def main(args):
info('loading embedding vec')
embedding = np.load(os.path.expanduser(args.embedding))
info('constructing config')
cfg = config(args, embedding)
info('constructing graph')
env = build_graph(cfg)
info('initializing session')
sess = tf.Session()
sess.run(tf.global_variables_initializer(),
feed_dict={cfg.embedding: embedding})
sess.run(tf.local_variables_initializer())
env.sess = sess
info('loading data')
(X_train, y_train), (X_test,
y_test), (X_valid,
y_valid) = load_data(cfg.data, cfg.bipolar)
info('training model')
train(env,
X_train,
y_train,
X_valid,
y_valid,
load=False,
batch_size=cfg.batch_size,
epochs=cfg.epochs,
name=cfg.name)
evaluate(env, X_test, y_test, batch_size=cfg.batch_size)
env.sess.close()
def main(args):
info('loading embedding vec')
embedding = np.load(os.path.expanduser(args.embedding))
info('constructing config')
cfg = config(args, embedding)
info('constructing graph')
env = build_graph(cfg)
env.cfg = cfg
info('initializing session')
sess = tf.Session()
sess.run(tf.global_variables_initializer(),
feed_dict={cfg.embedding: embedding})
sess.run(tf.local_variables_initializer())
env.sess = sess
info('loading data')
(_, _), (X_data, y_data) = load_data(cfg.data, cfg.bipolar, -1)
info('loading model')
train(env, load=True, name=cfg.name)
info('evaluating against clean test samples')
evaluate(env, X_data, y_data, batch_size=cfg.batch_size)
env.re = ReverseEmbedding(w2v_file=cfg.w2v, index_file=cfg.indexer)
info('making adversarial texts')
X_adv, X_sents = make_adversarial(env, X_data)
info('evaluating against adversarial texts')
evaluate(env, X_adv, y_data, batch_size=cfg.batch_size)
y_adv = predict(env, X_adv, batch_size=cfg.batch_size)
env.sess.close()
postfn(cfg, X_sents, y_data, y_adv)
def main(args):
info('loading embedding vec')
embedding = np.eye(args.vocab_size).astype(np.float32)
info('constructing config')
cfg = config(args, embedding)
info('constructing graph')
env = build_graph(cfg)
env.cfg = cfg
info('initializing session')
sess = tf.Session()
sess.run(tf.global_variables_initializer(),
feed_dict={cfg.embedding: embedding})
sess.run(tf.local_variables_initializer())
env.sess = sess
info('loading data')
(_, _), (X_data, y_data) = load_data(os.path.expanduser(cfg.data),
cfg.bipolar, validation_split=-1)
info('loading model')
train(env, load=True, name=cfg.name)
info('evaluating against clean test samples')
evaluate(env, X_data, y_data, batch_size=cfg.batch_size)
info('making adversarial texts')
X_adv = make_adversarial(env, X_data)
X_adv = np.reshape(X_adv, (-1, cfg.charlen))
y_data = np.tile(y_data, (cfg.beam_width, 1))
info('evaluating against adversarial texts')
evaluate(env, X_adv, y_data, batch_size=cfg.batch_size)
y_adv = predict(env, X_adv, batch_size=cfg.batch_size)
env.sess.close()
info('recover chars from indices')
X_sents = index2char(X_adv, unk=cfg.unk)
postfn(cfg, X_sents, y_data, y_adv)
def main(args):
info('loading embedding vec')
embedding = np.eye(args.vocab_size).astype(np.float32)
info('constructing config')
cfg = config(args, embedding)
info('constructing graph')
env = build_graph(cfg)
info('initializing session')
sess = tf.Session()
sess.run(tf.global_variables_initializer(),
feed_dict={cfg.embedding: embedding})
sess.run(tf.local_variables_initializer())
env.sess = sess
info('loading data {}'.format(cfg.data))
X_test, y_test = load_data(cfg.data, cfg.bipolar, -1)
info('training model')
train(env, load=True, name=cfg.name)
evaluate(env, X_test, y_test, batch_size=cfg.batch_size)
env.sess.close()
#NOTE: main loop for training
if __name__ == "__main__":
# Option
opt = parse_option(print_option=True)
# Data Loader
dataset_trn, dataset_val = get_dataloader(opt)
# Network
net = create_model(opt)
# Loss Function
criterion = get_loss_function(opt)
# Optimizer
optimizer = get_optimizer(net, criterion, opt)
scheduler = CosineAnnealingLR(optimizer, eta_min=opt.lr*opt.eta_min_ratio, T_max=(opt.max_epoch - opt.lr_warmup_epoch))
# Initial Best Score
best_dice, best_epoch = [0, 0]
for epoch in range(opt.start_epoch, opt.max_epoch):
# Train
train(net, dataset_trn, optimizer, criterion, epoch, opt)
# Evaluate
best_dice, best_epoch = validate(dataset_val, net, criterion, optimizer, epoch, opt, best_dice, best_epoch)
lr_update(epoch, opt, optimizer, scheduler)
print('Training done')