def check_features(data_type, network='Resnet50'):
"""Check if features file exists locally
Only one dataset, e.g. train, valid, test is checked
"""
data_dir = bfeatures.set_features_file(data_type, network,
return_type='dir')
bottleneck_file = bfeatures.set_features_file(data_type,
network, return_type='file')
bottleneck_path = bfeatures.set_features_file(data_type, network)
bottleneck_exists, _ = dutils.maybe_download_data(data_dir=data_dir,
data_file=bottleneck_file)
if not bottleneck_exists:
print("[INFO] %s was neither found nor downloaded. Trying to build. It may take time .. "
% bottleneck_file)
# check if directory with train, test, and valid images exists:
dutils.maybe_download_and_unzip()
bfeatures.build_features(data_type, network)
# Upload to nextcloud newly created file
bottleneck_exists = True if os.path.exists(bottleneck_path) else False
dest_dir = cfg.Dog_RemoteStorage.rstrip('/') + '/'+ data_dir
print("[INFO] Upload %s to %s" % (bottleneck_path, dest_dir))
dutils.rclone_copy(bottleneck_path, dest_dir)
return bottleneck_exists
def prepare_data(network='Resnet50'):
""" Function to prepare data
"""
# check if dog_names file exists locally, if not -> download,
# if not downloaded -> dutils.dog_names_create()
dog_names_file = cfg.Dog_LabelsFile.split('/')[-1]
status_dog_names, _ = dutils.maybe_download_data(data_dir='/data',
data_file=dog_names_file)
if not status_dog_names:
dutils.maybe_download_and_unzip()
dutils.dog_names_create()
else:
print("[INFO] %s exists" % (cfg.Dog_LabelsFile))
# check if bottleneck_features file exists locally
# if not -> download it, if not downloaded -> try to build
# train
status = { True: "exists", False: "does not exist"}
datasets = ['train', 'valid', 'test']
for dset in datasets:
status_targets = check_targets(dset)
print("[INFO] Targets file for %s %s" % (dset, status[status_targets]))
status_bottleneck = check_features(dset, network)
print("[INFO] Bottleneck file for %s (%s) %s" %
(dset, network, status[status_bottleneck]))
def check_targets(data_type):
"""Check if targets file exists locally
Only one dataset, e.g. train, valid, test is checked
"""
data_dir = '/data'
targets_file = 'Dogs_targets_' + data_type + '.npz'
targets_path = os.path.join(cfg.BASE_DIR, 'data', targets_file)
targets_exists, _ = dutils.maybe_download_data(data_dir=data_dir,
data_file=targets_file)
if not targets_exists:
print("[INFO] %s was neither found nor downloaded. Trying to build .. "
% targets_file)
# check if directory with train, test, and valid images exists:
dutils.maybe_download_and_unzip()
dutils.build_targets(data_type)
# Upload to nextcloud newly created file
targets_exists = True if os.path.exists(targets_path) else False
dest_dir = cfg.Dog_RemoteStorage.rstrip('/') + data_dir
print("[INFO] Upload %s to %s" % (targets_path, dest_dir))
dutils.rclone_copy(targets_path, dest_dir)
return targets_exists
def predict_file(img_path, network='Resnet50'):
"""
Function to make prediction which breed is closest
:param img_path: image to classify, full path
:param network: neural network to be used
:return: most probable dogs breed
"""
nets = {
'VGG16': bfeatures.extract_VGG16,
'VGG19': bfeatures.extract_VGG19,
'Resnet50': bfeatures.extract_Resnet50,
'InceptionV3': bfeatures.extract_InceptionV3,
}
# clear possible pre-existing sessions. IMPORTANT!
K.clear_session()
# check that all necessary data is there
#prepare_data(network)
weights_file = mutils.build_weights_filename(network)
saved_weights_path = os.path.join(cfg.BASE_DIR, 'models', weights_file)
# check if the weights file exists locally. if not -> try to download
status_weights, _ = dutils.maybe_download_data(data_dir='/models',
data_file=weights_file)
# attempt to download default weights file
if not status_weights and network in cfg.cnn_list:
print("[INFO] Trying to download weights from the public link")
url_weights = cfg.Dog_RemoteShare + weights_file
status_weights, _ = dutils.url_download(url_path=url_weights,
data_dir=os.path.join(
cfg.BASE_DIR, 'models'),
data_file=weights_file)
dog_names_file = cfg.Dog_LabelsFile.split('/')[-1]
# check if the labels file exists locally. if not -> try to download
status_dog_names, _ = dutils.maybe_download_data(data_dir='/data',
data_file=dog_names_file)
# attempt to download labels file
if not status_dog_names:
print("[INFO] Trying to download labels from the public link")
url_dog_names = cfg.Dog_RemoteShare + dog_names_file
status_weights, _ = dutils.url_download(url_path=url_dog_names,
data_dir=os.path.join(
cfg.BASE_DIR, 'data'),
data_file=dog_names_file)
if status_weights:
dog_names = dutils.dog_names_load()
net_model = build_model(network, len(dog_names))
net_model.load_weights(saved_weights_path)
# extract bottleneck features
bottleneck_feature = nets[network](dutils.path_to_tensor(img_path))
if debug_model:
print("[INFO] Bottleneck feature size: {}".format(
bottleneck_feature.shape))
# obtain predicted vector
predicted_vector = net_model.predict(bottleneck_feature)
print("[INFO] Sum: %f" % np.sum(predicted_vector))
# return dog breed that is predicted by the model
idxs = np.argsort(predicted_vector[0])[::-1][:5]
# dog_names_best = [ dog_names[i] for i in idxs ]
dog_names_best = []
probs_best = []
for i in idxs:
dog_names_best.append(dog_names[i])
probs_best.append(predicted_vector[0][i])
print("%s : %f" % (dog_names[i], predicted_vector[0][i]))
msg = mutils.format_prediction(dog_names_best, probs_best)
else:
msg = "ERROR in predict_file(). No weights file found! " + \
"Please first train the model with the " + network + " network!"
msg = {"Error": msg}
return msg