# Copyright 2019 Google LLC # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # 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. """Unsupervised data augmentation (UDA) """ import functools import os import numpy as np import tensorflow as tf from absl import app from absl import flags from tqdm import trange from libml import models, utils from libml.data import PAIR_DATASETS FLAGS = flags.FLAGS class UDA(models.MultiModel): TSA_MODES = 'no exp linear log'.split() def tsa_threshold(self, tsa, scale=5, tsa_pos=10, **kwargs): del kwargs # step ratio will be maxed at (2 ** 14) * (2 ** 10) ~ 16.8M updates step_ratio = tf.to_float(self.step) / tf.to_float(min(FLAGS.train_kimg, 1 << 14) << tsa_pos) if tsa == 'linear': coeff = step_ratio elif tsa == 'exp': # [exp(-5), exp(0)] = [1e-2, 1] coeff = tf.exp((step_ratio - 1) * scale) elif tsa == 'log': # [1 - exp(0), 1 - exp(-5)] = [0, 0.99] coeff = 1 - tf.exp((-step_ratio) * scale) elif tsa == 'no': coeff = tf.to_float(1.0) elif tsa != 'no': raise NotImplementedError(tsa) coeff = tf.math.minimum(coeff, 1.0) # bound the coefficient p_min = 1. / self.nclass return coeff * (1 - p_min) + p_min def tsa_loss_mask(self, tsa, logits, labels, tsa_pos, **kwargs): thresh = self.tsa_threshold(tsa, tsa_pos=tsa_pos, **kwargs) p_class = tf.nn.softmax(logits, axis=-1) p_correct = tf.reduce_sum(labels * p_class, axis=-1) loss_mask = tf.cast(p_correct <= thresh, tf.float32) # Ignore confident predictions. return tf.stop_gradient(loss_mask) @staticmethod def confidence_based_masking(logits, p_class=None, thresh=0.9): if logits is not None: p_class = tf.nn.softmax(logits, axis=-1) p_class_max = tf.reduce_max(p_class, axis=-1) loss_mask = tf.cast(p_class_max >= thresh, tf.float32) # Ignore unconfident predictions. return tf.stop_gradient(loss_mask) @staticmethod def softmax_temperature_controlling(logits, T): # this is essentially the same as sharpening in mixmatch logits = logits / T return tf.stop_gradient(logits) @staticmethod def kl_divergence_from_logits(p_logits, q_logits): p = tf.nn.softmax(p_logits) log_p = tf.nn.log_softmax(p_logits) log_q = tf.nn.log_softmax(q_logits) kl = tf.reduce_sum(p * (log_p - log_q), -1) return kl @staticmethod def entropy_from_logits(logits): log_prob = tf.nn.log_softmax(logits, axis=-1) prob = tf.exp(log_prob) ent = tf.reduce_sum(-prob * log_prob, axis=-1) return ent def train(self, train_nimg, report_nimg): if FLAGS.eval_ckpt: self.eval_checkpoint(FLAGS.eval_ckpt) return batch = FLAGS.batch train_labeled = self.dataset.train_labeled.repeat().shuffle(FLAGS.shuffle).parse().augment() train_labeled = train_labeled.batch(batch).prefetch(16).make_one_shot_iterator().get_next() train_unlabeled = self.dataset.train_unlabeled.repeat().shuffle(FLAGS.shuffle).parse().augment() train_unlabeled = train_unlabeled.batch(batch * self.params['uratio']).prefetch(16) train_unlabeled = train_unlabeled.make_one_shot_iterator().get_next() scaffold = tf.train.Scaffold(saver=tf.train.Saver(max_to_keep=FLAGS.keep_ckpt, pad_step_number=10)) with tf.Session(config=utils.get_config()) as sess: self.session = sess self.cache_eval() with tf.train.MonitoredTrainingSession( scaffold=scaffold, checkpoint_dir=self.checkpoint_dir, config=utils.get_config(), save_checkpoint_steps=FLAGS.save_kimg << 10, save_summaries_steps=report_nimg - batch) as train_session: self.session = train_session._tf_sess() gen_labeled = self.gen_labeled_fn(train_labeled) gen_unlabeled = self.gen_unlabeled_fn(train_unlabeled) self.tmp.step = self.session.run(self.step) while self.tmp.step < train_nimg: loop = trange(self.tmp.step % report_nimg, report_nimg, batch, leave=False, unit='img', unit_scale=batch, desc='Epoch %d/%d' % (1 + (self.tmp.step // report_nimg), train_nimg // report_nimg)) for _ in loop: self.train_step(train_session, gen_labeled, gen_unlabeled) while self.tmp.print_queue: loop.write(self.tmp.print_queue.pop(0)) while self.tmp.print_queue: print(self.tmp.print_queue.pop(0)) def model(self, batch, lr, wd, wu, we, confidence, uratio, temperature=1.0, tsa='no', tsa_pos=10, ema=0.999, **kwargs): hwc = [self.dataset.height, self.dataset.width, self.dataset.colors] xt_in = tf.placeholder(tf.float32, [batch] + hwc, 'xt') # For training x_in = tf.placeholder(tf.float32, [None] + hwc, 'x') y_in = tf.placeholder(tf.float32, [batch * uratio, 2] + hwc, 'y') l_in = tf.placeholder(tf.int32, [batch], 'labels') l = tf.one_hot(l_in, self.nclass) lrate = tf.clip_by_value(tf.to_float(self.step) / (FLAGS.train_kimg << 10), 0, 1) lr *= tf.cos(lrate * (7 * np.pi) / (2 * 8)) tf.summary.scalar('monitors/lr', lr) # Compute logits for xt_in and y_in classifier = lambda x, **kw: self.classifier(x, **kw, **kwargs).logits skip_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) x = utils.interleave(tf.concat([xt_in, y_in[:, 0], y_in[:, 1]], 0), 2 * uratio + 1) logits = utils.para_cat(lambda x: classifier(x, training=True), x) logits = utils.de_interleave(logits, 2 * uratio+1) post_ops = [v for v in tf.get_collection(tf.GraphKeys.UPDATE_OPS) if v not in skip_ops] logits_x = logits[:batch] logits_weak, logits_strong = tf.split(logits[batch:], 2) del logits, skip_ops # softmax temperature control logits_weak_tgt = self.softmax_temperature_controlling(logits_weak, T=temperature) # generate confidence mask based on sharpened distribution pseudo_labels = tf.stop_gradient(tf.nn.softmax(logits_weak)) pseudo_mask = self.confidence_based_masking(logits=None, p_class=pseudo_labels, thresh=confidence) tf.summary.scalar('monitors/mask', tf.reduce_mean(pseudo_mask)) tf.summary.scalar('monitors/conf_weak', tf.reduce_mean(tf.reduce_max(tf.nn.softmax(logits_weak), axis=1))) tf.summary.scalar('monitors/conf_strong', tf.reduce_mean(tf.reduce_max(tf.nn.softmax(logits_strong), axis=1))) kld = self.kl_divergence_from_logits(logits_weak_tgt, logits_strong) entropy = self.entropy_from_logits(logits_weak) loss_xeu = tf.reduce_mean(kld * pseudo_mask) tf.summary.scalar('losses/xeu', loss_xeu) loss_ent = tf.reduce_mean(entropy) tf.summary.scalar('losses/entropy', loss_ent) # supervised loss with TSA loss_mask = self.tsa_loss_mask(tsa=tsa, logits=logits_x, labels=l, tsa_pos=tsa_pos) loss_xe = tf.nn.softmax_cross_entropy_with_logits_v2(labels=l, logits=logits_x) loss_xe = tf.reduce_sum(loss_xe * loss_mask) / tf.math.maximum(tf.reduce_sum(loss_mask), 1.0) tf.summary.scalar('losses/xe', loss_xe) tf.summary.scalar('losses/mask_sup', tf.reduce_mean(loss_mask)) # L2 regularization loss_wd = sum(tf.nn.l2_loss(v) for v in utils.model_vars('classify') if 'kernel' in v.name) tf.summary.scalar('losses/wd', loss_wd) ema = tf.train.ExponentialMovingAverage(decay=ema) ema_op = ema.apply(utils.model_vars()) ema_getter = functools.partial(utils.getter_ema, ema) post_ops.append(ema_op) train_op = tf.train.MomentumOptimizer(lr, 0.9, use_nesterov=True).minimize( loss_xe + loss_xeu * wu + loss_ent * we + loss_wd * wd, colocate_gradients_with_ops=True) with tf.control_dependencies([train_op]): train_op = tf.group(*post_ops) return utils.EasyDict( xt=xt_in, x=x_in, y=y_in, label=l_in, train_op=train_op, classify_raw=tf.nn.softmax(classifier(x_in, training=False)), # No EMA, for debugging. classify_op=tf.nn.softmax(classifier(x_in, getter=ema_getter, training=False))) def main(argv): utils.setup_main() del argv # Unused. dataset = PAIR_DATASETS()[FLAGS.dataset]() log_width = utils.ilog2(dataset.width) model = UDA( os.path.join(FLAGS.train_dir, dataset.name), dataset, lr=FLAGS.lr, wd=FLAGS.wd, wu=FLAGS.wu, we=FLAGS.we, arch=FLAGS.arch, batch=FLAGS.batch, nclass=dataset.nclass, temperature=FLAGS.temperature, tsa=FLAGS.tsa, tsa_pos=FLAGS.tsa_pos, confidence=FLAGS.confidence, uratio=FLAGS.uratio, scales=FLAGS.scales or (log_width - 2), filters=FLAGS.filters, repeat=FLAGS.repeat) model.train(FLAGS.train_kimg << 10, FLAGS.report_kimg << 10) # 1024 epochs if __name__ == '__main__': utils.setup_tf() flags.DEFINE_float('wu', 1, 'Consistency weight.') flags.DEFINE_float('wd', 0.0005, 'Weight decay.') flags.DEFINE_float('we', 0, 'Entropy minimization weight.') flags.DEFINE_float('ema', 0.999, 'Exponential moving average of params.') flags.DEFINE_float('confidence', 0.95, 'Confidence threshold.') flags.DEFINE_float('temperature', 1, 'Softmax sharpening temperature.') flags.DEFINE_integer('scales', 0, 'Number of 2x2 downscalings in the classifier.') flags.DEFINE_integer('filters', 32, 'Filter size of convolutions.') flags.DEFINE_integer('repeat', 4, 'Number of residual layers per stage.') flags.DEFINE_integer('tsa_pos', 8, 'TSA change rate.') flags.DEFINE_integer('uratio', 7, 'Unlabeled batch size ratio.') flags.DEFINE_enum('tsa', 'no', UDA.TSA_MODES, 'TSA mode.') FLAGS.set_default('augment', 'd.d.rac') FLAGS.set_default('dataset', 'cifar10.3@250-1') FLAGS.set_default('batch', 64) FLAGS.set_default('lr', 0.03) FLAGS.set_default('train_kimg', 1 << 16) app.run(main)