pos in xrange(0, len(test_dataset), batch_size):
batch = itertools.islice(test_dataset, pos, pos + batch_size)
size = min(batch_size, len(test_dataset) - pos)
trainer.testOneDataBatch(size, converter(batch))
trainer.finishTestPeriod()
trainer.finishTrain()
def __init__(self,
train_conf,
use_gpu=True,
model_dir=None,
resize_dim=None,
crop_dim=None,
mean_file=None,
oversample=False,
is_color=True):
"""
train_conf: network configure.
model_dir: string, directory of model.
resize_dim: int, resized image size.
crop_dim: int, crop size.
mean_file: string, image mean file.
oversample: bool, oversample means multiple crops, namely five
patches (the four corner patches and the center
patch) as well as their horizontal reflections,
ten crops in all.
"""
self.train_conf = train_conf
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.resize_dim = resize_dim
self.crop_dims = [crop_dim, crop_dim]
self.oversample = oversample
self.is_color = is_color
self.transformer = image_util.ImageTransformer(is_color=is_color)
self.transformer.set_transpose((2, 0, 1))
self.mean_file = mean_file
mean = np.load(self.mean_file)['data_mean']
mean = mean.reshape(3, self.crop_dims[0], self.crop_dims[1])
self.transformer.set_mean(mean) # mean pixel
gpu = 1 if use_gpu else 0
conf_args = "is_test=1,use_gpu=%d,is_predict=1" % (gpu)
conf = parse_config(train_conf, conf_args)
swig_paddle.initPaddle("--use_gpu=%d" % (gpu))
self.network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(self.network, swig_paddle.GradientMachine)
self.network.loadParameters(self.model_dir)
data_size = 3 * self.crop_dims[0] * self.crop_dims[1]
slots = [DenseSlot(data_size)]
self.converter = util.DataProviderWrapperConverter(False, slots)
def __init__(self,
train_conf,
use_gpu=True,
model_dir=None,
resize_dim=None,
crop_dim=None,
mean_file=None,
oversample=False,
is_color=True):
"""
train_conf: network configure.
model_dir: string, directory of model.
resize_dim: int, resized image size.
crop_dim: int, crop size.
mean_file: string, image mean file.
oversample: bool, oversample means multiple crops, namely five
patches (the four corner patches and the center
patch) as well as their horizontal reflections,
ten crops in all.
"""
self.train_conf = train_conf
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.resize_dim = resize_dim
self.crop_dims = [crop_dim, crop_dim]
self.oversample = oversample
self.is_color = is_color
self.transformer = image_util.ImageTransformer(is_color=is_color)
self.transformer.set_transpose((2, 0, 1))
self.mean_file = mean_file
mean = np.load(self.mean_file)['data_mean']
mean = mean.reshape(3, self.crop_dims[0], self.crop_dims[1])
self.transformer.set_mean(mean) # mean pixel
gpu = 1 if use_gpu else 0
conf_args = "is_test=1,use_gpu=%d,is_predict=1" % (gpu)
conf = parse_config(train_conf, conf_args)
swig_paddle.initPaddle("--use_gpu=%d" % (gpu))
self.network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(self.network, swig_paddle.GradientMachine)
self.network.loadParameters(self.model_dir)
data_size = 3 * self.crop_dims[0] * self.crop_dims[1]
slots = [dense_vector(data_size)]
self.converter = DataProviderConverter(slots)
def predict(data):
path = os.path.split(os.path.realpath(__file__))[0]
conf = parse_config(path + "/trainer_config_age.py", "is_predict=1")
print conf.data_config.load_data_args
network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
network.loadParameters(path + "/output_age/pass-00099")
converter = DataProviderConverter([dense_vector(26)])
inArg = converter(data)
network.forwardTest(inArg)
output = network.getLayerOutputs("__fc_layer_0__")
#print output
prob = output["__fc_layer_0__"][0]
#print prob
lab = np.argsort(-prob)
#print lab
return lab[0]
def main():
trainer_config = parse_config("./testTrainConfig.py", "")
model = swig_paddle.GradientMachine.createFromConfigProto(
trainer_config.model_config)
trainer = swig_paddle.Trainer.create(trainer_config, model)
trainer.startTrain()
for train_pass in xrange(2):
trainer.startTrainPass()
num = 0
cost = 0
while True: # Train one batch
batch_size = 1000
data, atEnd = util.loadMNISTTrainData(batch_size)
if atEnd:
break
trainer.trainOneDataBatch(batch_size, data)
outs = trainer.getForwardOutput()
cost += sum(outs[0]['value'])
num += batch_size
trainer.finishTrainPass()
logger.info('train cost=%f' % (cost / num))
trainer.startTestPeriod()
num = 0
cost = 0
while True: # Test one batch
batch_size = 1000
data, atEnd = util.loadMNISTTrainData(batch_size)
if atEnd:
break
trainer.testOneDataBatch(batch_size, data)
outs = trainer.getForwardOutput()
cost += sum(outs[0]['value'])
num += batch_size
trainer.finishTestPeriod()
logger.info('test cost=%f' % (cost / num))
trainer.finishTrain()
def __init__(self, train_conf, dict_file, model_dir=None, label_file = None):
"""
train_conf: trainer configure.
dict_file: word dictionary file name.
model_dir: directory of model.
"""
self.train_conf = train_conf
self.dict_file = dict_file
self.word_dict = {}
self.dict_dim = self.load_dict()
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.label = None
if label_file is not None:
self.load_label(label_file)
conf = parse_config(train_conf, "is_predict=1")
self.network = swig_paddle.GradientMachine.createFromConfigProto(conf.model_config)
self.network.loadParameters(self.model_dir)
input_types = [integer_value_sequence(self.dict_dim)]
self.converter = DataProviderConverter(input_types)
def __init__(self, train_conf, dict_file, model_dir=None, label_file=None):
"""
train_conf: trainer configure.
dict_file: word dictionary file name.
model_dir: directory of model.
"""
self.train_conf = train_conf
self.dict_file = dict_file
self.word_dict = {}
self.dict_dim = self.load_dict()
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.label = None
if label_file is not None:
self.load_label(label_file)
conf = parse_config(train_conf, "is_predict=1")
self.network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
self.network.loadParameters(self.model_dir)
input_types = [integer_value_sequence(self.dict_dim)]
self.converter = DataProviderConverter(input_types)
def parse_network_config(network_conf, config_arg_str=''):
config = config_parser.parse_config(network_conf, config_arg_str)
return config.model_config
def parse_trainer_config(trainer_conf, config_arg_str):
return config_parser.parse_config(trainer_conf, config_arg_str)
def test_parse(self):
a = parse_config('trainer_config_helpers/tests/layers_test_config.py',
'')
b = parse_config('trainer_config_helpers/tests/layers_test_config.py',
'')
self.assertEqual(a, b)
def make_diagram(config_file, dot_file, config_arg_str):
config = parse_config(config_file, config_arg_str)
make_diagram_from_proto(config.model_config, dot_file)
def parse_optimizer_config(optimizer_conf, config_arg_str=''):
config = config_parser.parse_config(optimizer_conf, config_arg_str)
return config.opt_config
from py_paddle import swig_paddle, DataProviderConverter
from common_utils import *
from paddle.trainer.config_parser import parse_config
try:
import cPickle as pickle
except ImportError:
import pickle
import sys
if __name__ == '__main__':
model_path = sys.argv[1]
swig_paddle.initPaddle('--use_gpu=0')
conf = parse_config("trainer_config.py", "is_predict=1")
network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(network, swig_paddle.GradientMachine)
network.loadParameters(model_path)
with open('./data/meta.bin', 'rb') as f:
meta = pickle.load(f)
headers = [h[1] for h in meta_to_header(meta, 'movie')]
headers.extend([h[1] for h in meta_to_header(meta, 'user')])
cvt = DataProviderConverter(headers)
while True:
movie_id = int(raw_input("Input movie_id: "))
user_id = int(raw_input("Input user_id: "))
movie_meta = meta['movie'][movie_id] # Query Data From Meta.
user_meta = meta['user'][user_id]
data = [movie_id - 1]
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer.config_parser import parse_config
from paddle.proto import TrainerConfig_pb2
import sys
__all__ = []
if __name__ == '__main__':
whole_conf = False
if len(sys.argv) == 2:
conf = parse_config(sys.argv[1], '')
elif len(sys.argv) == 3:
conf = parse_config(sys.argv[1], sys.argv[2])
elif len(sys.argv) == 4:
conf = parse_config(sys.argv[1], sys.argv[2])
if sys.argv[3] == '--whole':
whole_conf = True
else:
raise RuntimeError()
assert isinstance(conf, TrainerConfig_pb2.TrainerConfig)
if whole_conf:
print conf
else:
print conf.model_config
def main():
parser = argparse.ArgumentParser()
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()
use_gpu = args.use_gpu
assert use_gpu in ["0", "1"]
if not os.path.exists("./samples/"):
os.makedirs("./samples/")
if not os.path.exists("./params/"):
os.makedirs("./params/")
api.initPaddle('--use_gpu=' + use_gpu, '--dot_period=10',
'--log_period=1000', '--gpu_id=' + args.gpu_id,
'--save_dir=' + "./params/")
conf = "vae_conf.py"
trainer_conf = parse_config(conf, "is_generating=False")
gener_conf = parse_config(conf, "is_generating=True")
batch_size = trainer_conf.opt_config.batch_size
noise_dim = get_layer_size(gener_conf.model_config, "noise")
mnist = dataloader.MNISTloader(batch_size=batch_size)
mnist.load_data()
training_machine = api.GradientMachine.createFromConfigProto(
trainer_conf.model_config)
generator_machine = api.GradientMachine.createFromConfigProto(
gener_conf.model_config)
trainer = api.Trainer.create(trainer_conf, training_machine)
trainer.startTrain()
for train_pass in xrange(100):
trainer.startTrainPass()
mnist.reset_pointer()
i = 0
it = 0
while mnist.pointer != 0 or i == 0:
X = mnist.next_batch().astype('float32')
inputs = api.Arguments.createArguments(1)
inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(X))
trainer.trainOneDataBatch(batch_size, inputs)
if it % 1000 == 0:
outputs = api.Arguments.createArguments(0)
training_machine.forward(inputs, outputs, api.PASS_TEST)
loss = np.mean(outputs.getSlotValue(0).copyToNumpyMat())
print "\niter: {}".format(str(it).zfill(3))
print "VAE loss: {}".format(str(loss).zfill(3))
#Sync parameters between networks (GradientMachine) at the beginning
copy_shared_parameters(training_machine, generator_machine)
z_samples = np.random.randn(batch_size,
noise_dim).astype('float32')
samples = get_fake_samples(generator_machine, batch_size,
z_samples)
#Generating the first 16 images for a picture.
figure = plot_samples(samples[:16])
plt.savefig(
"./samples/{}_{}.png".format(
str(train_pass).zfill(3), str(i).zfill(3)),
bbox_inches='tight')
plt.close(figure)
i += 1
it += 1
trainer.finishTrainPass()
trainer.finishTrain()
def make_diagram(config_file, dot_file, config_arg_str):
config = parse_config(config_file, config_arg_str)
# print >> sys.stderr, config
name2id = {}
f = open(dot_file, 'w')
submodel_layers = set()
def make_link(link):
return 'l%s -> l%s;' % (name2id[link.layer_name],
name2id[link.link_name])
def make_mem(mem):
s = ''
if mem.boot_layer_name:
s += 'l%s -> l%s;\n' % (name2id[mem.boot_layer_name],
name2id[mem.layer_name])
s += 'l%s -> l%s [style=dashed];' % (name2id[mem.layer_name],
name2id[mem.link_name])
return s
print >> f, 'digraph graphname {'
print >> f, 'node [width=0.375,height=0.25];'
for i in xrange(len(config.model_config.layers)):
l = config.model_config.layers[i]
name2id[l.name] = i
i = 0
for sub_model in config.model_config.sub_models:
if sub_model.name == 'root':
continue
print >> f, 'subgraph cluster_%s {' % i
print >> f, 'style=dashed;'
label = '%s ' % sub_model.name
if sub_model.reversed:
label += '<=='
print >> f, 'label = "%s";' % label
i += 1
submodel_layers.add(sub_model.name)
for layer_name in sub_model.layer_names:
submodel_layers.add(layer_name)
lid = name2id[layer_name]
layer_config = config.model_config.layers[lid]
label = make_layer_label(layer_config)
print >> f, 'l%s [label="%s", shape=box];' % (lid, label)
print >> f, '}'
for i in xrange(len(config.model_config.layers)):
l = config.model_config.layers[i]
if l.name not in submodel_layers:
label = make_layer_label(l)
print >> f, 'l%s [label="%s", shape=box];' % (i, label)
for sub_model in config.model_config.sub_models:
if sub_model.name == 'root':
continue
for link in sub_model.in_links:
print >> f, make_link(link)
for link in sub_model.out_links:
print >> f, make_link(link)
for mem in sub_model.memories:
print >> f, make_mem(mem)
for i in xrange(len(config.model_config.layers)):
for l in config.model_config.layers[i].inputs:
print >> f, 'l%s -> l%s [label="%s"];' % (
name2id[l.input_layer_name], i, l.input_parameter_name)
print >> f, '}'
f.close()
def make_diagram(config_file, dot_file, config_arg_str):
config = parse_config(config_file, config_arg_str)
# print >> sys.stderr, config
name2id = {}
f = open(dot_file, 'w')
submodel_layers = set()
def make_link(link):
return 'l%s -> l%s;' % (
name2id[link.layer_name], name2id[link.link_name])
def make_mem(mem):
s = ''
if mem.boot_layer_name:
s += 'l%s -> l%s;\n' % (
name2id[mem.boot_layer_name], name2id[mem.layer_name])
s += 'l%s -> l%s [style=dashed];' % (
name2id[mem.layer_name], name2id[mem.link_name])
return s
print >> f, 'digraph graphname {'
print >> f, 'node [width=0.375,height=0.25];'
for i in xrange(len(config.model_config.layers)):
l = config.model_config.layers[i]
name2id[l.name] = i
i = 0
for sub_model in config.model_config.sub_models:
if sub_model.name == 'root':
continue
print >> f, 'subgraph cluster_%s {' % i
print >> f, 'style=dashed;'
label = '%s ' % sub_model.name
if sub_model.reversed:
label += '<=='
print >> f, 'label = "%s";' % label
i += 1
submodel_layers.add(sub_model.name)
for layer_name in sub_model.layer_names:
submodel_layers.add(layer_name)
lid = name2id[layer_name]
layer_config = config.model_config.layers[lid]
label = make_layer_label(layer_config)
print >> f, 'l%s [label="%s", shape=box];' % (lid, label)
print >> f, '}'
for i in xrange(len(config.model_config.layers)):
l = config.model_config.layers[i]
if l.name not in submodel_layers:
label = make_layer_label(l)
print >> f, 'l%s [label="%s", shape=box];' % (i, label)
for sub_model in config.model_config.sub_models:
if sub_model.name == 'root':
continue
for link in sub_model.in_links:
print >> f, make_link(link)
for link in sub_model.out_links:
print >> f, make_link(link)
for mem in sub_model.memories:
print >> f, make_mem(mem)
for i in xrange(len(config.model_config.layers)):
for l in config.model_config.layers[i].inputs:
print >> f, 'l%s -> l%s [label="%s"];' % (
name2id[l.input_layer_name], i, l.input_parameter_name)
print >> f, '}'
f.close()
def __init__(self,
train_conf,
model_dir=None,
resize_dim=256,
crop_dim=224,
use_gpu=True,
mean_file=None,
output_layer=None,
oversample=False,
is_color=True):
"""
train_conf: network configure.
model_dir: string, directory of model.
resize_dim: int, resized image size.
crop_dim: int, crop size.
mean_file: string, image mean file.
oversample: bool, oversample means multiple crops, namely five
patches (the four corner patches and the center
patch) as well as their horizontal reflections,
ten crops in all.
"""
self.train_conf = train_conf
self.model_dir = model_dir
if model_dir is None:
self.model_dir = os.path.dirname(train_conf)
self.resize_dim = resize_dim
self.crop_dims = [crop_dim, crop_dim]
self.oversample = oversample
self.is_color = is_color
self.output_layer = output_layer
if self.output_layer:
assert isinstance(self.output_layer, basestring)
self.output_layer = self.output_layer.split(",")
self.transformer = image_util.ImageTransformer(is_color=is_color)
self.transformer.set_transpose((2, 0, 1))
self.transformer.set_channel_swap((2, 1, 0))
self.mean_file = mean_file
if self.mean_file is not None:
mean = np.load(self.mean_file)['data_mean']
mean = mean.reshape(3, self.crop_dims[0], self.crop_dims[1])
self.transformer.set_mean(mean) # mean pixel
else:
# if you use three mean value, set like:
# this three mean value is calculated from ImageNet.
self.transformer.set_mean(np.array([103.939, 116.779, 123.68]))
conf_args = "is_test=1,use_gpu=%d,is_predict=1" % (int(use_gpu))
conf = parse_config(train_conf, conf_args)
swig_paddle.initPaddle("--use_gpu=%d" % (int(use_gpu)))
self.network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(self.network, swig_paddle.GradientMachine)
self.network.loadParameters(self.model_dir)
data_size = 3 * self.crop_dims[0] * self.crop_dims[1]
slots = [dense_vector(data_size)]
self.converter = DataProviderConverter(slots)
def parse_network_config(network_conf):
config = config_parser.parse_config(network_conf, '')
return config.model_config
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")
args = parser.parse_args()
data_source = args.data_source
use_gpu = args.use_gpu
assert data_source in ["mnist", "cifar", "uniform"]
assert use_gpu in ["0", "1"]
if not os.path.exists("./%s_samples/" % data_source):
os.makedirs("./%s_samples/" % data_source)
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)
if data_source == "uniform":
conf = "gan_conf.py"
num_iter = 10000
else:
conf = "gan_conf_image.py"
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)
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")
if data_source == "mnist":
data_np = load_mnist_data("./data/mnist_data/train-images-idx3-ubyte")
elif data_source == "cifar":
data_np = load_cifar_data("./data/cifar-10-batches-py/")
else:
data_np = load_uniform_data()
# this creates a gradient machine for discriminator
dis_training_machine = api.GradientMachine.createFromConfigProto(
dis_conf.model_config)
# this create a gradient machine for generator
gen_training_machine = api.GradientMachine.createFromConfigProto(
gen_conf.model_config)
# generator_machine is used to generate data only, which is used for
# training discriminator
logger.info(str(generator_conf.model_config))
generator_machine = api.GradientMachine.createFromConfigProto(
generator_conf.model_config)
dis_trainer = api.Trainer.create(dis_conf, dis_training_machine)
gen_trainer = api.Trainer.create(gen_conf, gen_training_machine)
dis_trainer.startTrain()
gen_trainer.startTrain()
# Sync parameters between networks (GradientMachine) at the beginning
copy_shared_parameters(gen_training_machine, dis_training_machine)
copy_shared_parameters(gen_training_machine, generator_machine)
# constrain that either discriminator or generator can not be trained
# consecutively more than MAX_strike times
curr_train = "dis"
curr_strike = 0
MAX_strike = 5
for train_pass in xrange(100):
dis_trainer.startTrainPass()
gen_trainer.startTrainPass()
for i in xrange(num_iter):
# Do forward pass in discriminator to get the dis_loss
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)
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 = (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)
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_loss is %s g_loss is %s" % (dis_loss, gen_loss)
# Decide which network to train based on the training history
# And the relative size of the loss
if (not (curr_train == "dis" and curr_strike == MAX_strike)) and \
((curr_train == "gen" and curr_strike == MAX_strike) or dis_loss > gen_loss):
if curr_train == "dis":
curr_strike += 1
else:
curr_train = "dis"
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)
else:
if curr_train == "gen":
curr_strike += 1
else:
curr_train = "gen"
curr_strike = 1
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, generator_machine)
dis_trainer.finishTrainPass()
gen_trainer.finishTrainPass()
# 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))
else:
save_images(fake_samples, "./%s_samples/train_pass%s.png" %
(data_source, train_pass))
dis_trainer.finishTrain()
gen_trainer.finishTrain()
def parse_optimizer_config(optimizer_conf):
config = config_parser.parse_config(optimizer_conf, '')
return config.opt_config
from py_paddle import swig_paddle, DataProviderConverter
from common_utils import *
from paddle.trainer.config_parser import parse_config
try:
import cPickle as pickle
except ImportError:
import pickle
import sys
if __name__ == '__main__':
model_path = sys.argv[1]
swig_paddle.initPaddle('--use_gpu=0')
conf = parse_config("trainer_config.py", "is_predict=1")
network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(network, swig_paddle.GradientMachine)
network.loadParameters(model_path)
with open('./data/meta.bin', 'rb') as f:
meta = pickle.load(f)
headers = list(meta_to_header(meta, 'movie'))
headers.extend(list(meta_to_header(meta, 'user')))
cvt = DataProviderConverter(headers)
while True:
movie_id = int(raw_input("Input movie_id: "))
user_id = int(raw_input("Input user_id: "))
movie_meta = meta['movie'][movie_id] # Query Data From Meta.
user_meta = meta['user'][user_id]
data = [movie_id - 1]
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")
args = parser.parse_args()
data_source = args.data_source
use_gpu = args.use_gpu
assert data_source in ["mnist", "cifar", "uniform"]
assert use_gpu in ["0", "1"]
if not os.path.exists("./%s_samples/" % data_source):
os.makedirs("./%s_samples/" % data_source)
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)
if data_source == "uniform":
conf = "gan_conf.py"
num_iter = 10000
else:
conf = "gan_conf_image.py"
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)
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")
if data_source == "mnist":
data_np = load_mnist_data("./data/mnist_data/train-images-idx3-ubyte")
elif data_source == "cifar":
data_np = load_cifar_data("./data/cifar-10-batches-py/")
else:
data_np = load_uniform_data()
# this creates a gradient machine for discriminator
dis_training_machine = api.GradientMachine.createFromConfigProto(
dis_conf.model_config)
# this create a gradient machine for generator
gen_training_machine = api.GradientMachine.createFromConfigProto(
gen_conf.model_config)
# generator_machine is used to generate data only, which is used for
# training discriminator
logger.info(str(generator_conf.model_config))
generator_machine = api.GradientMachine.createFromConfigProto(
generator_conf.model_config)
dis_trainer = api.Trainer.create(dis_conf, dis_training_machine)
gen_trainer = api.Trainer.create(gen_conf, gen_training_machine)
dis_trainer.startTrain()
gen_trainer.startTrain()
# Sync parameters between networks (GradientMachine) at the beginning
copy_shared_parameters(gen_training_machine, dis_training_machine)
copy_shared_parameters(gen_training_machine, generator_machine)
# constrain that either discriminator or generator can not be trained
# consecutively more than MAX_strike times
curr_train = "dis"
curr_strike = 0
MAX_strike = 5
for train_pass in xrange(100):
dis_trainer.startTrainPass()
gen_trainer.startTrainPass()
for i in xrange(num_iter):
# Do forward pass in discriminator to get the dis_loss
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)
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 = (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)
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_loss is %s g_loss is %s" % (dis_loss, gen_loss)
# Decide which network to train based on the training history
# And the relative size of the loss
if (not (curr_train == "dis" and curr_strike == MAX_strike)) and \
((curr_train == "gen" and curr_strike == MAX_strike) or dis_loss > gen_loss):
if curr_train == "dis":
curr_strike += 1
else:
curr_train = "dis"
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)
else:
if curr_train == "gen":
curr_strike += 1
else:
curr_train = "gen"
curr_strike = 1
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, generator_machine)
dis_trainer.finishTrainPass()
gen_trainer.finishTrainPass()
# 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))
else:
save_images(
fake_samples,
"./%s_samples/train_pass%s.png" % (data_source, train_pass))
dis_trainer.finishTrain()
gen_trainer.finishTrain()
def make_diagram(config_file, dot_file, config_arg_str):
config = parse_config(config_file, config_arg_str)
make_diagram_from_proto(config.model_config, dot_file)
