# Generate inception features
img_side = 299
num_channels = 3
num_train_ex_per_class = 900
num_test_ex_per_class = 300
batch_size = 100
tf.reset_default_graph()
full_model_name = '%s_inception' % dataset_name
full_model = BinaryInceptionModel(img_side=img_side,
num_channels=num_channels,
weight_decay=weight_decay,
num_classes=num_classes,
batch_size=batch_size,
data_sets=image_data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=True,
train_dir='output',
log_dir='log',
model_name=full_model_name)
train_inception_features_val = generate_inception_features(
full_model,
image_data_sets.train.x,
image_data_sets.train.labels,
batch_size=batch_size)
test_inception_features_val = generate_inception_features(
full_model,
image_data_sets.test.x,
# dataset_name = 'dogfish_orig_and_koda'
# data_sets = load_dogfish_with_orig_and_koda()
full_graph = tf.Graph()
top_graph = tf.Graph()
print('*** Full:')
with full_graph.as_default():
full_model_name = '%s_inception_wd-%s' % (dataset_name, weight_decay)
full_model = BinaryInceptionModel(
img_side=img_side,
num_channels=num_channels,
weight_decay=weight_decay,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=True,
train_dir='output',
log_dir='log',
model_name=full_model_name)
for data_set, label in [(data_sets.train, 'train'),
(data_sets.test, 'test')]:
inception_features_path = 'output/%s_inception_features_new_%s.npz' % (
dataset_name, label)
if not os.path.exists(inception_features_path):
print('Inception features do not exist. Generating %s...' % label)