def main():
if FLAGS.datasource == 'sinusoid':
if FLAGS.train:
test_num_updates = 5
else:
test_num_updates = 10
else:
if FLAGS.datasource == 'miniimagenet' or FLAGS.datasource == 'cifar100':
if FLAGS.train == True:
test_num_updates = 1 # eval on at least one update during training
else:
test_num_updates = 10
else:
test_num_updates = 10
if FLAGS.train == False:
orig_meta_batch_size = FLAGS.meta_batch_size
# always use meta batch size of 1 when testing.
FLAGS.meta_batch_size = 1
if FLAGS.datasource == 'sinusoid':
data_generator = DataGenerator(FLAGS.update_batch_size * 2,
FLAGS.meta_batch_size)
else:
if FLAGS.metatrain_iterations == 0: #and FLAGS.datasource == 'miniimagenet':
assert FLAGS.meta_batch_size == 1
assert FLAGS.update_batch_size == 1
data_generator = DataGenerator(
1, FLAGS.meta_batch_size) # only use one datapoint,
else:
if FLAGS.datasource == 'miniimagenet': #or FLAGS.datasource == 'cifar100': # TODO - use 15 val examples for imagenet?
if FLAGS.train:
data_generator = DataGenerator(
FLAGS.update_batch_size + 15, FLAGS.meta_batch_size
) # only use one datapoint for testing to save memory
else:
data_generator = DataGenerator(
FLAGS.update_batch_size * 2, FLAGS.meta_batch_size
) # only use one datapoint for testing to save memory
else:
data_generator = DataGenerator(
FLAGS.update_batch_size * 2, FLAGS.meta_batch_size
) # only use one datapoint for testing to save memory
dim_output = data_generator.dim_output
if FLAGS.baseline == 'oracle':
assert FLAGS.datasource == 'sinusoid'
dim_input = 3
FLAGS.pretrain_iterations += FLAGS.metatrain_iterations
FLAGS.metatrain_iterations = 0
else:
dim_input = data_generator.dim_input
if FLAGS.datasource == 'miniimagenet' or FLAGS.datasource == 'omniglot' or FLAGS.datasource == 'cifar100':
tf_data_load = True
num_classes = data_generator.num_classes
if FLAGS.train: # only construct training model if needed
random.seed(5)
image_tensor, label_tensor = data_generator.make_data_tensor()
inputa = tf.slice(image_tensor, [0, 0, 0],
[-1, num_classes * FLAGS.update_batch_size, -1])
inputb = tf.slice(image_tensor,
[0, num_classes * FLAGS.update_batch_size, 0],
[-1, -1, -1])
labela = tf.slice(label_tensor, [0, 0, 0],
[-1, num_classes * FLAGS.update_batch_size, -1])
labelb = tf.slice(label_tensor,
[0, num_classes * FLAGS.update_batch_size, 0],
[-1, -1, -1])
input_tensors = {
'inputa': inputa,
'inputb': inputb,
'labela': labela,
'labelb': labelb
}
random.seed(6)
image_tensor, label_tensor = data_generator.make_data_tensor(
train=False)
inputa = tf.slice(image_tensor, [0, 0, 0],
[-1, num_classes * FLAGS.update_batch_size, -1])
inputb = tf.slice(image_tensor,
[0, num_classes * FLAGS.update_batch_size, 0],
[-1, -1, -1])
labela = tf.slice(label_tensor, [0, 0, 0],
[-1, num_classes * FLAGS.update_batch_size, -1])
labelb = tf.slice(label_tensor,
[0, num_classes * FLAGS.update_batch_size, 0],
[-1, -1, -1])
metaval_input_tensors = {
'inputa': inputa,
'inputb': inputb,
'labela': labela,
'labelb': labelb
}
else:
tf_data_load = False
input_tensors = None
random.seed(7)
X = data_generator.make_autoencoder_data_tensor(train=True)
Y = data_generator.make_autoencoder_data_tensor(train=False)
ims_for_encodings = data_generator.make_autoencoder_test_data_tensor(
train=True)
# print(X.get_shape())
autoencoder_input_tensors = {'X': X, 'Y': Y}
dim_s = 32
autoencoder = Autoencoder(dim_input, dim_s)
model = MAML(autoencoder,
dim_input,
dim_output,
test_num_updates=test_num_updates)
if FLAGS.train or not tf_data_load:
autoencoder.construct_autoencoder(
input_tensors=autoencoder_input_tensors,
prefix='autoencoder_train')
if tf_data_load:
model.construct_model(input_tensors=metaval_input_tensors,
prefix='metaval_')
model.summ_op = tf.summary.merge_all()
autoencoder.summ_op = tf.summary.merge_all()
saver = loader = tf.train.Saver(tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES),
max_to_keep=10)
sess = tf.InteractiveSession()
if FLAGS.train == False:
# change to original meta batch size when loading model.
FLAGS.meta_batch_size = orig_meta_batch_size
if FLAGS.train_update_batch_size == -1:
FLAGS.train_update_batch_size = FLAGS.update_batch_size
if FLAGS.train_update_lr == -1:
FLAGS.train_update_lr = FLAGS.update_lr
exp_string = 'cls_' + str(FLAGS.num_classes) + '.mbs_' + str(
FLAGS.meta_batch_size) + '.ubs_' + str(
FLAGS.train_update_batch_size) + '.numstep' + str(
FLAGS.num_updates) + '.updatelr' + str(FLAGS.train_update_lr)
if FLAGS.num_filters != 64:
exp_string += 'hidden' + str(FLAGS.num_filters)
if FLAGS.max_pool:
exp_string += 'maxpool'
if FLAGS.stop_grad:
exp_string += 'stopgrad'
if FLAGS.baseline:
exp_string += FLAGS.baseline
if FLAGS.norm == 'batch_norm':
exp_string += 'batchnorm'
elif FLAGS.norm == 'layer_norm':
exp_string += 'layernorm'
elif FLAGS.norm == 'None':
exp_string += 'nonorm'
else:
print('Norm setting not recognized.')
resume_itr = 0
model_file = None
tf.global_variables_initializer().run()
tf.train.start_queue_runners()
if FLAGS.resume or not FLAGS.train:
model_file = tf.train.latest_checkpoint(FLAGS.logdir + '/' +
exp_string)
if FLAGS.test_iter > 0:
model_file = model_file[:model_file.index('model'
)] + 'model' + str(
FLAGS.test_iter)
if model_file:
ind1 = model_file.index('model')
resume_itr = int(model_file[ind1 + 5:])
print("Restoring model weights from " + model_file)
saver.restore(sess, model_file)
if FLAGS.train:
train(model, autoencoder, saver, sess, exp_string, data_generator,
resume_itr)
else:
test(model, autoencoder, saver, sess, exp_string, data_generator,
test_num_updates)
