def main():
parser = argparse.ArgumentParser(
description=
'Creates the dataset to train the matching classifier from triplet annotations'
)
parser.add_argument('--downsample_rate',
help='number of times that the input is downsampled',
default=16,
type=int)
parser.add_argument('--image_size',
help='input image size',
default=512,
type=int)
parser.add_argument('--no_random',
help='true too apply data augmentation',
action='store_false')
parser.add_argument('--add_appearance',
help='true add appearance features',
action='store_true')
parser.add_argument('--path_dataset',
help='path of the Imagenet VID dataset',
type=str)
args = parser.parse_args()
downsample_rate = args.downsample_rate
image_size = (args.image_size, args.image_size)
rndm = args.no_random
add_appearace_similarity = args.add_appearance
path_dataset = args.path_dataset
print(downsample_rate, image_size, rndm, add_appearace_similarity,
path_dataset)
if add_appearace_similarity:
import roi_nn
import train_utils
base_path_model = './demos/YOLOv3/pretrained_models/ILSVRC/1203_1758_model_8/'
path_roi_model = base_path_model + 'embedding_model/'
path_weights = train_utils.get_best_weights(base_path_model)
backbone = roi_nn.get_backbone(path_weights,
downsample_rate=downsample_rate)
branch_model = roi_nn.load_branch_body(path_roi_model)
else:
backbone, branch_model = None, None
np.random.seed(0)
path_annotations = './data_annotations/triplet_annotations/triplet_annotations_train.txt'
get_and_store_dataset_aug(path_annotations, path_dataset, image_size,
backbone, branch_model, downsample_rate, rndm,
'train')
path_annotations = './data_annotations/triplet_annotations/triplet_annotations_val.txt'
get_and_store_dataset_aug(path_annotations, path_dataset, image_size,
backbone, branch_model, downsample_rate, rndm,
'val')
def get_best_weights(model_folder, train_params, score=0, iou=0.5):
annotations_file = train_params['path_annotations'][1]
eval_t = '{}stats_{}_score{}_iou{}.json'.format(model_folder, annotations_file.split('/')[-1][:-4], score, iou)
eval_f = '{}stats_stage2_{}_score{}_iou{}.json'.format(model_folder, annotations_file.split('/')[-1][:-4], score, iou)
eval_t = json.load(open(eval_t, 'r'))
eval_f = json.load(open(eval_f, 'r'))
if eval_t['total'][1] > eval_f['total'][1]:
# best_weights
return train_utils.get_best_weights(model_folder)
else:
# Final weights
return model_folder + 'weights/trained_weights_final.h5'
def predict_and_store_from_annotations(model_folder,
train_params,
annotations_file,
preds_filename,
score,
iou,
best_weights=True):
if os.path.isfile(preds_filename): return None, -1
if best_weights:
model_path = train_utils.get_best_weights(model_folder)
else:
model_path = model_folder + 'weights/trained_weights_final.h5'
model = EYOLO(
model_image_size=tuple(train_params['input_shape']),
model_path=model_path,
classes_path=train_params['path_classes'],
anchors_path=train_params['path_anchors'],
score=score, # 0.3
iou=iou, # 0.5
td_len=train_params.get('td_len', None),
mode=train_params.get('mode', None),
spp=train_params.get('spp', False))
# Create pred file
with open(annotations_file, 'r') as f:
annotations = f.read().splitlines()
preds = []
t = time.time()
total = len(annotations)
for ann in tqdm(annotations[:total], total=total, file=sys.stdout):
img = ann.split()[0]
# image = cv2.imread(train_params['path_dataset'] + img)
# image = Image.fromarray(image)
images = [
Image.open(train_params['path_dataset'] + img)
for img in img.split(',')
]
img_size = images[0].size
# images = images[0] if len(images) == 1 else np.stack(images)
boxes, scores, classes = model.get_prediction(images)
for i in range(len(boxes)):
left, bottom, right, top = [int(b) for b in boxes[i].tolist()]
left, bottom = max(0, left), max(0, bottom)
right, top = min(img_size[1], right), min(img_size[0], top)
width = top - bottom
height = right - left
preds.append({
'image_id': img[:-4],
'category_id': int(classes[i]),
# 'bbox': [ x_min, y_min, width, height ],
'bbox': [bottom, left, width, height],
'score': float(scores[i]),
})
fps = len(annotations) / (time.time() - t)
json.dump(preds, open(preds_filename, 'w'))
return model, fps
logging,
checkpoint,
reduce_lr_2,
# reduce_lr
early_stopping
])
model.save_weights(path_model + 'weights/trained_weights_final.h5')
print('=' * print_line)
## %%
train_utils.remove_worst_weights(path_model)
## %%
best_weights = train_utils.get_best_weights(path_model)
model.load_weights(best_weights)
# %%
log('EVALUATING')
model_num = int(path_model.split('/')[-2].split('_')[-1])
#dataset_name, path_results = 'kitchen', '/mnt/hdd/egocentric_results/'
#evaluate_model.main(path_results, dataset_name, model_num, score, iou, num_annotation_file, plot=True, full=True)
iou = 0.5
if dataset_name == 'kitchen':
annotation_files = [(0.005, 1), (0.005, 0)]
else:
annotation_files = [(0, 1), (evaluate_model.MIN_SCORE, 0)]
def predict_and_store_from_annotations(model_folder,
train_params,
annotations_file,
output_dir,
image_size,
score,
nms_iou,
best_weights=True,
raw_eval='def',
load=False,
model_path=None):
if raw_eval == 'raw_scores':
preds_filename = '{}preds_{}_{}_is{}_score{}_raw_score.json'.format(
output_dir, 'bw' if best_weights else 'stage2',
annotations_file.split('/')[-1][:-4], image_size[0], score,
nms_iou)
elif raw_eval == 'raw':
preds_filename = '{}preds_{}_{}_is{}_score{}_raw.json'.format(
output_dir, 'bw' if best_weights else 'stage2',
annotations_file.split('/')[-1][:-4], image_size[0], score)
elif 'def':
preds_filename = '{}preds_{}_{}_is{}_score{}_iou{}.json'.format(
output_dir, 'bw' if best_weights else 'stage2',
annotations_file.split('/')[-1][:-4], image_size[0], score,
nms_iou)
else:
raise ValueError('Unrecognized eval error')
if os.path.isfile(preds_filename):
print(' * Loading:', preds_filename)
if load: return json.load(open(preds_filename, 'r')), preds_filename
else: return None, preds_filename
print(' * Predicting:', preds_filename)
if model_path is None:
if best_weights:
model_path = train_utils.get_best_weights(model_folder)
else:
model_path = model_folder + 'weights/trained_weights_final.h5'
model = EYOLO(
model_image_size=image_size,
model_path=model_path,
classes_path=train_params['path_classes'],
anchors_path=train_params['path_anchors'],
score=score,
iou=nms_iou, # 0.5
td_len=train_params.get('td_len', None),
mode=train_params.get('mode', None),
spp=train_params.get('spp', False),
raw_eval=raw_eval)
# Create pred file
with open(annotations_file, 'r') as f:
annotations = f.read().splitlines()
preds = []
prediction_time = 0
total = len(annotations)
total_prediction_time = time.time()
for ann in tqdm(annotations[:total], total=total, file=sys.stdout):
img = ann.split()[0]
# image = cv2.imread(train_params['path_dataset'] + img)
# image = Image.fromarray(image)
images = [
Image.open(train_params['path_dataset'] + img)
for img in img.split(',')
]
img_size = images[0].size
max_size = max(img_size[0], img_size[1])
# images = images[0] if len(images) == 1 else np.stack(images)
t = time.time()
boxes, scores, classes = model.get_prediction(images)
prediction_time += time.time() - t
for i in range(len(boxes)):
left, bottom, right, top = [
min(int(b), max_size) for b in boxes[i].tolist()
]
left, bottom = max(0, left), max(0, bottom)
right, top = min(img_size[1], right), min(img_size[0], top)
width = top - bottom
height = right - left
preds.append({
'image_id': '.'.join(img.split('.')[:-1]),
'category_id': int(classes[i]),
# 'bbox': [ x_min, y_min, width, height ],
'bbox': [bottom, left, width, height],
'score': float(scores[i]),
})
print('Total prediction time: ', time.time() - total_prediction_time)
print('Prediction time: {} secs || {:.2f}'.format(
prediction_time, prediction_time / len(annotations)))
# keras.backend.clear_session()
model.yolo_model = None
del model.yolo_model
model = None
del model
# fps = len(annotations)/(time.time()-t)
json.dump(preds, open(preds_filename, 'w'))
if load: return preds, preds_filename
else: return None, preds_filename
model_num = 1
model_folder = train_utils.get_model_path('/mnt/hdd/egocentric_results/',
'kitchen', model_num)
train_params = json.load(open(model_folder + 'train_params.json', 'r'))
# classes_path = './dataset_scripts/kitchen/kitchen_classes{}.txt'.format(version)
classes_path = train_params['path_classes']
path_base = ''
# anchors_path = 'base_models/yolo_anchors.txt'
anchors_path = train_params['path_anchors']
model_image_size = [416, 416]
else:
raise ValueError('Model not recognized')
if model_path is None:
model_path = train_utils.get_best_weights(model_folder)
print('Loading:', model_path)
print('=' * 80)
model = EYOLO(
model_image_size=(img_size, img_size),
model_path=model_path,
anchors_path=anchors_path,
classes_path=classes_path,
score=0.1,
iou=0.5,
# gpu_num = 2
td_len=train_params['td_len'],
mode=train_params['mode'],
spp=train_params.get('spp', False))