def build_dataloader(config, be, frcn_rois_per_img):
"""
Builds the dataloader for the Faster-RCNN network using our aeon loader.
Besides, the base loader, we add several operations:
1. Cast the image data into float32 format
2. Subtract the BGRMean from the image. We used pre-defined means from training
the VGG network.
3. Repack the data for Faster-RCNN model. This model has several nested branches, so
The buffers have to repacked into nested tuples to match the branch leafs. Additionally,
buffers for training the RCNN portion of the model are also allocated and provisioned
to the model.
Arguments:
config (dict): dataloader configuration
be (backend): compute backend
frcn_rois_per_img (int): Number of ROIs to use for training the RCNN portion of the
model. This is used to create the target buffers for RCNN.
Returns:
dataloader object.
"""
dl = DataLoader(config, be)
dl = TypeCast(dl, index=0, dtype=np.float32) # cast image to float
dl = BGRMeanSubtract(dl, index=0,
pixel_mean=util.FRCN_PIXEL_MEANS) # subtract means
dl = ObjectLocalization(
dl, frcn_rois_per_img=frcn_rois_per_img) # repack for faster-rcnn
return dl
def make_test_loader(manifest_file, manifest_root, backend_obj):
aeon_config = common_config(manifest_file, manifest_root, backend_obj.bsz)
aeon_config['type'] = 'audio' # No labels provided
aeon_config.pop('label', None)
dl = AeonDataLoader(aeon_config, backend_obj)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
def data_transform(dl):
""" Data is loaded from Aeon as a 4-tuple. We need to cast the audio
(index 0) from int8 to float32 and repack the data into (audio, 3-tuple).
"""
dl = TypeCast(dl, index=0, dtype=np.float32)
dl = Retuple(dl, data=(0,), target=(1, 2, 3))
return dl
def wrap_dataloader(dl):
dl = OneHot(dl, index=1, nclasses=10)
dl = TypeCast(dl, index=0, dtype=np.float32)
dl = ValueNormalize(dl,
index=0,
source_range=[0., 255.],
target_range=[-1., 1.])
return dl
def make_inference_loader(manifest_file, backend_obj):
manifest_root = "" # This is used for demo script which generates abs path manifest
aeon_config = common_config(manifest_file, manifest_root, backend_obj.bsz)
aeon_config['type'] = 'video' # No labels provided
aeon_config.pop('label', None)
dl = AeonDataLoader(aeon_config, backend_obj)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
def make_test_loader(manifest_file,
manifest_root,
backend_obj,
subset_pct=100):
aeon_config = common_config(manifest_file, manifest_root, backend_obj.bsz)
aeon_config['subset_fraction'] = float(subset_pct / 100.0)
dl = AeonDataLoader(aeon_config, backend_obj)
dl = OneHot(dl, index=1, nclasses=101)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
def make_inference_loader(manifest_file, backend_obj):
manifest_root = "" # This is used for demo script which generates abs path manifest
aeon_config = common_config(manifest_file, manifest_root, backend_obj.bsz)
video_config = {
"type": "video",
"max_frame_count": 16,
"frame": {
"height": 112,
"width": 112
}
}
aeon_config['etl'] = [video_config]
dl = AeonDataLoader(aeon_config)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
def make_train_loader(manifest_file,
manifest_root,
backend_obj,
subset_pct=100,
random_seed=0):
aeon_config = common_config(manifest_file, manifest_root, backend_obj.bsz)
aeon_config['subset_fraction'] = float(subset_pct / 100.0)
aeon_config['shuffle_manifest'] = True
aeon_config['shuffle_every_epoch'] = True
aeon_config['random_seed'] = random_seed
aeon_config['video']['frame']['center'] = False
aeon_config['video']['frame']['flip_enable'] = True
dl = AeonDataLoader(aeon_config, backend_obj)
dl = OneHot(dl, index=1, nclasses=101)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
def test_anchor_target_layer(backend_default, fargs):
(height, width) = fargs
manifest_path = os.environ['PASCAL_MANIFEST_PATH']
assert manifest_path is not None, "Please set the PASCAL_MANIFEST_PATH variable."
manifest_root = os.environ['PASCAL_MANIFEST_ROOT']
assert manifest_root is not None, "Please set the PASCAL_MANIFEST_ROOT variable."
config = PASCALVOC(manifest_path, manifest_root, cache_dir='',
height=height, width=width, inference=False)
config['subset_fraction'] = 0.1
dl = DataLoader(config, backend_default)
dl = TypeCast(dl, index=0, dtype=np.float32)
train_set = ObjectLocalization(dl, frcn_rois_per_img=128)
for idx, (X, Y) in enumerate(train_set):
reference_test(train_set, X, Y)
def build_dataloader(config,
manifest_root,
batch_size,
subset_pct=100,
PIXEL_MEANS=np.array([104, 117, 123])):
"""
Builds the dataloader for the Faster-RCNN network using our aeon loader.
Besides, the base loader, we add several operations:
1. Cast the image data into float32 format
2. Subtract the BGRMean from the image. We used pre-defined means from training
the VGG network.
3. Repack the data for Faster-RCNN model. This model has several nested branches, so
The buffers have to repacked into nested tuples to match the branch leafs. Additionally,
buffers for training the RCNN portion of the model are also allocated and provisioned
to the model.
Arguments:
config (dict): dataloader configuration
be (backend): compute backend
Returns:
dataloader object.
"""
# assert config['minibatch_size'] == be.bsz,
# 'Dataloader config\'s minibatch size not matching backend bsz'
config["manifest_root"] = manifest_root
config["batch_size"] = batch_size
config["subset_fraction"] = float(subset_pct / 100.0)
dl = DataLoaderAdapter(DataLoader(config))
dl = TypeCast(dl, index=5, dtype=np.float32) # cast image to float
dl = BGRMeanSubtract(dl, index=5, pixel_mean=PIXEL_MEANS) # subtract means
dl = ObjectLocalization(dl)
dl.set_classes(config['etl'][0]['class_names'])
dl.shape = dl.shapes()[5]
return dl
def wrap_dataloader(dl):
dl = OneHot(dl, index=1, nclasses=2)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
def transformers(dl):
dl = OneHot(dl, nclasses=10, index=1)
dl = TypeCast(dl, index=0, dtype=np.float32)
dl = BGRMeanSubtract(dl, index=0, pixel_mean=bgr_means)
return dl
brightness=(0.9, 1.1),
scale=(0.75, 0.75),
fixed_aspect_ratio=True)
label_config = dict(binary=False)
config = dict(type="image,label",
image=image_config,
label=label_config,
manifest_filename='manifest_all_but_9.csv',
minibatch_size=args.batch_size,
macrobatch_size=128,
cache_directory='cache_dir',
shuffle_manifest=True,
shuffle_every_epoch=True)
#shuffle_every_epoch = True)
train_set = DataLoader(config, be)
train_set = TypeCast(train_set, index=0,
dtype=np.float32) # cast image to float
# Set up the validation set to load via aeon
image_config = dict(height=patch_size, width=patch_size, channels=3)
label_config = dict(binary=False)
config = dict(type="image,label",
image=image_config,
label=label_config,
manifest_filename='manifest_subset9_augmented.csv',
minibatch_size=args.batch_size)
valid_set = DataLoader(config, be)
valid_set = TypeCast(valid_set, index=0,
dtype=np.float32) # cast image to float
class WeightedBinaryCrossEntropyCost(Cost):
def wrap_dataloader(dl):
dl = OneHot(dl, index=1, nclasses=10)
dl = TypeCast(dl, index=0, dtype=np.float32)
dl = BGRMeanSubtract(dl, index=0)
return dl
def wrap_dataloader(aeon_config):
dl = AeonDataLoader(aeon_config)
dl = OneHot(dl, index=1, nclasses=101)
dl = TypeCast(dl, index=0, dtype=np.float32)
return dl
# setup backend
be = gen_backend(**extract_valid_args(args, gen_backend))
be.enable_winograd = 4 # default to winograd 4 for fast autotune
# Set up the testset to load via aeon
image_config = dict(height=64, width=64, channels=1)
label_config = dict(binary=False)
config = dict(type="image,label",
image=image_config,
label=label_config,
manifest_filename=testFileName,
minibatch_size=args.batch_size,
subset_fraction=1)
test_set = DataLoader(config, be)
test_set = TypeCast(test_set, index=0, dtype=np.float32) # cast image to float
relu = Rectlin()
init_uni = GlorotUniform()
convp1 = dict(init=init_uni, batch_norm=True, activation=relu, padding=1)
layers = [
Conv((5, 5, 24), **convp1),
Pooling(2, op='max'),
Conv((3, 3, 32), **convp1),
Pooling(2, op='max'),
Conv((3, 3, 48), **convp1),
Dropout(keep=.6),
Pooling(2, op='max'),
Affine(nout=64, init=init_uni, activation=relu),
Dropout(keep=.4),
