def main():
"""
Load data and train a model on it.
"""
args = argument_parser().parse_args()
random.seed(args.seed)
train_set, test_set = split_dataset(read_dataset(DATA_DIR))
train_set = list(augment_dataset(train_set))
test_set = list(test_set)
model = OmniglotModel(args.classes, **model_kwargs(args))
with tf.Session() as sess:
if not args.pretrained:
print('Training...')
train(sess, model, train_set, test_set, args.checkpoint, **train_kwargs(args))
else:
print('Restoring from checkpoint...')
tf.train.Saver().restore(sess, tf.train.latest_checkpoint(args.checkpoint))
print('Evaluating...')
eval_kwargs = evaluate_kwargs(args)
print('Train accuracy: ' + str(evaluate(sess, model, train_set, **eval_kwargs)))
print('Test accuracy: ' + str(evaluate(sess, model, test_set, **eval_kwargs)))
def main():
"""
Load data and train a model on it.
"""
args = argument_parser().parse_args()
random.seed(args.seed)
train_set, test_set = split_dataset(read_dataset(DATA_DIR))
train_set = list(augment_dataset(train_set))
test_set = list(test_set)
model = OmniglotModel(args.classes, **model_kwargs(args))
with tf.Session() as sess:
eval_kwargs = evaluate_kwargs(args)
if not args.pretrained:
print('Training...')
train(sess,
model,
train_set,
test_set,
args.checkpoint,
eval_kwargs=eval_kwargs,
**train_kwargs(args))
else:
print('Restoring from checkpoint...')
tf.train.Saver().restore(
sess, tf.train.latest_checkpoint(args.checkpoint))
print('Evaluating...')
eval_kwargs = evaluate_kwargs(args)
print('Train accuracy: ' +
str(evaluate(sess, model, train_set, **eval_kwargs)))
print('Test accuracy: ' +
str(evaluate(sess, model, test_set, **eval_kwargs)))
def main():
"""
Load data and train a model on it.
"""
args = argument_parser().parse_args()
random.seed(args.seed)
train_set, test_set = split_dataset(read_dataset(DATA_DIR))
train_set = list(augment_dataset(train_set))
test_set = list(test_set)
model = OmniglotModel(args.classes, **model_kwargs(args))
with tf.Session() as sess:
resume_itr = 0 # Zero iterations have already been trained.
if args.pretrained or args.test: # It must be pretrained to test it.
print('Restoring from checkpoint...')
saved_model = tf.train.latest_checkpoint(args.checkpoint)
tf.train.Saver().restore(sess, saved_model)
resume_itr = int(saved_model[saved_model.index('model.ckpt') + 11:])
if not args.test:
print('Training...')
train(sess, model, train_set, test_set, args.checkpoint, resume_itr, **train_kwargs(args))
print('Evaluating...')
eval_kwargs = evaluate_kwargs(args)
print('Train accuracy: ' + str(evaluate(sess, model, train_set, **eval_kwargs)))
print('Test accuracy: ' + str(evaluate(sess, model, test_set, **eval_kwargs)))
def main():
"""
Load data and train a model on it.
"""
context = neptune.Context()
context.integrate_with_tensorflow()
final_train_channel = context.create_channel('final_train_accuracy',
neptune.ChannelType.NUMERIC)
final_test_channel = context.create_channel('final_test_accuracy',
neptune.ChannelType.NUMERIC)
args = neptune_args(context)
print('args:\n', args)
random.seed(args.seed)
train_set, test_set = split_dataset(read_dataset(args.omniglot_src))
train_set = list(augment_dataset(train_set))
test_set = list(test_set)
model = ProgressiveOmniglotModel(args.classes, **model_kwargs(args))
config = tf.ConfigProto()
config.gpu_options.allow_growth = args.allow_growth
with tf.Session(config=config) as sess:
if not args.pretrained:
print('Training...')
train(sess, model, train_set, test_set, args.checkpoint,
**train_kwargs(args))
else:
print('Restoring from checkpoint...')
tf.train.Saver().restore(
sess, tf.train.latest_checkpoint(args.checkpoint))
print('Evaluating...')
eval_kwargs = evaluate_kwargs(args)
final_train_accuracy = evaluate(sess, model, train_set, **eval_kwargs)
print('final_train_accuracy:', final_train_accuracy)
final_train_channel.send(final_train_accuracy)
final_test_accuracy = evaluate(sess, model, test_set, **eval_kwargs)
print('final_test_accuracy:', final_test_accuracy)
final_test_channel.send(final_test_accuracy)
def main():
parser = argument_parser()
parser.add_argument('--dataset',
help='which dataset to use',
default='omniglot',
type=str)
parser.add_argument('--trials',
help='number of seeds',
default=3,
type=int)
parser.add_argument('--val-samples',
help='number of validation samples',
default=100,
type=int)
parser.add_argument('--restore',
help='restore from final training checkpoint',
action='store_true')
args = parser.parse_args()
os.makedirs(args.checkpoint, exist_ok=True)
with open(os.path.join(args.checkpoint, 'args.json'), 'w') as f:
json.dump(vars(args), f, indent=4)
random.seed(args.seed)
if args.dataset == 'omniglot':
train_set, test_set = split_dataset(
read_dataset(DATA_DIR[args.dataset]))
trainval_set = list(augment_dataset(train_set))
val_set = list(train_set[-200:])
train_set = list(augment_dataset(train_set[:-200]))
test_set = list(test_set)
else:
train_set, val_set, test_set = mini_dataset(DATA_DIR[args.dataset])
trainval_set = train_set
print('Training...')
modes = ['train', 'test']
metrics = ['acc', 'loss']
try:
with open(os.path.join(args.checkpoint, 'results.pkl'), 'rb') as f:
results = pickle.load(f)
except FileNotFoundError:
results = {
'regular':
{mode: {metric: {}
for metric in metrics}
for mode in modes},
'trials': set()
}
if args.transductive:
results['transductive'] = {
mode: {metric: {}
for metric in metrics}
for mode in modes
}
if args.adaptive:
results['adaptive'] = {
mode: {metric: {}
for metric in metrics}
for mode in modes
}
if args.transductive:
results['adaptive_transductive'] = {
mode: {metric: {}
for metric in metrics}
for mode in modes
}
for j in range(args.trials):
if j in results['trials']:
print('skipping trial', j)
continue
tf.reset_default_graph()
model = OmniglotModel(args.classes, **model_kwargs(
args)) if args.dataset == 'omniglot' else MiniImageNetModel(
args.classes, **model_kwargs(args))
with tf.Session() as sess:
checkpoint = os.path.join(args.checkpoint, 'final' + str(j))
if args.restore:
print('Trial', j, 'Restoring...')
tf.train.Saver().restore(
sess, tf.train.latest_checkpoint(checkpoint))
args.restore = False
else:
os.makedirs(checkpoint, exist_ok=True)
print('Trial', j, 'Training...')
train(sess, model, trainval_set, test_set, checkpoint,
**train_kwargs(args))
print('Trial', j, 'Evaluating...')
for ev in results.keys():
if ev == 'trials':
continue
evkw = evaluate_kwargs(args)
evkw['transductive'] = 'transductive' in ev
evkw['adaptive'] = args.adaptive if 'adaptive' in ev else 0.0
for mode, dset in zip(modes, [trainval_set, test_set]):
results[ev][mode]['acc'][j], results[ev][mode]['loss'][
j] = evaluate(sess, model, dset, **evkw)
results['trials'].add(j)
with open(os.path.join(args.checkpoint, 'results.pkl'), 'wb') as f:
pickle.dump(results, f)
with open(os.path.join(args.checkpoint, 'results.json'), 'w') as f:
tr = results.pop('trials')
json.dump(results, f, indent=4)
results['trials'] = tr
print('Train Acc:',
sum(results['regular']['train']['acc'].values()) / args.trials)
print('Test Acc:',
sum(results['regular']['test']['acc'].values()) / args.trials)
if args.transductive:
print(
'Transductive Train Acc:',
sum(results['transductive']['train']['acc'].values()) /
args.trials)
print(
'Transductive Test Acc:',
sum(results['transductive']['test']['acc'].values()) / args.trials)
if args.adaptive:
print('Adaptive Train Acc:',
sum(results['adaptive']['train']['acc'].values()) / args.trials)
print('Adaptive Test Acc:',
sum(results['adaptive']['test']['acc'].values()) / args.trials)
if args.transductive:
print(
'Adaptive Transductive Train Acc:',
sum(results['adaptive_transductive']['train']['acc'].values())
/ args.trials)
print(
'Adaptive Transductive Test Acc:',
sum(results['adaptive_transductive']['test']['acc'].values()) /
args.trials)