def train():
loss = net['cent'] + tf.add_n(net['kl']) / float(N) + net['wd']
global_step = tf.train.get_or_create_global_step()
bdr = [int(n_train_batches * (args.n_epochs - 1) * r) for r in [0.5, 0.75]]
vals = [1e-2, 1e-3, 1e-4]
lr = tf.train.piecewise_constant(tf.cast(global_step, tf.int32), bdr, vals)
train_op1 = tf.train.AdamOptimizer(lr).minimize(loss,
var_list=net['qpi_vars'],
global_step=global_step)
train_op2 = tf.train.AdamOptimizer(0.1 * lr).minimize(
loss, var_list=net['weights'])
train_op = tf.group(train_op1, train_op2)
pretrain_saver = tf.train.Saver(net['weights'])
saver = tf.train.Saver(net['weights'] + net['qpi_vars'])
logfile = open(os.path.join(savedir, 'train.log'), 'w', 0)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
pretrain_saver.restore(sess, os.path.join(pretraindir, 'model'))
train_logger = Accumulator('cent', 'acc')
train_to_run = [train_op, net['cent'], net['acc']]
test_logger = Accumulator('cent', 'acc')
test_to_run = [tnet['cent'], tnet['acc']]
for i in range(args.n_epochs):
line = 'Epoch %d start, learning rate %f' % (i + 1, sess.run(lr))
print(line)
logfile.write(line + '\n')
train_logger.clear()
start = time.time()
for j in range(n_train_batches):
bx, by = mnist.train.next_batch(batch_size)
train_logger.accum(sess.run(train_to_run, {x: bx, y: by}))
train_logger.print_(header='train',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
test_logger.clear()
for j in range(n_test_batches):
bx, by = mnist.test.next_batch(batch_size)
test_logger.accum(sess.run(test_to_run, {x: bx, y: by}))
test_logger.print_(header='test',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
line = 'kl: ' + str(sess.run(tnet['kl'])) + '\n'
line += 'n_active: ' + str(sess.run(tnet['n_active'])) + '\n'
print(line)
logfile.write(line + '\n')
if (i + 1) % args.save_freq == 0:
saver.save(sess, os.path.join(savedir, 'model'))
logfile.close()
saver.save(sess, os.path.join(savedir, 'model'))
def train():
import random
loss = -net['elbo'] + net['proto_loss'] * args.lamb
global_step = tf.train.get_or_create_global_step()
lr_step = args.n_trn_epsd / 3
lr = tf.train.piecewise_constant(tf.cast(global_step,
tf.int32), [lr_step, lr_step * 2],
[1e-3, 1e-3 * 0.5, 1e-3 * 0.5 * 0.5])
train_op = tf.train.AdamOptimizer(lr).minimize(loss,
global_step=global_step)
saver = tf.train.Saver(net['weights'])
logfile = open(os.path.join(savedir, 'train.log'), 'w', 0)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# to_run
train_logger = Accumulator('elbo', 'proto_loss')
train_to_run = [train_op, net['elbo'], net['proto_loss']]
for i in range(args.n_trn_epsd):
#train feed_dict
cidx = random.sample(xrange(len(nxtr)), args.way)
fdtr = data_queue(args, xtr, nxtr, cidx)
# train
train_logger.clear()
start = time.time()
train_logger.accum(sess.run(train_to_run, feed_dict=fdtr))
if i % 100 == 0:
train_logger.print_(header='train',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
line = 'Epoch %d start, learning rate %f' % (i + 1, sess.run(lr))
print('\n' + line)
logfile.write('\n' + line + '\n')
accu = sess.run(net['acc'], feed_dict=fdtr)
print("test accu ", np.mean(accu))
saver.save(sess, os.path.join(savedir, 'model'))
logfile.close()
saver.save(sess, os.path.join(savedir, 'model'))
def train():
loss = net['cent'] + net['wd']
global_step = tf.train.get_or_create_global_step()
lr = tf.train.piecewise_constant(tf.cast(global_step, tf.int32),
[n_train_batches * args.n_epochs / 2],
[1e-4, 1e-5])
train_op = tf.train.AdamOptimizer(lr).minimize(loss,
global_step=global_step)
saver = tf.train.Saver(net['weights'])
logfile = open(os.path.join(savedir, 'train.log'), 'w', 0)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
train_logger = Accumulator('cent', 'acc')
train_to_run = [train_op, net['cent'], net['acc']]
test_logger = Accumulator('cent', 'acc')
test_to_run = [tnet['cent'], tnet['acc']]
for i in range(args.n_epochs):
line = 'Epoch %d start, learning rate %f' % (i + 1, sess.run(lr))
print(line)
logfile.write(line + '\n')
train_logger.clear()
start = time.time()
for j in range(n_train_batches):
bx, by = mnist.train.next_batch(batch_size)
train_logger.accum(sess.run(train_to_run, {x: bx, y: by}))
train_logger.print_(header='train',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
test_logger.clear()
for j in range(n_test_batches):
bx, by = mnist.test.next_batch(batch_size)
test_logger.accum(sess.run(test_to_run, {x: bx, y: by}))
test_logger.print_(header='test',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
print()
logfile.write('\n')
if (i + 1) % args.save_freq == 0:
saver.save(sess, os.path.join(savedir, 'model'))
logfile.close()
saver.save(sess, os.path.join(savedir, 'model'))
def test():
import random
sess = tf.Session()
saver = tf.train.Saver(tnet['weights'])
saver.restore(sess, os.path.join(savedir, 'model'))
test_logger = Accumulator('elbo', 'proto_loss', 'acc')
test_to_run = [tnet['elbo'], tnet['proto_loss'], tnet['acc']]
test_logger.clear()
for i in range(args.n_tst_epsd):
#train feed_dict
cidx = random.sample(xrange(len(nxte)), args.way)
fdte = data_queue(args, xte, nxte, cidx)
# test
test_logger.accum(sess.run(test_to_run, feed_dict=fdte))
if (i + 1) % 100 == 0:
test_logger.print_(header='test', epoch=i + 1)
def train():
loss = net['cent'] + tf.add_n(net['kl']) / float(N) + net['wd']
global_step = tf.train.get_or_create_global_step()
bdr = [int(n_train_batches * (args.n_epochs - 1) * r) for r in [0.5, 0.75]]
vals = [1e-2, 1e-3, 1e-4]
lr = tf.train.piecewise_constant(tf.cast(global_step, tf.int32), bdr, vals)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op1 = tf.train.AdamOptimizer(lr).minimize(
loss, var_list=net['pzx_vars'], global_step=global_step)
train_op2 = tf.train.AdamOptimizer(0.1 * lr).minimize(
loss, var_list=net['weights'])
train_op = tf.group(train_op1, train_op2)
pretrain_saver = tf.train.Saver(net['weights'] + net['qpi_vars'])
saver = tf.train.Saver(net['weights'] + net['qpi_vars'] + net['pzx_vars'])
logfile = open(os.path.join(savedir, 'train.log'), 'w', 0)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
pretrain_saver.restore(sess, os.path.join(pretraindir, 'model'))
train_logger = Accumulator('cent', 'acc')
train_to_run = [train_op, net['cent'], net['acc']]
test_logger = Accumulator('cent', 'acc')
test_to_run = [tnet['cent'], tnet['acc']] + tnet['n_active']
for i in range(args.n_epochs):
line = 'Epoch %d start, learning rate %f' % (i + 1, sess.run(lr))
print(line)
logfile.write(line + '\n')
train_logger.clear()
start = time.time()
for j in range(n_train_batches):
bx, by = mnist.train.next_batch(batch_size)
train_logger.accum(sess.run(train_to_run, {x: bx, y: by}))
train_logger.print_(header='train',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
test_logger.clear()
np_n_active = [0] * n_drop
for j in range(n_test_batches):
bx, by = mnist.test.next_batch(batch_size)
res = sess.run(test_to_run, {x: bx, y: by})
test_logger.accum(res[:-n_drop])
np_n_active = [a + b for a, b in zip(np_n_active, res[-n_drop:])]
test_logger.print_(header='test',
epoch=i + 1,
time=time.time() - start,
logfile=logfile)
np_n_active = [int(a / n_test_batches) for a in np_n_active]
line = 'kl: ' + str(sess.run(tnet['kl'])) + '\n'
line += 'n_active: ' + str(np_n_active) + '\n'
print(line)
logfile.write(line + '\n')
if (i + 1) % args.save_freq == 0:
saver.save(sess, os.path.join(savedir, 'model'))
if (i + 1) % args.vis_freq == 0:
fig = _visualize(sess)
fig.savefig(os.path.join(figdir, 'epoch%d.png' % (i + 1)), dpi=200)
saver.save(sess, os.path.join(savedir, 'model'))
logfile.close()