def choose_degree(seq_len, rng):
degree_probabilities = []
for i in range(1, seq_len + 1):
degree_probabilities.append(1.0 / i)
degree_chooser = RandomSampler(degree_probabilities)
return degree_chooser.next(lambda: rng.next_double()) + 1
def test_random_sampler(self):
probs = [ 1, 2, 4, 8 ]
sampler = RandomSampler(probs)
rng = Xoshiro256.from_string("Wolf")
samples = []
f = lambda: rng.next_double()
for i in range(500):
samples.append(sampler.next(f))
expected_samples = [3, 3, 3, 3, 3, 3, 3, 0, 2, 3, 3, 3, 3, 1, 2, 2, 1, 3, 3, 2, 3, 3, 1, 1, 2, 1, 1, 3, 1, 3, 1, 2, 0, 2, 1, 0, 3, 3, 3, 1, 3, 3, 3, 3, 1, 3, 2, 3, 2, 2, 3, 3, 3, 3, 2, 3, 3, 0, 3, 3, 3, 3, 1, 2, 3, 3, 2, 2, 2, 1, 2, 2, 1, 2, 3, 1, 3, 0, 3, 2, 3, 3, 3, 3, 3, 3, 3, 3, 2, 3, 1, 3, 3, 2, 0, 2, 2, 3, 1, 1, 2, 3, 2, 3, 3, 3, 3, 2, 3, 3, 3, 3, 3, 2, 3, 1, 2, 1, 1, 3, 1, 3, 2, 2, 3, 3, 3, 1, 3, 3, 3, 3, 3, 3, 3, 3, 2, 3, 2, 3, 3, 1, 2, 3, 3, 1, 3, 2, 3, 3, 3, 2, 3, 1, 3, 0, 3, 2, 1, 1, 3, 1, 3, 2, 3, 3, 3, 3, 2, 0, 3, 3, 1, 3, 0, 2, 1, 3, 3, 1, 1, 3, 1, 2, 3, 3, 3, 0, 2, 3, 2, 0, 1, 3, 3, 3, 2, 2, 2, 3, 3, 3, 3, 3, 2, 3, 3, 3, 3, 2, 3, 3, 2, 0, 2, 3, 3, 3, 3, 2, 1, 1, 1, 2, 1, 3, 3, 3, 2, 2, 3, 3, 1, 2, 3, 0, 3, 2, 3, 3, 3, 3, 0, 2, 2, 3, 2, 2, 3, 3, 3, 3, 1, 3, 2, 3, 3, 3, 3, 3, 2, 2, 3, 1, 3, 0, 2, 1, 3, 3, 3, 3, 3, 3, 3, 3, 1, 3, 3, 3, 3, 2, 2, 2, 3, 1, 1, 3, 2, 2, 0, 3, 2, 1, 2, 1, 0, 3, 3, 3, 2, 2, 3, 2, 1, 2, 0, 0, 3, 3, 2, 3, 3, 2, 3, 3, 3, 3, 3, 2, 2, 2, 3, 3, 3, 3, 3, 1, 1, 3, 2, 2, 3, 1, 1, 0, 1, 3, 2, 3, 3, 2, 3, 3, 2, 3, 3, 2, 2, 2, 2, 3, 2, 2, 2, 2, 2, 1, 2, 3, 3, 2, 2, 2, 2, 3, 3, 2, 0, 2, 1, 3, 3, 3, 3, 0, 3, 3, 3, 3, 2, 2, 3, 1, 3, 3, 3, 2, 3, 3, 3, 2, 3, 3, 3, 3, 2, 3, 2, 1, 3, 3, 3, 3, 2, 2, 0, 1, 2, 3, 2, 0, 3, 3, 3, 3, 3, 3, 1, 3, 3, 2, 3, 2, 2, 3, 3, 3, 3, 3, 2, 2, 3, 3, 2, 2, 2, 1, 3, 3, 3, 3, 1, 2, 3, 2, 3, 3, 2, 3, 2, 3, 3, 3, 2, 3, 1, 2, 3, 2, 1, 1, 3, 3, 2, 3, 3, 2, 3, 3, 0, 0, 1, 3, 3, 2, 3, 3, 3, 3, 1, 3, 3, 0, 3, 2, 3, 3, 1, 3, 3, 3, 3, 3, 3, 3, 0, 3, 3, 2]
assert(samples == expected_samples)
def __init__(self, args, max_outstanding=64):
self.app = Flask(__name__)
self.mode = args.mode
if self.mode == 'validation':
f = h5py.File(args.inpath)
preds, labs, y = f['preds'].value, f['labs'].value, f['y'].value
sampler = ValidationSampler(preds=preds,
labs=labs,
y=y,
no_permute=args.no_permute)
elif self.mode == 'las':
sampler = SimpleLASSampler(
crow=args.inpath,
seeds=args.seeds if args.seeds else None,
prefix=args.img_dir,
n=args.n_return)
elif self.mode == 'random':
sampler = RandomSampler(crow=args.inpath, prefix=args.img_dir)
elif self.mode == 'uncertainty':
f = h5py.File(args.inpath)
X, y, labs = f['X'].value, f['y'].value, f['labs'].value
sampler = UncertaintySampler(X=X, y=y, labs=labs)
else:
raise Exception('TaglessServer: unknown mode %s' % args.mode,
file=sys.stderr)
self.sampler = sampler
self.app.add_url_rule('/', 'view_1', self.index)
self.app.add_url_rule('/<path:x>', 'view_2',
lambda x: send_file('/' + x))
self.app.add_url_rule('/label', 'view_3', self.label, methods=['POST'])
self.app.add_url_rule('/twitter_tag',
'view_4',
self.twitter_tag,
methods=['POST'])
self.n_las = args.n_las
self.outpath = args.outpath
self.sent = set([])
self.outstanding = 0
self.max_outstanding = max_outstanding
def save():
self.sampler.save(self.outpath)
atexit.register(save)
def __init__(self,
config,
num_gpus,
replicas,
optimizer,
global_step,
summary_writer=None):
"""Builds TF graph nodes for model updating (via _ModelUpdateGraph).
Args:
config: the model config (an argparse.Namespace)
num_gpus: the number of available GPUs.
replicas: a list of RNNModel or Transformer objects.
optimizer: a TensorFlow optimizer.
global_step: a tf.Variable to be updated by optimizer.
summary_writer: a tf.summary.FileWriter object.
"""
assert len(replicas) > 0
assert (len(replicas) == num_gpus
or (len(replicas) == 1 and num_gpus == 0))
self._config = config
self._replicas = replicas
self._summary_writer = summary_writer
self._graph = _ModelUpdateGraph(config, num_gpus, replicas, optimizer,
global_step)
if config.loss_function == 'MRT':
if config.sample_way == 'beam_search':
self._mrt_sampler = BeamSearchSampler(
models=[replicas[0]],
configs=[config],
beam_size=config.samplesN)
else:
assert config.sample_way == 'randomly_sample'
# Set beam_size to config.samplesN instead of using
# np.repeat to expand input in full_sampler()
self._mrt_sampler = RandomSampler(models=[replicas[0]],
configs=[config],
beam_size=config.samplesN)
def main(settings):
"""
Translates a source language file (or STDIN) into a target language file
(or STDOUT).
"""
# Create the TensorFlow session.
g = tf.Graph()
with g.as_default():
tf_config = tf.compat.v1.ConfigProto()
tf_config.allow_soft_placement = True
session = tf.compat.v1.Session(config=tf_config)
# Load config file for each model.
configs = []
for model in settings.models:
config = load_config_from_json_file(model)
setattr(config, 'reload', model)
setattr(config, 'translation_maxlen', settings.translation_maxlen)
configs.append(config)
# Create the model graphs.
logging.debug("Loading models\n")
models = []
for i, config in enumerate(configs):
with tf.compat.v1.variable_scope("model%d" % i) as scope:
if config.model_type == "transformer":
model = TransformerModel(
config, consts_config_str=settings.config_str)
else:
model = rnn_model.RNNModel(config)
model.sampling_utils = SamplingUtils(settings)
models.append(model)
# Add smoothing variables (if the models were trained with smoothing).
# FIXME Assumes either all models were trained with smoothing or none were.
if configs[0].exponential_smoothing > 0.0:
smoothing = ExponentialSmoothing(configs[0].exponential_smoothing)
# Restore the model variables.
for i, config in enumerate(configs):
with tf.compat.v1.variable_scope("model%d" % i) as scope:
_ = model_loader.init_or_restore_variables(
config, session, ensemble_scope=scope)
# Swap-in the smoothed versions of the variables.
if configs[0].exponential_smoothing > 0.0:
session.run(fetches=smoothing.swap_ops)
max_translation_len = settings.translation_maxlen
# Create a BeamSearchSampler / RandomSampler.
if settings.translation_strategy == 'beam_search':
sampler = BeamSearchSampler(models, configs, settings.beam_size)
else:
assert settings.translation_strategy == 'sampling'
sampler = RandomSampler(models, configs, settings.beam_size)
# Warn about the change from neg log probs to log probs for the RNN.
if settings.n_best:
model_types = [config.model_type for config in configs]
if 'rnn' in model_types:
logging.warn(
'n-best scores for RNN models have changed from '
'positive to negative (as of commit 95793196...). '
'If you are using the scores for reranking etc, then '
'you may need to update your scripts.')
# Translate the source file.
translate_utils.translate_file(
input_file=settings.input,
output_file=settings.output,
session=session,
sampler=sampler,
config=configs[0],
max_translation_len=max_translation_len,
normalization_alpha=settings.normalization_alpha,
consts_config_str=settings.config_str,
nbest=settings.n_best,
minibatch_size=settings.minibatch_size,
maxibatch_size=settings.maxibatch_size)
def train(config, sess):
assert (config.prior_model != None and (tf.train.checkpoint_exists(os.path.abspath(config.prior_model))) or (config.map_decay_c==0.0)), \
"MAP training requires a prior model file: Use command-line option --prior_model"
# Construct the graph, with one model replica per GPU
num_gpus = len(tf_utils.get_available_gpus())
num_replicas = max(1, num_gpus)
if config.loss_function == 'MRT':
assert config.gradient_aggregation_steps == 1
assert config.max_sentences_per_device == 0, "MRT mode does not support sentence-based split"
if config.max_tokens_per_device != 0:
assert (config.samplesN * config.maxlen <= config.max_tokens_per_device), "need to make sure candidates of a sentence could be " \
"feed into the model"
else:
assert num_replicas == 1, "MRT mode does not support sentence-based split"
assert (config.samplesN * config.maxlen <= config.token_batch_size), "need to make sure candidates of a sentence could be " \
"feed into the model"
logging.info('Building model...')
replicas = []
for i in range(num_replicas):
device_type = "GPU" if num_gpus > 0 else "CPU"
device_spec = tf.DeviceSpec(device_type=device_type, device_index=i)
with tf.device(device_spec):
with tf.variable_scope(tf.get_variable_scope(), reuse=(i>0)):
if config.model_type == "transformer":
model = TransformerModel(config)
else:
model = rnn_model.RNNModel(config)
replicas.append(model)
init = tf.zeros_initializer(dtype=tf.int32)
global_step = tf.get_variable('time', [], initializer=init, trainable=False)
if config.learning_schedule == "constant":
schedule = learning_schedule.ConstantSchedule(config.learning_rate)
elif config.learning_schedule == "transformer":
schedule = learning_schedule.TransformerSchedule(
global_step=global_step,
dim=config.state_size,
warmup_steps=config.warmup_steps)
elif config.learning_schedule == "warmup-plateau-decay":
schedule = learning_schedule.WarmupPlateauDecaySchedule(
global_step=global_step,
peak_learning_rate=config.learning_rate,
warmup_steps=config.warmup_steps,
plateau_steps=config.plateau_steps)
else:
logging.error('Learning schedule type is not valid: {}'.format(
config.learning_schedule))
sys.exit(1)
if config.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=schedule.learning_rate,
beta1=config.adam_beta1,
beta2=config.adam_beta2,
epsilon=config.adam_epsilon)
else:
logging.error('No valid optimizer defined: {}'.format(config.optimizer))
sys.exit(1)
if config.summary_freq:
summary_dir = (config.summary_dir if config.summary_dir is not None
else os.path.abspath(os.path.dirname(config.saveto)))
writer = tf.summary.FileWriter(summary_dir, sess.graph)
else:
writer = None
updater = ModelUpdater(config, num_gpus, replicas, optimizer, global_step,
writer)
if config.exponential_smoothing > 0.0:
smoothing = ExponentialSmoothing(config.exponential_smoothing)
saver, progress = model_loader.init_or_restore_variables(
config, sess, train=True)
global_step.load(progress.uidx, sess)
if config.sample_freq:
random_sampler = RandomSampler(
models=[replicas[0]],
configs=[config],
beam_size=1)
if config.beam_freq or config.valid_script is not None:
beam_search_sampler = BeamSearchSampler(
models=[replicas[0]],
configs=[config],
beam_size=config.beam_size)
#save model options
write_config_to_json_file(config, config.saveto)
text_iterator, valid_text_iterator = load_data(config)
_, _, num_to_source, num_to_target = util.load_dictionaries(config)
total_loss = 0.
n_sents, n_words = 0, 0
last_time = time.time()
logging.info("Initial uidx={}".format(progress.uidx))
# set epoch = 1 if print per-token-probability
if config.print_per_token_pro:
config.max_epochs = progress.eidx+1
for progress.eidx in range(progress.eidx, config.max_epochs):
logging.info('Starting epoch {0}'.format(progress.eidx))
for source_sents, target_sents in text_iterator:
if len(source_sents[0][0]) != config.factors:
logging.error('Mismatch between number of factors in settings ({0}), and number in training corpus ({1})\n'.format(config.factors, len(source_sents[0][0])))
sys.exit(1)
x_in, x_mask_in, y_in, y_mask_in = util.prepare_data(
source_sents, target_sents, config.factors, maxlen=None)
if x_in is None:
logging.info('Minibatch with zero sample under length {0}'.format(config.maxlen))
continue
write_summary_for_this_batch = config.summary_freq and ((progress.uidx % config.summary_freq == 0) or (config.finish_after and progress.uidx % config.finish_after == 0))
(factors, seqLen, batch_size) = x_in.shape
output = updater.update(
sess, x_in, x_mask_in, y_in, y_mask_in, num_to_target,
write_summary_for_this_batch)
if config.print_per_token_pro == False:
total_loss += output
else:
# write per-token probability into the file
f = open(config.print_per_token_pro, 'a')
for pro in output:
pro = str(pro) + '\n'
f.write(pro)
f.close()
n_sents += batch_size
n_words += int(numpy.sum(y_mask_in))
progress.uidx += 1
# Update the smoothed version of the model variables.
# To reduce the performance overhead, we only do this once every
# N steps (the smoothing factor is adjusted accordingly).
if config.exponential_smoothing > 0.0 and progress.uidx % smoothing.update_frequency == 0:
sess.run(fetches=smoothing.update_ops)
if config.disp_freq and progress.uidx % config.disp_freq == 0:
duration = time.time() - last_time
disp_time = datetime.now().strftime('[%Y-%m-%d %H:%M:%S]')
logging.info('{0} Epoch: {1} Update: {2} Loss/word: {3} Words/sec: {4} Sents/sec: {5}'.format(disp_time, progress.eidx, progress.uidx, total_loss/n_words, n_words/duration, n_sents/duration))
last_time = time.time()
total_loss = 0.
n_sents = 0
n_words = 0
if config.sample_freq and progress.uidx % config.sample_freq == 0:
x_small = x_in[:, :, :10]
x_mask_small = x_mask_in[:, :10]
y_small = y_in[:, :10]
samples = translate_utils.translate_batch(
sess, random_sampler, x_small, x_mask_small,
config.translation_maxlen, 0.0)
assert len(samples) == len(x_small.T) == len(y_small.T), \
(len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
sample = util.seq2words(ss[0][0], num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
logging.info('SAMPLE: {}'.format(sample))
if config.beam_freq and progress.uidx % config.beam_freq == 0:
x_small = x_in[:, :, :10]
x_mask_small = x_mask_in[:, :10]
y_small = y_in[:,:10]
samples = translate_utils.translate_batch(
sess, beam_search_sampler, x_small, x_mask_small,
config.translation_maxlen, config.normalization_alpha)
assert len(samples) == len(x_small.T) == len(y_small.T), \
(len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
for i, (sample_seq, cost) in enumerate(ss):
sample = util.seq2words(sample_seq, num_to_target)
msg = 'SAMPLE {}: {} Cost/Len/Avg {}/{}/{}'.format(
i, sample, cost, len(sample), cost/len(sample))
logging.info(msg)
if config.valid_freq and progress.uidx % config.valid_freq == 0:
if config.exponential_smoothing > 0.0:
sess.run(fetches=smoothing.swap_ops)
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
sess.run(fetches=smoothing.swap_ops)
else:
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
if (len(progress.history_errs) == 0 or
valid_ce < min(progress.history_errs)):
progress.history_errs.append(valid_ce)
progress.bad_counter = 0
save_non_checkpoint(sess, saver, config.saveto)
progress_path = '{0}.progress.json'.format(config.saveto)
progress.save_to_json(progress_path)
else:
progress.history_errs.append(valid_ce)
progress.bad_counter += 1
if progress.bad_counter > config.patience:
logging.info('Early Stop!')
progress.estop = True
break
if config.valid_script is not None:
if config.exponential_smoothing > 0.0:
sess.run(fetches=smoothing.swap_ops)
score = validate_with_script(sess, beam_search_sampler)
sess.run(fetches=smoothing.swap_ops)
else:
score = validate_with_script(sess, beam_search_sampler)
need_to_save = (score is not None and
(len(progress.valid_script_scores) == 0 or
score > max(progress.valid_script_scores)))
if score is None:
score = 0.0 # ensure a valid value is written
progress.valid_script_scores.append(score)
if need_to_save:
progress.bad_counter = 0
save_path = config.saveto + ".best-valid-script"
save_non_checkpoint(sess, saver, save_path)
write_config_to_json_file(config, save_path)
progress_path = '{}.progress.json'.format(save_path)
progress.save_to_json(progress_path)
if config.save_freq and progress.uidx % config.save_freq == 0:
saver.save(sess, save_path=config.saveto, global_step=progress.uidx)
write_config_to_json_file(config, "%s-%s" % (config.saveto, progress.uidx))
progress_path = '{0}-{1}.progress.json'.format(config.saveto, progress.uidx)
progress.save_to_json(progress_path)
if config.finish_after and progress.uidx % config.finish_after == 0:
logging.info("Maximum number of updates reached")
saver.save(sess, save_path=config.saveto, global_step=progress.uidx)
write_config_to_json_file(config, "%s-%s" % (config.saveto, progress.uidx))
progress.estop=True
progress_path = '{0}-{1}.progress.json'.format(config.saveto, progress.uidx)
progress.save_to_json(progress_path)
break
if progress.estop:
break
self.imgs[i])).convert('RGB')
img = self.transforms(img)
return (img, self.ids[i])
transforms_ = transforms.Compose([
transforms.Resize(228),
transforms.RandomCrop((224, 224)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
data = Fashion_inshop(type="train", transform=transforms_)
sampler = RandomSampler(data, args.batch)
batch_sampler = BatchSampler(sampler, args.batch, True)
dataset = torch.utils.data.DataLoader(
data,
batch_sampler=batch_sampler,
#batch_sampler=BalancedBatchSampler(data, batch_size=data.number_of_classes, batch_k=args.batch_k, length=args.num_batch),
num_workers=args.num_workers)
data_test = Fashion_inshop(type="test", transform=transforms_)
test_sampler = RandomSampler(data_test, args.batch)
test_batch_sampler = BatchSampler(test_sampler, args.batch, True)
dataset_test = torch.utils.data.DataLoader(
data_test,
batch_sampler=test_batch_sampler,
#batch_sampler=BalancedBatchSampler(data_test, batch_size=data.number_of_classes, batch_k=args.batch_k, length=args.num_batch//2)