def parse_args_and_setup(self):
defaults = self.defaults
args = self.parser.parse_args()
args_dict = vars(args)
# Convert 0/1 to boolean, keep None
for argname in self.boolean_args:
if not args_dict[argname] is None:
args_dict[argname] = bool(args_dict[argname])
self.args = args
# Process while defaults not set yet
if args.shuffle_examples is None and not args.shuffle_examples_capacity is None:
raise ValueError(
'Shuffle capacity given without --shuffle_examples')
if args.encode_lstm is None and not args.encode_hidden_units is None:
raise ValueError('Encode hidden units given without --encode_lstm')
# In order to make it possible to save the absence of a value
# (args.cson), allow for 'none' and convert to None
# Kinda silly since a lot of code works with 'none', but oh well
# none_updates = {}
# for k, v in args.__dict__.items():
# if isinstance(v, str) and v.lower() == 'none':
# none_updates[v] = None
# args.__dict__.update(none_updates)
defaults.update(
{k: v
for (k, v) in args.__dict__.items() if v is not None})
args.__dict__.update(defaults)
args.debug = self.debug
if args.gen_log_dir:
args.log_dir = self._gen_log_dir(args.log_dir, args.prefix)
else:
os.makedirs(args.log_dir, exist_ok=True)
self.setup_log(args.log_dir)
log = Log.get_logger(__name__)
if not args.gen_log_dir and args.restore_if_possible is None:
args.restore_if_possible = True
log.info('Restoring if possible since a log directory is provided')
if 'LD_PRELOAD' in os.environ and 'tcmalloc' in os.environ[
'LD_PRELOAD']:
log.info('Using tcmalloc, good on you!')
else:
log.warning('NOT using tcmalloc')
import tensorflow as tf
import numpy as np
from time import time
from collections import OrderedDict
from .util import *
from attend.log import Log; log = Log.get_logger(__name__)
from attend import util, tf_util
import attend
class AttendSolver():
def __init__(self, model, update_rule, learning_rate, stats_every):
self.model = model
self.update_rule = update_rule
self.learning_rate = learning_rate
self.summary_producer = None
if self.update_rule == 'adam':
self.optimizer = tf.train.AdamOptimizer
else:
raise Exception()
self.loss_names = ['mse', 'pearson_r', 'icc']
from attend import SummaryProducer
self.summary_producer = SummaryProducer(self.loss_names)
self.stats_every = stats_every
# TODO
# This will contain the whole validation set, for losses
import tensorflow as tf
import numpy as np
import functools
from attend.log import Log
log = Log.get_logger(__name__)
class Encoder():
ALLOWED_CONV_IMPLS = ['small', 'convnet', 'resnet', 'vggface', 'none']
ACTIVATIONS = dict(relu=tf.nn.relu)
def __init__(self,
batch_size,
encode_hidden_units=0,
time_steps=None,
debug=True,
conv_impl=None,
dense_layer=0,
dropout=.75,
use_dropout=False,
use_maxnorm=False,
dense_spec=None,
encode_lstm=None,
lstm_impl=None,
use_batch_norm=True,
batch_norm_decay=None,
use_batch_renorm=False):
self.use_batch_norm = use_batch_norm
self.batch_norm_decay = batch_norm_decay
self.use_batch_renorm = use_batch_renorm
def save_interesting_things(self):
Log.save_args(self.args)
Log.save_pid()
Log.save_env()
Log.save_git_version()
Log.save_hostname()
Log.save_condor()
def setup_log(self, log_dir=None):
if log_dir is None:
log_dir = self.args.log_dir
from attend.log import Log
Log.setup(log_dir, self.debug)