train_val_test_split=(1200 / 1622, 211 / 1622, 211 / 1622),
samples_per_iter=1,
num_workers=4,
data_path="datasets/omniglot_data",
name="omniglot_data",
index_of_folder_indicating_class=-2,
reset_stored_filepaths=False,
num_samples_per_class=args.samples_per_class,
num_classes_per_set=args.classes_per_set,
label_as_int=True)
experiment = ExperimentBuilder(data)
one_shot_omniglot, losses, c_error_opt_op, init = experiment.build_experiment(
args.batch_size,
args.classes_per_set,
args.samples_per_class,
args.use_full_context_embeddings,
full_context_unroll_k=args.full_context_unroll_k,
args=args)
total_train_batches = args.total_iter_per_epoch
total_val_batches = args.total_iter_per_epoch
total_test_batches = args.total_iter_per_epoch
saved_models_filepath, logs_filepath = build_experiment_folder(
args.experiment_title)
save_statistics(logs_filepath, [
"epoch", "total_train_c_loss_mean", "total_train_c_loss_std",
"total_train_accuracy_mean", "total_train_accuracy_std",
"total_val_c_loss_mean", "total_val_c_loss_std", "total_val_accuracy_mean",
"total_val_accuracy_std", "total_test_c_loss_mean",
batch_size = int(32 // sp) # default 32 for 5way1shot
classes_per_set = opt.way #20
samples_per_class = opt.shot
# N-way, K-shot
continue_from_epoch = opt.ckp # use -1 to start from scratch
logs_path = "one_shot_outputs/"
experiment_name = "LGM-Net_{}way{}shot".format(classes_per_set,
samples_per_class)
# Experiment builder
data = dataset.MiniImageNetDataSet(batch_size=batch_size,
classes_per_set=classes_per_set,
samples_per_class=samples_per_class,
shuffle_classes=True)
experiment = ExperimentBuilder(data)
one_shot_miniImagenet, losses, c_error_opt_op, init = experiment.build_experiment(
batch_size, classes_per_set, samples_per_class)
total_epochs = 120
total_train_batches = 1000
total_val_batches = int(250 * sp)
total_test_batches = int(250 * sp)
logs = "{}way{}shot learning problems, with {} tasks per task batch".format(
classes_per_set, samples_per_class, batch_size)
save_statistics(experiment_name, ["Experimental details: {}".format(logs)])
save_statistics(experiment_name, [
"epoch", "train_c_loss", "train_c_accuracy", "val_loss",
"val_accuracy", "test_c_loss", "test_c_accuracy", "learning_rate"
])
# Experiment initialization and running
config = tf.ConfigProto()
logs_path = "one_shot_outputs_ADNI/"
experiment_name = "one_shot_ADNI_{}fold_{}_{}".format(k, samples_per_class,
classes_per_set)
# Experiment builder
# data = dataset.OmniglotNShotDataset(batch_size=batch_size, classes_per_set=classes_per_set,
# samples_per_class=samples_per_class)
data = dataset.ADNIDataset(k=k,
batch_size=batch_size,
classes_per_set=classes_per_set,
samples_per_class=samples_per_class)
experiment = ExperimentBuilder(data)
one_shot_omniglot, losses, c_error_opt_op, init = experiment.build_experiment(
batch_size, classes_per_set, samples_per_class, fce)
total_epochs = 100
total_train_batches = 5
total_val_batches = 2
total_test_batches = 2
save_statistics(experiment_name, [
"epoch", "train_c_loss", "train_c_accuracy", "val_loss", "val_accuracy",
"test_c_loss", "test_c_accuracy"
])
# summary_path = "/summary/%d" % (int(time.time()))
# Experiment initialization and running
with tf.Session() as sess:
data = dataset.CIFAR_100(batch_size=batch_size,
samples_per_class=samples_per_class,
queries_per_class=queries_per_class)
elif data_name == "omniglot":
channels = 1
image_size = 28
augment = True
data = dataset.OmniglotNShotDataset(batch_size=batch_size,\
classes_per_set=classes_per_set, samples_per_class=samples_per_class, queries_per_class = queries_per_class)
else:
print("Unsupported dataset.")
assert False
experiment = ExperimentBuilder(data)
one_shot, losses, c_error_opt_op, init = experiment.build_experiment(batch_size,\
classes_train, classes_test, samples_per_class, queries_per_class, channels, image_size, fce, network_name)
#
total_epochs = 300
total_train_batches = 1000
total_val_batches = 100
total_test_batches = 250
save_statistics(experiment_name, [
"epoch", "train_c_loss", "train_c_accuracy", "val_loss", "val_accuracy",
"test_c_loss", "test_c_accuracy"
])
# Experiment initialization and running
with tf.Session() as sess:
sess.run(init)
