def common_config(manifest_file,
manifest_root,
batch_size,
subset_pct,
h=96,
w=96,
scale=[1., 1.]):
cache_root = get_data_cache_or_nothing('stl10-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'subset_fraction': float(subset_pct / 100.0),
'macrobatch_size': 5000,
'type': 'image,label',
'cache_directory': cache_root,
'image': {
'height': h,
'width': w,
'scale': scale
},
'label': {
'binary': False
}
}
def common_config(manifest_file, batch_size):
cache_root = get_data_cache_or_nothing('mnist-cache/')
return {'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'minibatch_size': batch_size,
'macrobatch_size': 25000,
'cache_directory': cache_root,
'type': 'image,label',
'image': {'height': 28,
'width': 28,
'channels': 1},
'label': {'binary': False}}
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('ucf-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'macrobatch_size': batch_size * 12,
'type': 'video,label',
'cache_directory': cache_root,
'video': {'max_frame_count': 16,
'frame': {'height': 112,
'width': 112,
'scale': [0.875, 0.875]}},
'label': {'binary': False}
}
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('whale-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'macrobatch_size': batch_size * 12,
'type': 'audio,label',
'cache_directory': cache_root,
'audio': {'sample_freq_hz': 2000,
'max_duration': '2 seconds',
'frame_length': '80 milliseconds',
'frame_stride': '40 milliseconds'},
'label': {'binary': False}
}
def common_config(manifest_file, manifest_root, batch_size, subset_pct):
cache_root = get_data_cache_or_nothing('cifar-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'subset_fraction': float(subset_pct/100.0),
'macrobatch_size': 5000,
'type': 'image,label',
'cache_directory': cache_root,
'image': {'height': 32,
'width': 32,
'scale': [0.8, 0.8]},
'label': {'binary': False}
}
def PASCALVOC(manifest_file,
manifest_root,
rois_per_img=256,
height=1000,
width=1000,
inference=False):
"""
Returns the aeon dataloader configuration for PASCAL VOC dataset.
"""
CLASSES = ('__background__', 'aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable',
'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep',
'sofa', 'train', 'tvmonitor')
# if inference, we do not shuffle or flip
do_transforms = not inference
image_config = {"type": "image", "height": height, "width": width}
localization_config = {
"type": "localization_rcnn",
"height": height,
"width": width,
"rois_per_image": rois_per_img,
"class_names": CLASSES,
"scaling_factor": 1. / 16
}
augmentation = {
"type": "image",
"fixed_aspect_ratio": True,
"flip_enable": do_transforms,
"crop_enable": False
}
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'etl': [image_config, localization_config],
'cache_directory': get_data_cache_or_nothing(subdir='pascalvoc_cache'),
'shuffle_enable': do_transforms,
'shuffle_manifest': do_transforms,
'batch_size': 1,
'block_size': 100,
'augmentation': [augmentation]
}
def common_config(manifest_file, manifest_root, batch_size, subset_pct):
cache_root = get_data_cache_or_nothing('cifar-cache/')
image_config = {"type": "image", "height": 32, "width": 32}
label_config = {"type": "label", "binary": False}
augmentation = {"type": "image", "scale": [0.8, 0.8], "crop_enable": True}
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'batch_size': batch_size,
'block_size': 5000,
'subset_fraction': float(subset_pct / 100.0),
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]
}
def KITTI(manifest_file,
manifest_root,
rois_per_img=256,
height=375,
width=1242,
inference=False):
"""
Returns the aeon dataloader configuration for KITTI dataset.
"""
CLASSES = ('__background__', 'Car', 'Van', 'Truck', 'Pedestrian',
'Person_sitting', 'Cyclist', 'Tram', 'Misc', 'DontCare')
# if inference, we do not shuffle or flip
do_transforms = not inference
image_config = {"type": "image", "height": height, "width": width}
localization_config = {
"type": "localization_rcnn",
"rois_per_image": rois_per_img,
"height": height,
"width": width,
"class_names": CLASSES,
"scaling_factor": 1. / 16
}
augmentation = {
"type": "image",
"fixed_aspect_ratio": True,
"flip_enable": do_transforms,
"crop_enable": False
}
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'etl': [image_config, localization_config],
'cache_directory': get_data_cache_or_nothing(subdir='kitti_cache'),
'shuffle_enable': do_transforms,
'shuffle_manifest': do_transforms,
'batch_size': 1,
'block_size': 100,
'augmentation': [augmentation]
}
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('whale-cache/')
audio_config = {"type": "audio",
"sample_freq_hz": 2000,
"max_duration": '2 seconds',
"frame_length": '80 milliseconds',
"frame_stride": '40 milliseconds'}
label_config = {"type": "label",
"binary": False}
return {'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'batch_size': batch_size,
'block_size': batch_size * 12,
'cache_directory': cache_root,
'etl': [audio_config, label_config]}
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('whale-cache/')
audio_config = {"type": "audio",
"sample_freq_hz": 2000,
"max_duration": '2 seconds',
"frame_length": '80 milliseconds',
"frame_stride": '40 milliseconds'}
label_config = {"type": "label",
"binary": False}
return {'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'batch_size': batch_size,
'block_size': batch_size * 12,
'cache_directory': cache_root,
'etl': [audio_config, label_config]}
def common_config(manifest_file, manifest_root, batch_size, subset_pct):
cache_root = get_data_cache_or_nothing('lsun_cache/')
image_config = {"type": "image",
"height": 64,
"width": 64}
label_config = {"type": "label",
"binary": False}
augmentation = {"type": "image",
"scale": [1., 1.]}
return {'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'batch_size': batch_size,
'subset_fraction': float(subset_pct/100.0),
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]}
def common_config(manifest_file, manifest_root, batch_size, subset_pct):
cache_root = get_data_cache_or_nothing('cifar-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'subset_fraction': float(subset_pct / 100.0),
'macrobatch_size': 5000,
'type': 'image,label',
'cache_directory': cache_root,
'image': {
'height': 32,
'width': 32,
'scale': [0.8, 0.8]
},
'label': {
'binary': False
}
}
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('whale-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'macrobatch_size': batch_size * 12,
'type': 'audio,label',
'cache_directory': cache_root,
'audio': {
'sample_freq_hz': 2000,
'max_duration': '2 seconds',
'frame_length': '80 milliseconds',
'frame_stride': '40 milliseconds'
},
'label': {
'binary': False
}
}
def common_config(manifest_file, manifest_root, batch_size, subset_pct):
# export NEON_DATA_CACHE_DIR=</data/cache/>
cache_root = get_data_cache_or_nothing('i1k-cache/')
image_config = {"type": "image",
"height": 224,
"width": 224}
label_config = {"type": "label",
"binary": False}
augmentation = {"type": "image",
"scale": [0.875, 0.875],
"crop_enable": True}
return {'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'batch_size': batch_size,
'subset_fraction': float(subset_pct/100.0),
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]}
def PASCALVOC(manifest_file, manifest_root, rois_per_img=256,
height=1000, width=1000, inference=False):
"""
Returns the aeon dataloader configuration for PASCAL VOC dataset.
"""
CLASSES = ('__background__',
'aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair',
'cow', 'diningtable', 'dog', 'horse',
'motorbike', 'person', 'pottedplant',
'sheep', 'sofa', 'train', 'tvmonitor')
# if inference, we do not shuffle or flip
do_transforms = not inference
image_config = {"type": "image",
"height": height,
"width": width}
localization_config = {"type": "localization_rcnn",
"height": height,
"width": width,
"rois_per_image": rois_per_img,
"class_names": CLASSES,
"scaling_factor": 1. / 16}
augmentation = {"type": "image",
"fixed_aspect_ratio": True,
"flip_enable": do_transforms,
"crop_enable": False}
return {'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'etl': [image_config, localization_config],
'cache_directory': get_data_cache_or_nothing(subdir='pascalvoc_cache'),
'shuffle_enable': do_transforms,
'shuffle_manifest': do_transforms,
'batch_size': 1,
'block_size': 100,
'augmentation': [augmentation]}
def common_config(manifest_file, batch_size):
cache_root = get_data_cache_or_nothing('cifar10-cache/')
image_config = {"type": "image", "height": 32, "width": 32}
label_config = {"type": "label", "binary": False}
augmentation = {
"type": "image",
"scale": [0.8, 0.8],
"center": False,
"flip_enable": True
}
return {
'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]
}
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('ucf-cache/')
return {
'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'minibatch_size': batch_size,
'macrobatch_size': batch_size * 12,
'type': 'video,label',
'cache_directory': cache_root,
'video': {
'max_frame_count': 16,
'frame': {
'height': 112,
'width': 112,
'scale': [0.875, 0.875]
}
},
'label': {
'binary': False
}
}
def KITTI(manifest_file, manifest_root, rois_per_img=256,
height=375, width=1242, inference=False):
"""
Returns the aeon dataloader configuration for KITTI dataset.
"""
CLASSES = ('__background__',
'Car', 'Van', 'Truck',
'Pedestrian', 'Person_sitting', 'Cyclist', 'Tram',
'Misc', 'DontCare')
# if inference, we do not shuffle or flip
do_transforms = not inference
image_config = {"type": "image",
"height": height,
"width": width}
localization_config = {"type": "localization_rcnn",
"rois_per_image": rois_per_img,
"height": height,
"width": width,
"class_names": CLASSES,
"scaling_factor": 1. / 16}
augmentation = {"type": "image",
"fixed_aspect_ratio": True,
"flip_enable": do_transforms,
"crop_enable": False}
return {'manifest_filename': manifest_file,
'manifest_root': manifest_root,
'etl': [image_config, localization_config],
'cache_directory': get_data_cache_or_nothing(subdir='kitti_cache'),
'shuffle_enable': do_transforms,
'shuffle_manifest': do_transforms,
'batch_size': 1,
'block_size': 100,
'augmentation': [augmentation]}
def PASCALVOC(manifest_file,
manifest_root,
rois_per_img=256,
height=1000,
width=1000,
inference=False):
"""
Returns the aeon dataloader configuration for PASCAL VOC dataset.
"""
CLASSES = ('__background__', 'aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable',
'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep',
'sofa', 'train', 'tvmonitor')
# if inference, we do not shuffle or flip
do_transforms = not inference
image_decode_cfg = dict(height=height,
width=width,
flip_enable=do_transforms,
crop_enable=False,
fixed_aspect_ratio=True)
localization_cfg = dict(rois_per_image=rois_per_img,
class_names=CLASSES,
scaling_factor=1. / 16)
return dict(
manifest_filename=manifest_file,
manifest_root=manifest_root,
type='image,localization',
image=image_decode_cfg,
cache_directory=get_data_cache_or_nothing(subdir='pascalvoc_cache'),
shuffle_every_epoch=do_transforms,
shuffle_manifest=do_transforms,
minibatch_size=1,
macrobatch_size=100,
localization=localization_cfg,
)
def common_config(manifest_file, manifest_root, batch_size):
cache_root = get_data_cache_or_nothing('ucf-cache/')
video_config = {
"type": "video",
"max_frame_count": 16,
"frame": {
"height": 112,
"width": 112
}
}
label_config = {"type": "label", "binary": True}
return {
"manifest_filename": manifest_file,
"manifest_root": manifest_root,
"batch_size": batch_size,
"block_size": 5000,
"cache_directory": cache_root,
"etl": [video_config, label_config],
"augmentation": []
}
def KITTI(manifest_file,
manifest_root,
rois_per_img=256,
height=375,
width=1242,
inference=False):
"""
Returns the aeon dataloader configuration for KITTI dataset.
"""
CLASSES = ('__background__', 'Car', 'Van', 'Truck', 'Pedestrian',
'Person_sitting', 'Cyclist', 'Tram', 'Misc', 'DontCare')
# if inference, we do not shuffle or flip
do_transforms = not inference
image_decode_cfg = dict(height=height,
width=width,
flip_enable=do_transforms,
crop_enable=False,
fixed_aspect_ratio=True)
localization_cfg = dict(rois_per_image=rois_per_img,
class_names=CLASSES,
scaling_factor=1. / 16)
return dict(
manifest_filename=manifest_file,
manifest_root=manifest_root,
type='image,localization',
image=image_decode_cfg,
cache_directory=get_data_cache_or_nothing(subdir='kitti_cache'),
shuffle_every_epoch=do_transforms,
shuffle_manifest=do_transforms,
minibatch_size=1,
macrobatch_size=100,
localization=localization_cfg,
)
def get_ssd_config(img_reshape, inference=False):
ssd_config = OrderedDict()
ssd_config['batch_size'] = 32
ssd_config['shuffle_enable'] = True
ssd_config['shuffle_manifest'] = True
if inference:
ssd_config['batch_size'] = 1
ssd_config['block_size'] = 50
ssd_config['cache_directory'] = get_data_cache_or_nothing(
subdir='pascalvoc_cache')
ssd_config["etl"] = [{
"type":
"localization_ssd",
"height":
img_reshape[0],
"width":
img_reshape[1],
"max_gt_boxes":
500,
"class_names": [
"__background__", "aeroplane", "bicycle", "bird", "boat", "bottle",
"bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse",
"motorbike", "person", "pottedplant", "sheep", "sofa", "train",
"tvmonitor"
]
}, {
"type": "image",
"height": img_reshape[0],
"width": img_reshape[1],
"channels": 3
}]
if not inference:
ssd_config["augmentation"] = [{
"type":
"image",
"crop_enable":
False,
"flip_enable":
True,
"expand_ratio": [1., 4.],
"expand_probability":
0.5,
# "emit_constraint_type": "center", TODO: enable when adds support for no gt boxes
"brightness": [0.9, 1.1],
"hue": [-18, 18],
"saturation": [0.9, 1.1],
"contrast": [0.9, 1.1],
"batch_samplers": [{
"max_sample": 1,
"max_trials": 1
}, {
"max_sample": 1,
"max_trials": 50,
"sampler": {
"scale": [0.3, 1.0],
"aspect_ratio": [0.5, 2.0]
},
"sample_constraint": {
"min_jaccard_overlap": 0.1
}
}, {
"max_sample": 1,
"max_trials": 50,
"sampler": {
"scale": [0.3, 1.0],
"aspect_ratio": [0.5, 2.0]
},
"sample_constraint": {
"min_jaccard_overlap": 0.3
}
}, {
"max_sample": 1,
"max_trials": 50,
"sampler": {
"scale": [0.3, 1.0],
"aspect_ratio": [0.5, 2.0]
},
"sample_constraint": {
"min_jaccard_overlap": 0.5
}
}, {
"max_sample": 1,
"max_trials": 50,
"sampler": {
"scale": [0.3, 1.0],
"aspect_ratio": [0.5, 2.0]
},
"sample_constraint": {
"min_jaccard_overlap": 0.7
}
}, {
"max_sample": 1,
"max_trials": 50,
"sampler": {
"scale": [0.3, 1.0],
"aspect_ratio": [0.5, 2.0]
},
"sample_constraint": {
"min_jaccard_overlap": 0.9
}
}, {
"max_sample": 1,
"max_trials": 50,
"sampler": {
"scale": [0.3, 1.0],
"aspect_ratio": [0.5, 2.0]
},
"sample_constraint": {
"max_jaccard_overlap": 1.0,
"min_jaccard_overlap": 0.1
}
}]
}]
if img_reshape == (300, 300):
ssd_config['ssd_config'] = OrderedDict([('conv4_3', {
'min_sizes': 30.0,
'max_sizes': 60.0,
'aspect_ratios': 2.0,
'step': 8,
'normalize': True
}),
('fc7', {
'min_sizes': 60.0,
'max_sizes': 111.0,
'aspect_ratios':
(2.0, 3.0),
'step': 16
}),
('conv6_2', {
'min_sizes': 111.0,
'max_sizes': 162.0,
'aspect_ratios':
(2.0, 3.0),
'step': 32
}),
('conv7_2', {
'min_sizes': 162.0,
'max_sizes': 213.0,
'aspect_ratios':
(2.0, 3.0),
'step': 64
}),
('conv8_2', {
'min_sizes': 213.0,
'max_sizes': 264.0,
'aspect_ratios': 2.0,
'step': 100
}),
('conv9_2', {
'min_sizes': 264.0,
'max_sizes': 315.0,
'aspect_ratios': 2.0,
'step': 300
})])
elif img_reshape == (512, 512):
ssd_config['ssd_config'] = OrderedDict([('conv4_3', {
'min_sizes': 35.84,
'max_sizes': 76.80,
'aspect_ratios': 2.0,
'step': 8,
'normalize': True
}),
('fc7', {
'min_sizes': 76.80,
'max_sizes': 153.6,
'aspect_ratios':
(2.0, 3.0),
'step': 16
}),
('conv6_2', {
'min_sizes': 153.6,
'max_sizes': 230.4,
'aspect_ratios':
(2.0, 3.0),
'step': 32
}),
('conv7_2', {
'min_sizes': 230.4,
'max_sizes': 307.2,
'aspect_ratios':
(2.0, 3.0),
'step': 64
}),
('conv8_2', {
'min_sizes': 307.2,
'max_sizes': 384.0,
'aspect_ratios': 2.0,
'step': 128
}),
('conv9_2', {
'min_sizes': 384.0,
'max_sizes': 460.8,
'aspect_ratios': 2.0,
'step': 256
}),
('conv10_2', {
'min_sizes': 460.8,
'max_sizes': 537.8,
'aspect_ratios': 2.0,
'step': 512
})])
else:
raise ValueError(
"Image shape of {} not supported.".format(img_reshape))
return ssd_config
def get_ssd_config(img_reshape, inference=False):
ssd_config = OrderedDict()
ssd_config['batch_size'] = 32
if inference:
ssd_config['batch_size'] = 1
ssd_config['block_size'] = 50
ssd_config['cache_directory'] = get_data_cache_or_nothing(subdir='kitti_cache')
ssd_config["etl"] = [{
"type": "localization_ssd",
"height": img_reshape[0],
"width": img_reshape[1],
"max_gt_boxes": 500,
"class_names": ['__background__', 'Car', 'Van', 'Truck', 'Pedestrian',
'Person_sitting', 'Cyclist', 'Tram', 'Misc', 'DontCare']
}, {
"type": "image",
"height": img_reshape[0],
"width": img_reshape[1],
"channels": 3
}]
if not inference:
ssd_config["augmentation"] = [{
"type": "image",
"batch_samplers":
[
{
"max_sample": 1,
"max_trials": 1
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.1}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.3}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.5}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.7}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.9}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"max_jaccard_overlap": 1.0, "min_jaccard_overlap": 0.1}
}
]
}]
ssd_config['ssd_config'] = OrderedDict(
[('conv4_3', {'min_sizes': 30.0, 'max_sizes': 60.0,
'aspect_ratios': 2.0, 'step': 8, 'normalize': True}),
('fc7', {'min_sizes': 60.0, 'max_sizes': 111.0,
'aspect_ratios': (2.0, 3.0), 'step': 16}),
('conv6_2', {'min_sizes': 111.0, 'max_sizes': 162.0,
'aspect_ratios': (2.0, 3.0), 'step': 32}),
('conv7_2', {'min_sizes': 162.0, 'max_sizes': 213.0,
'aspect_ratios': (2.0, 3.0), 'step': 64}),
('conv8_2', {'min_sizes': 213.0, 'max_sizes': 264.0,
'aspect_ratios': 2.0, 'step': 100}),
('conv9_2', {'min_sizes': 264.0, 'max_sizes': 315.0,
'aspect_ratios': 2.0, 'step': {'step_h': 300, 'step_w': 100}})])
return ssd_config
def get_ssd_config(img_reshape, inference=False):
ssd_config = OrderedDict()
ssd_config['batch_size'] = 32
ssd_config['shuffle_enable'] = True
ssd_config['shuffle_manifest'] = True
if inference:
ssd_config['batch_size'] = 1
ssd_config['block_size'] = 50
ssd_config['cache_directory'] = get_data_cache_or_nothing(subdir='pascalvoc_cache')
ssd_config["etl"] = [{
"type": "localization_ssd",
"height": img_reshape[0],
"width": img_reshape[1],
"max_gt_boxes": 500,
"class_names": ["__background__", "aeroplane", "bicycle", "bird", "boat",
"bottle", "bus", "car", "cat", "chair", "cow", "diningtable",
"dog", "horse", "motorbike", "person", "pottedplant",
"sheep", "sofa", "train", "tvmonitor"]
}, {
"type": "image",
"height": img_reshape[0],
"width": img_reshape[1],
"channels": 3
}]
if not inference:
ssd_config["augmentation"] = [{
"type": "image",
"crop_enable": False,
"flip_enable": True,
"expand_ratio": [1., 4.],
"expand_probability": 0.5,
# "emit_constraint_type": "center", TODO: enable when adds support for no gt boxes
"brightness": [0.9, 1.1],
"hue": [-18, 18],
"saturation": [0.9, 1.1],
"contrast": [0.9, 1.1],
"batch_samplers":
[
{
"max_sample": 1,
"max_trials": 1
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.1}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.3}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.5}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.7}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.9}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"max_jaccard_overlap": 1.0, "min_jaccard_overlap": 0.1}
}
]
}]
if img_reshape == (300, 300):
ssd_config['ssd_config'] = OrderedDict(
[('conv4_3', {'min_sizes': 30.0, 'max_sizes': 60.0,
'aspect_ratios': 2.0, 'step': 8, 'normalize': True}),
('fc7', {'min_sizes': 60.0, 'max_sizes': 111.0,
'aspect_ratios': (2.0, 3.0), 'step': 16}),
('conv6_2', {'min_sizes': 111.0, 'max_sizes': 162.0,
'aspect_ratios': (2.0, 3.0), 'step': 32}),
('conv7_2', {'min_sizes': 162.0, 'max_sizes': 213.0,
'aspect_ratios': (2.0, 3.0), 'step': 64}),
('conv8_2', {'min_sizes': 213.0, 'max_sizes': 264.0,
'aspect_ratios': 2.0, 'step': 100}),
('conv9_2', {'min_sizes': 264.0, 'max_sizes': 315.0,
'aspect_ratios': 2.0, 'step': 300})])
elif img_reshape == (512, 512):
ssd_config['ssd_config'] = OrderedDict(
[('conv4_3', {'min_sizes': 35.84, 'max_sizes': 76.80,
'aspect_ratios': 2.0, 'step': 8, 'normalize': True}),
('fc7', {'min_sizes': 76.80, 'max_sizes': 153.6,
'aspect_ratios': (2.0, 3.0), 'step': 16}),
('conv6_2', {'min_sizes': 153.6, 'max_sizes': 230.4,
'aspect_ratios': (2.0, 3.0), 'step': 32}),
('conv7_2', {'min_sizes': 230.4, 'max_sizes': 307.2,
'aspect_ratios': (2.0, 3.0), 'step': 64}),
('conv8_2', {'min_sizes': 307.2, 'max_sizes': 384.0,
'aspect_ratios': 2.0, 'step': 128}),
('conv9_2', {'min_sizes': 384.0, 'max_sizes': 460.8,
'aspect_ratios': 2.0, 'step': 256}),
('conv10_2', {'min_sizes': 460.8, 'max_sizes': 537.8,
'aspect_ratios': 2.0, 'step': 512})])
else:
raise ValueError("Image shape of {} not supported.".format(img_reshape))
return ssd_config
def get_ssd_config(img_reshape, inference=False):
ssd_config = OrderedDict()
ssd_config['batch_size'] = 32
if inference:
ssd_config['batch_size'] = 1
ssd_config['block_size'] = 50
ssd_config['cache_directory'] = get_data_cache_or_nothing(subdir='spacenet_cache')
ssd_config["etl"] = [{
"type": "localization_ssd",
"height": img_reshape[0],
"width": img_reshape[1],
"max_gt_boxes": 500,
"class_names": ['__background__', 'building']
}, {
"type": "image",
"height": img_reshape[0],
"width": img_reshape[1],
"channels": 3
}]
if not inference:
ssd_config["augmentation"] = [{
"type": "image",
"batch_samplers":
[
{
"max_sample": 1,
"max_trials": 1
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.1}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.3}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.5}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.7}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"min_jaccard_overlap": 0.9}
},
{
"max_sample": 1,
"max_trials": 50,
"sampler": {"scale": [0.3, 1.0], "aspect_ratio": [0.5, 2.0]},
"sample_constraint": {"max_jaccard_overlap": 1.0, "min_jaccard_overlap": 0.1}
}
]
}]
ssd_config['ssd_config'] = OrderedDict(
[('conv4_3', {'min_sizes': 35.84, 'max_sizes': 76.80,
'aspect_ratios': 2.0, 'step': 8, 'normalize': True}),
('fc7', {'min_sizes': 76.80, 'max_sizes': 153.6,
'aspect_ratios': (2.0, 3.0), 'step': 16}),
('conv6_2', {'min_sizes': 153.6, 'max_sizes': 230.4,
'aspect_ratios': (2.0, 3.0), 'step': 32}),
('conv7_2', {'min_sizes': 230.4, 'max_sizes': 307.2,
'aspect_ratios': (2.0, 3.0), 'step': 64}),
('conv8_2', {'min_sizes': 307.2, 'max_sizes': 384.0,
'aspect_ratios': 2.0, 'step': 128}),
('conv9_2', {'min_sizes': 384.0, 'max_sizes': 460.8,
'aspect_ratios': 2.0, 'step': 256}),
('conv10_2', {'min_sizes': 460.8, 'max_sizes': 537.8,
'aspect_ratios': 2.0, 'step': 512})])
return ssd_config
def common_config(manifest_file, batch_size, valid_set=False):
if(dataset == "cifar10"):
# Define Cache
cache_root = get_data_cache_or_nothing('cifar10-cache/')
# Define image properties
image_config = {"type": "image",
"height": 32,
"width": 32}
# Define label properties
label_config = {"type": "label",
"binary": False}
# Define Augmentations
augmentation = {"type": "image",
"padding": 4,
"crop_enable": False,
"flip_enable": True}
# Don't enable augmentations if it is test set
if(valid_set):
return {'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config]}
return {'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]}
elif(dataset == "i1k"):
# Define cache
cache_root = get_data_cache_or_nothing("i1k-cache/")
# Define image properties
image_config = {"type": "image",
"height": 224,
"width": 224}
# Define label properties
label_config = {"type": "label",
"binary": False}
# Define Augmentations
augmentation = {"type": "image",
"center": False,
"flip_enable": True,
"scale": [0.08, 1.0],
"do_area_scale": True,
"horizontal_distortion": [0.75, 1.33],
"lighting": [0.0, 0.01],
"contrast": [0.9, 1.1],
"brightness": [0.9, 1.1],
"saturation": [0.9, 1.1]
}
# Dont do augemtations on test set
if(valid_set):
return{'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config]}
# Do augmentations on training set
return{'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]}
elif(dataset == "cifar100"):
# Define Cache
cache_root = get_data_cache_or_nothing('cifar100-cache/')
# Define image properties
image_config = {"type": "image",
"height": 32,
"width": 32}
# Define label properties
label_config = {"type": "label",
"binary": False}
# Define Augmentations
augmentation = {"type": "image",
"padding": 4,
"crop_enable": False,
"flip_enable": True}
# Don't enable augmentations if it is test set
if(valid_set):
return {'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config]}
return {'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]}
elif(dataset == "i1k100"):
# Define cache
cache_root = get_data_cache_or_nothing("i1k-cache/")
# Define image properties
image_config = {"type": "image",
"height": 224,
"width": 224}
# Define label properties
label_config = {"type": "label",
"binary": False}
# Define Augmentations
augmentation = {"type": "image",
"center": False,
"flip_enable": True,
"scale": [0.08, 1.0],
"do_area_scale": True,
"horizontal_distortion": [0.75, 1.33],
"lighting": [0.0, 0.01],
"contrast": [0.9, 1.1],
"brightness": [0.9, 1.1],
"saturation": [0.9, 1.1]
}
# Dont do augemtations on test set
if(valid_set):
return{'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config]}
# Do augmentations on training set
return{'manifest_filename': manifest_file,
'manifest_root': os.path.dirname(manifest_file),
'batch_size': batch_size,
'block_size': 5000,
'cache_directory': cache_root,
'etl': [image_config, label_config],
'augmentation': [augmentation]}
else:
raise NameError("Unkown dataset.Choose correct dataset")
