if use_gpu:
trained_cnn = trained_cnn.cuda()
print("=CNN state loaded=")
print("Extracting distractors features...")
# Dump the features to then load them
distractors_features_folder_name = save_features(trained_cnn,
target_shapes_dataset,
cnn_dump_id)
# Load data
if should_train_visual:
assert False
_train_data, _valid_data, _test_data = load_images(
'shapes/{}'.format(target_shapes_dataset), BATCH_SIZE, K)
else:
n_pretrained_image_features, _t, _v, test_data = load_pretrained_features_zero_shot(
target_features_folder_name, distractors_features_folder_name,
BATCH_SIZE, K)
assert n_pretrained_image_features == n_image_features
# Create onehot metadata if not created yet - only target is needed
if not does_shapes_onehot_metadata_exist(target_shapes_dataset):
create_shapes_onehot_metadata(target_shapes_dataset)
# Load metadata - only target is needed
_train_metadata, _valid_metadata, target_test_metadata = load_shapes_onehot_metadata(
target_shapes_dataset)
# Settings
# Load metadata
train_metadata, valid_metadata, test_metadata, noise_metadata = load_shapes_onehot_metadata(
shapes_dataset)
else:
train_metadata = None
valid_metadata = None
test_metadata = None
noise_metadata = None
print("loaded metadata")
print("loading data")
# Load data
if not shapes_dataset is None:
if not use_symbolic_input:
if should_train_visual:
train_data, valid_data, test_data, noise_data = load_images(
'shapes/{}'.format(shapes_dataset), BATCH_SIZE, K)
else:
n_pretrained_image_features, train_data, valid_data, test_data, noise_data = load_pretrained_features(
features_folder_name, BATCH_SIZE, K)
assert n_pretrained_image_features == n_image_features
else:
n_image_features, train_data, valid_data, test_data, noise_data = load_pretrained_features(
'shapes/{}'.format(shapes_dataset),
BATCH_SIZE,
K,
use_symbolic=True)
else:
n_image_features, train_data, valid_data, test_data, noise_data = load_pretrained_features(
'data/mscoco', BATCH_SIZE, K)
print('\nUsing {} image features\n'.format(n_image_features))
import neonetwork as nn
import dataloader as dl
def foo(net, images, labels, which):
net.set_input(dl.img_to_array(images[which]))
net.forward_prop()
# dl.show_img_arr(dl.img_to_array(images[which]))
print(net.get_out())
print(labels[which])
if __name__ == "__main__":
labels = dl.load_labels(
"http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz", 60000)
images = dl.load_images(
"http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz", 60000)
d = dl.create_data(images, labels)
train, test = dl.divide_train_test(d, 48000)
net = nn.Network([0 for i in range(784)])
net.train(train, 10)
net.test(test)