from __future__ import absolute_import, division, print_function, unicode_literals
from builtins import (object)
import os
import argparse
import logging
import tensorflow as tf
logger = logging.getLogger(__name__)
# Hyper parameters
class hyper(object):
# operation
evaluation = False
warm_up = False
# shapes
node_type_num = 40 # total number of node types
word_dim = 40 # dimension of the feature vector for each node
conv_dim = 40 # number of conv feature detectors
fc_dim = 1024 # dimension of fc output
output_dim = 2 # number of output classes
# structure
use_relu = False
# learning
learning_rate = 0.00002 # learning rate
weight_decay = 0.0002
batch_size = 128
num_epochs = 100
# directories
log_dir = '/tmp/workspace/tf_log'
train_dir = '/tmp/workspace/tf_log'
embedding_dir = '/tmp/workspace/embedding_checkpoints/model'
# variables
variable_scope = ''
@classmethod
def dump(clz):
d = {k: v for k, v in vars(clz).items() if not k.startswith('__')}
with open(os.path.join(clz.work_dir, 'params.txt'), 'w') as f:
print(d, file=f)
@classmethod
def initialize(clz, from_cmd=True, **kwargs):
if from_cmd:
parser = argparse.ArgumentParser()
parser.add_argument('--embedding_dir',
help="""Saved embedding metrix,
defaults to /tmp/workspace/embedding_checkpoints/model
Only used when train tbcnn.
""",
default='/tmp/workspace/embedding_checkpoints/model')
parser.add_argument('--work_dir',
help='directory for saving files, defaults to /tmp/workspace/tf',
default='/tmp/workspace/tflogs')
parser.add_argument('--log_dir',
help='directory for tensorboard logs, defaults to WORK_DIR/logs')
parser.add_argument('--train_dir',
help='directory for model checkpoints, defaults to WORK_DIR/model')
parser.add_argument('--num_epochs', help='total number of epochs', type=int, default=100)
parser.add_argument('--batch_size', help='batch size', type=int, default=128)
parser.add_argument('--learning_rate', help='learning rate', type=float, default=0.0002)
parser.add_argument('--weight_decay', help='weight decay factor', type=float, default=0.0002)
parser.add_argument('--word_dim', help='dimension of node feature', type=int, default=100)
parser.add_argument('--conv_dim',
help='dimension of conv feature detectors', type=int, default=50)
parser.add_argument('--fc_dim',
help='dimension of fc layer', type=int, default=1024)
parser.add_argument('--output_dim',
help='number of output classes, default to 2. Should not need to change',
type=int, default=2)
parser.add_argument('--node_type_num', help='total number of node types', type=int, default=20)
parser.add_argument('--use_relu', type=bool, default=False)
parser.add_argument('--evaluation', help='Evaluating rather than training', action='store_true')
parser.add_argument('--warm_up', help='Do warming_up before evaluation, used for timing',
action='store_true')
args = parser.parse_args()
if not os.path.exists(args.work_dir):
os.makedirs(args.work_dir)
if args.log_dir is None:
args.log_dir = os.path.join(args.work_dir, 'logs')
if args.train_dir is None:
args.train_dir = os.path.join(args.work_dir, 'model')
for k, v in vars(args).items():
setattr(clz, k, v)
for k, v in kwargs.items():
setattr(clz, k, v)
# weights management
class param(object):
@classmethod
def create_variable(clz, name, shape, initializer, dtype=tf.float32, trainable=True):
with tf.variable_scope(hyper.variable_scope):
return tf.get_variable(name, shape=shape, dtype=dtype,
trainable=trainable, initializer=initializer)
@classmethod
def get(clz, name):
with tf.variable_scope(hyper.variable_scope, reuse=True):
return tf.get_variable(name)
@classmethod
def initialize_embedding_weights(clz):
clz.create_variable('Wl', (hyper.word_dim, hyper.word_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('Wr', (hyper.word_dim, hyper.word_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('B', (hyper.word_dim,),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('We', (hyper.node_type_num, hyper.word_dim),
tf.random_uniform_initializer(-.2, .2))
@classmethod
def initialize_tbcnn_weights(clz):
clz.initialize_embedding_weights()
# Don't train We
tf.get_collection_ref(tf.GraphKeys.TRAINABLE_VARIABLES).remove(clz.get('We'))
clz.create_variable('Wcomb1', (hyper.word_dim, hyper.word_dim),
tf.constant_initializer(-.2, .2))
clz.create_variable('Wcomb2', (hyper.word_dim, hyper.word_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('Wconvt', (hyper.word_dim, hyper.conv_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('Wconvl', (hyper.word_dim, hyper.conv_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('Wconvr', (hyper.word_dim, hyper.conv_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('Bconv', (hyper.conv_dim,),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('FC1/weight', (hyper.conv_dim, hyper.fc_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('FC1/bias', (hyper.fc_dim,),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('FC2/weight', (hyper.fc_dim, hyper.output_dim),
tf.random_uniform_initializer(-.2, .2))
clz.create_variable('FC2/bias', (hyper.output_dim, ),
tf.random_uniform_initializer(-.2, .2))
