def load_labels(restart_valid_train):
# Generate consistent train and validation sets
## Only do this if the train / validation has not been generated yet
original_labels = dataset_loaders.groundlabels_dataframe()
if restart_valid_train == True:
os.system('rm -f %s/%s/train_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name))
os.system('rm -f %s/%s/valid_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name))
if not os.path.exists(
'%s/%s/train_df.csv' %
(settings.DATAMODEL_PATH,
experiment_folder_name)) and not os.path.exists(
'%s/%s/valid_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name)):
cases = original_labels['image'].unique()
np.random.shuffle(cases)
train_data = original_labels[original_labels.image.isin(
cases[num_valid_cases:])]
valid_data = original_labels[~original_labels.image.
isin(cases[num_valid_cases:])]
train_data.to_csv('%s/%s/train_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name),
index=False)
valid_data.to_csv('%s/%s/valid_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name),
index=False)
train_data = pd.read_csv('%s/%s/train_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name))
valid_data = pd.read_csv('%s/%s/valid_df.csv' %
(settings.DATAMODEL_PATH, experiment_folder_name))
return original_labels, train_data, valid_data
patches.append(
scipy.misc.imresize(img[xi - wnd:xi + wnd, yi - wnd:yi + wnd],
[image_size_nn, image_size_nn]) / 255)
if len(patches) == batch_size:
yield np.array(patches).transpose([0, 3, 1, 2])
patches = []
if len(patches) > 0:
yield np.array(patches).transpose([0, 3, 1, 2])
########################################################
### Initilize things
########################################################
# We load the cases as they were originally...if needed
original_labels = dataset_loaders.groundlabels_dataframe()
original_labels = dataset_loaders.map_labels(original_labels, map_category)
## train_data = pd.read_csv('%s/%s/train_df.csv' % (settings.DATAMODEL_PATH, experiment_folder_name))
## valid_data = pd.read_csv('%s/%s/valid_df.csv' % (settings.DATAMODEL_PATH, experiment_folder_name))
# Initialize these cases
## We add background as label to the rest of existing labels
existing_labels = np.concatenate(
[original_labels['class'].unique(), ['background']])
labelencoder = LabelEncoding(existing_labels)
########################################################
### Load model
########################################################
model = ResnetBuilder().build_resnet_50((3, image_size_nn, image_size_nn),
len(existing_labels))