def test_model(model, save):
nframes = 128
def image_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose([
tv.transforms.ToTensor(),
tv.transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
])
return transform(pic)
def label_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose([LabelToLongTensor()])
label = transform(pic)
return label
datasets = {
'DAVIS16_val':
DAVIS17V2(config['davis_path'], '2016', 'val', image_read, label_read,
nframes),
'DAVIS17_val':
DAVIS17V2(config['davis_path'], '2017', 'val', image_read, label_read,
nframes)
}
for key, dataset in datasets.items():
if key == 'DAVIS16_val':
evaluator = evaluation.VOSEvaluator(dataset, 'cuda', save)
print("\n-- DAVIS16 dataset initialization started.")
elif key == 'DAVIS17_val':
evaluator = evaluation.VOSEvaluator(dataset, 'cuda', save)
print("\n-- DAVIS17 dataset initialization started.")
result_fpath = os.path.join(config['output_path'])
evaluator.evaluate(model, os.path.join(result_fpath, key))
def forward(model,
path,
prediction_path,
debug_sequences_dict=None,
save_prediction=False):
size = (480, 854)
nframes = 128
# pre-processing the input
# function for reading a single image
def image_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose([
tv.transforms.ToTensor(),
tv.transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
])
return transform(pic)
def label_read(path):
if (os.path.exists(path)):
pic = Image.open(path)
transform = tv.transforms.Compose([LabelToLongTensor()])
label = transform(pic)
else:
label = torch.LongTensor(1, *size).fill_(
255) # Put label that will be ignored
return label
# our init : (self, root_path, image_set, image_read=get_default_image_read(),
# anno_read=get_default_anno_read(),
# joint_transform=None, samplelen=4, obj_selection=get_sample_all(), min_num_obj=1,
# start_frame='random')
# YTVOS init : (self, root_path, split, image_set, impath='JPEGImages', image_read=get_default_image_read(),
# anno_read=get_default_anno_read(), joint_transform=None,
# samplelen=4, obj_selection=get_sample_all(), min_num_obj=1, start_frame='random')
datasets = {
"new-botinok":
RANDATA(root_path=path,
image_set='val',
image_read=image_read,
anno_read=label_read,
joint_transform=None,
samplelen=nframes,
start_frame="first")
}
if debug_sequences_dict is None:
debug_sequences_dict = {key: () for key in datasets.keys()}
for key, dataset in datasets.items():
evaluator = evaluation.VOSEvaluator(dataset, 'cuda', 'all',
True, False,
debug_sequences_dict.get(key))
result_fpath = os.path.join(
path,
os.path.splitext(os.path.basename(__file__))[0])
evaluator.evaluate(model, os.path.join(result_fpath, key))
def test_model(model, debug_sequences_dict=None, save_predictions=False, prediction_path=None):
nframes = 128
def image_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose(
[tv.transforms.ToTensor(),
tv.transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)])
return transform(pic)
def label_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose(
[LabelToLongTensor()])
label = transform(pic)
return label
datasets = {
'DAVIS16_train': DAVIS17V2(config['davis17_path'], '2016', 'train', image_read, label_read, None, nframes),
'DAVIS16_val': DAVIS17V2(config['davis17_path'], '2016', 'val', image_read, label_read, None, nframes),
'DAVIS17_val': DAVIS17V2(config['davis17_path'], '2017', 'val', image_read, label_read, None, nframes),
'YTVOS_jval': YTVOSV2(config['ytvos_path'], 'train', 'val_joakim', 'JPEGImages', image_read, label_read,
None, nframes),
'YTVOS_val': YTVOSV2(config['ytvos_path'], 'valid', None, 'JPEGImages', image_read, label_read,
None, nframes)
}
multitarget_sets = ('DAVIS17_val', 'YTVOS_jval', 'YTVOS_val')
if debug_sequences_dict is None:
debug_sequences_dict = {key: () for key in datasets.keys()}
for key, dataset in datasets.items():
if key == 'YTVOS_val':
evaluator = evaluation.VOSEvaluator(dataset, 'cuda', 'all', True, False, debug_sequences_dict.get(key))
elif key == 'DAVIS17_val':
evaluator = evaluation.VOSEvaluator(dataset, 'cuda', 'all', True, True, debug_sequences_dict.get(key))
else:
evaluator = evaluation.VOSEvaluator(dataset, 'cuda', 'all', False, True, debug_sequences_dict.get(key))
result_fpath = os.path.join(config['output_path'], os.path.splitext(os.path.basename(__file__))[0])
if key in multitarget_sets: # Only apply the multitarget aggregation if we have mult. targets
model.update_with_softmax_aggregation = True
else:
model.update_with_softmax_aggregation = False
evaluator.evaluate(model, os.path.join(result_fpath, key))
def test_model(model,
debug_sequences_dict=None,
save_predictions=False,
prediction_path=None):
nframes = 128
def image_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose([
tv.transforms.ToTensor(),
tv.transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
])
return transform(pic)
def label_read(path):
pic = Image.open(path)
transform = tv.transforms.Compose([LabelToLongTensor()])
label = transform(pic)
return label
datasets = {
"DAVIS16_train":
DAVIS17V2(
config["davis17_path"],
"2016",
"train",
image_read,
label_read,
None,
nframes,
),
"DAVIS16_val":
DAVIS17V2(config["davis17_path"], "2016", "val", image_read,
label_read, None, nframes),
"DAVIS17_val":
DAVIS17V2(config["davis17_path"], "2017", "val", image_read,
label_read, None, nframes),
"YTVOS_jval":
YTVOSV2(
config["ytvos_path"],
"train",
"val_joakim",
"JPEGImages",
image_read,
label_read,
None,
nframes,
),
"YTVOS_val":
YTVOSV2(
config["ytvos_path"],
"valid",
None,
"JPEGImages",
image_read,
label_read,
None,
nframes,
),
}
multitarget_sets = ("DAVIS17_val", "YTVOS_jval", "YTVOS_val")
if debug_sequences_dict is None:
debug_sequences_dict = {key: () for key in datasets.keys()}
for key, dataset in datasets.items():
if key == "YTVOS_val":
evaluator = evaluation.VOSEvaluator(dataset, "cuda", "all", True,
False,
debug_sequences_dict.get(key))
elif key == "DAVIS17_val":
evaluator = evaluation.VOSEvaluator(dataset, "cuda", "all", True,
True,
debug_sequences_dict.get(key))
else:
evaluator = evaluation.VOSEvaluator(dataset, "cuda", "all", False,
True,
debug_sequences_dict.get(key))
result_fpath = os.path.join(
config["output_path"],
os.path.splitext(os.path.basename(__file__))[0])
if (
key in multitarget_sets
): # Only apply the multitarget aggregation if we have mult. targets
model.update_with_softmax_aggregation = True
else:
model.update_with_softmax_aggregation = False
evaluator.evaluate(model, os.path.join(result_fpath, key))
