def build_coco_dsets(args):
dset_kwargs = {
'image_dir': args.coco_train_image_dir,
'instances_json': args.coco_train_instances_json,
'stuff_json': args.coco_train_stuff_json,
'stuff_only': args.coco_stuff_only,
'image_size': args.image_size,
'mask_size': args.mask_size,
'max_samples': args.num_train_samples,
'min_object_size': args.min_object_size,
'min_objects_per_image': args.min_objects_per_image,
'instance_whitelist': args.instance_whitelist,
'stuff_whitelist': args.stuff_whitelist,
'include_other': args.coco_include_other,
'include_relationships': args.include_relationships,
}
train_dset = CocoSceneGraphDataset(**dset_kwargs)
num_objs = train_dset.total_objects()
num_imgs = len(train_dset)
print('Training dataset has %d images and %d objects' %
(num_imgs, num_objs))
print('(%.2f objects per image)' % (float(num_objs) / num_imgs))
dset_kwargs['image_dir'] = args.coco_val_image_dir
dset_kwargs['instances_json'] = args.coco_val_instances_json
dset_kwargs['stuff_json'] = args.coco_val_stuff_json
dset_kwargs['max_samples'] = args.num_val_samples
val_dset = CocoSceneGraphDataset(**dset_kwargs)
assert train_dset.vocab == val_dset.vocab
vocab = json.loads(json.dumps(train_dset.vocab))
return vocab, train_dset, val_dset
def build_coco_dsets(args):
dset_kwargs = {
'image_dir': args.coco_train_image_dir,
'instances_json': args.coco_train_instances_json,
'stuff_json': args.coco_train_stuff_json,
'stuff_only': args.coco_stuff_only,
'image_size': args.image_size,
'mask_size': args.mask_size,
'max_samples': args.num_train_samples,
'min_object_size': args.min_object_size,
'min_objects_per_image': args.min_objects_per_image,
'instance_whitelist': args.instance_whitelist,
'stuff_whitelist': args.stuff_whitelist,
'include_other': args.coco_include_other,
'include_relationships': args.include_relationships,
'seed': 0 # randomize for train
#'heuristics_ordering' : args.heuristics_ordering
}
train_dset = None
if args.coco_view_validation_error is False:
train_dset = CocoSceneGraphDataset(**dset_kwargs)
num_objs = train_dset.total_objects()
num_imgs = len(train_dset)
print('Training dataset has %d images and %d objects' %
(num_imgs, num_objs))
print('(%.2f objects per image)' % (float(num_objs) / num_imgs))
dset_kwargs['image_dir'] = args.coco_val_image_dir
dset_kwargs['instances_json'] = args.coco_val_instances_json
dset_kwargs['stuff_json'] = args.coco_val_stuff_json
dset_kwargs['max_samples'] = args.num_val_samples
# *deactivate* randomization for val (for consistent results)
dset_kwargs['seed'] = args.random_seed
val_dset = CocoSceneGraphDataset(**dset_kwargs)
if args.coco_view_validation_error is True:
print('Using val dataset at train dataset to view validation error')
train_dset = val_dset
num_objs = train_dset.total_objects()
num_imgs = len(train_dset)
print('Valdiation dataset has %d images and %d objects' %
(num_imgs, num_objs))
print('(%.2f objects per image)' % (float(num_objs) / num_imgs))
assert train_dset.vocab == val_dset.vocab
vocab = json.loads(json.dumps(train_dset.vocab))
return vocab, train_dset, val_dset
def build_coco_dset(args, checkpoint):
checkpoint_args = checkpoint['args']
dset_kwargs = {
'image_dir': args.coco_train_image_dir,
'instances_json': args.coco_train_instances_json,
'stuff_json': args.coco_train_stuff_json,
'stuff_only': checkpoint_args['coco_stuff_only'],
'oldlist_txt': args.coco_train_oldlist_txt,
'image_size': args.image_size,
'mask_size': checkpoint_args['mask_size'],
'max_samples': args.num_samples,
'min_object_size': checkpoint_args['min_object_size'],
'min_objects_per_image': checkpoint_args['min_objects_per_image'],
'instance_whitelist': checkpoint_args['instance_whitelist'],
'stuff_whitelist': checkpoint_args['stuff_whitelist'],
'include_other': checkpoint_args.get('coco_include_other', True),
}
# training set
if (args.which_data == 'train'):
dset = CocoSceneGraphDataset(**dset_kwargs)
else: # validation set
dset_kwargs['image_dir'] = args.coco_val_image_dir
dset_kwargs['instances_json'] = args.coco_val_instances_json
dset_kwargs['stuff_json'] = args.coco_val_stuff_json
dset_kwargs['oldlist_txt'] = args.coco_val_oldlist_txt
dset = CocoSceneGraphDataset(**dset_kwargs)
num_objs = dset.total_objects()
num_imgs = len(dset)
print('Training dataset has %d images and %d objects' %
(num_imgs, num_objs))
print('(%.2f objects per image)' % (float(num_objs) / num_imgs))
print('Training dataset has %d images and %d objects' %
(num_imgs, num_objs))
print('(%.2f objects per image)' % (float(num_objs) / num_imgs))
return dset
def build_coco_dset(args, checkpoint):
checkpoint_args = checkpoint['args']
print('include other: ', checkpoint_args.get('coco_include_other'))
dset_kwargs = {
'image_dir': args.coco_image_dir,
'instances_json': args.instances_json,
'stuff_json': args.stuff_json,
'stuff_only': checkpoint_args['coco_stuff_only'],
'image_size': args.image_size,
'mask_size': checkpoint_args['mask_size'],
'max_samples': args.num_samples,
'min_object_size': checkpoint_args['min_object_size'],
'min_objects_per_image': checkpoint_args['min_objects_per_image'],
'instance_whitelist': checkpoint_args['instance_whitelist'],
'stuff_whitelist': checkpoint_args['stuff_whitelist'],
'include_other': checkpoint_args.get('coco_include_other', True),
}
dset = CocoSceneGraphDataset(**dset_kwargs)
return dset
def build_coco_dsets(args):
dset_kwargs = {
'image_dir': args.coco_train_image_dir,
'instances_json': args.coco_train_instances_json,
'stuff_json': args.coco_train_stuff_json,
'stuff_only': args.coco_stuff_only,
'image_size': args.image_size,
'mask_size': args.mask_size,
'max_samples': args.num_train_samples,
'min_object_size': args.min_object_size,
'min_objects_per_image': args.min_objects_per_image,
'instance_whitelist': args.instance_whitelist,
'stuff_whitelist': args.stuff_whitelist,
'include_other': args.coco_include_other,
'include_relationships': args.include_relationships,
'seed': args.random_seed
}
train_dset = None
#train_dset = CocoSceneGraphDataset(**dset_kwargs)
#num_objs = train_dset.total_objects()
#num_imgs = len(train_dset)
#print('Training dataset has %d images and %d objects' % (num_imgs, num_objs))
#print('(%.2f objects per image)' % (float(num_objs) / num_imgs))
dset_kwargs['image_dir'] = args.coco_val_image_dir
dset_kwargs['instances_json'] = args.coco_val_instances_json
dset_kwargs['stuff_json'] = args.coco_val_stuff_json
dset_kwargs['max_samples'] = args.num_val_samples
# *deactivate* randomization for val (for consistent results)
dset_kwargs['seed'] = args.random_seed
val_dset = CocoSceneGraphDataset(**dset_kwargs)
#assert train_dset.vocab == val_dset.vocab
#vocab = json.loads(json.dumps(train_dset.vocab))
vocab = json.loads(json.dumps(val_dset.vocab))
return vocab, train_dset, val_dset